Docs: Update for release FEX-2509.1

This takes an f80 as input and returns an f32. A copy-paste error had
this truncating the input float if !TMP_ABIARGS.

.github/workflows/ccpp.yml

@@ -13,6 +13,7 @@ env:
   BUILD_TYPE: Release
   CC: clang
   CXX: clang++
+  FEX_PORTABLE: 1
 
 jobs:
   build_plus_test:
@@ -136,6 +137,9 @@ jobs:
     - name: FEXLinuxTests
       working-directory: ${{runner.workspace}}/build
       shell: bash
+      env:
+        # These tests require non-portable install due to thunks.
+        FEX_PORTABLE: 0
       run: cmake --build . --config $BUILD_TYPE --target fex_linux_tests_all
 
     - name: FEXLinuxTests Results move

.github/workflows/glibc_fault.yml

@@ -20,6 +20,7 @@ env:
   BUILD_TYPE: Release
   CC: clang
   CXX: clang++
+  FEX_PORTABLE: 1
 
 jobs:
   glibc_fault_test:

.github/workflows/hostrunner.yml

@@ -13,6 +13,7 @@ env:
   BUILD_TYPE: Release
   CC: clang
   CXX: clang++
+  FEX_PORTABLE: 1
 
 jobs:
   hostrunner_tests:

.github/workflows/pr-code-format.yml

@@ -60,7 +60,6 @@ jobs:
           START_REV: ${{ github.event.pull_request.base.sha }}
           END_REV: ${{ github.event.pull_request.head.sha }}
           CHANGED_FILES: ${{ steps.changed-files.outputs.all_changed_files }}
-        # Using --diff_from_common_commit option available in clang-format-19
         run: |
           python ./External/code-format-helper/code-format-helper.py \
             --repo "FEX-emu/FEX" \

.github/workflows/vixl_simulator.yml

@@ -13,6 +13,7 @@ env:
   BUILD_TYPE: Release
   CC: clang
   CXX: clang++
+  FEX_PORTABLE: 1
 
 jobs:
   vixl_simulator:

.gitmodules

@@ -46,3 +46,6 @@
 [submodule "External/tracy"]
 	path = External/tracy
 	url = https://github.com/wolfpld/tracy
+[submodule "External/range-v3"]
+	path = External/range-v3
+	url = https://github.com/ericniebler/range-v3.git

CMakeLists.txt

@@ -354,6 +354,12 @@ if (NOT fmt_FOUND)
   add_subdirectory(External/fmt/)
 endif()
 
+find_package(range-v3 QUIET)
+if (NOT range-v3_FOUND)
+  add_subdirectory(External/range-v3/)
+  target_compile_definitions(range-v3 INTERFACE RANGES_DISABLE_DEPRECATED_WARNINGS)
+endif()
+
 add_subdirectory(External/tiny-json/)
 include_directories(External/tiny-json/)
 

CodeEmitter/CodeEmitter/ALUOps.inl

@@ -174,7 +174,7 @@ public:
   // Logical immediate
   void and_(ARMEmitter::Size s, ARMEmitter::Register rd, ARMEmitter::Register rn, uint64_t Imm) {
     uint32_t n, immr, imms;
-    [[maybe_unused]] const auto IsImm = IsImmLogical(Imm, RegSizeInBits(s), &n, &imms, &immr);
+    const auto IsImm = IsImmLogical(Imm, RegSizeInBits(s), &n, &imms, &immr);
     LOGMAN_THROW_A_FMT(IsImm, "Couldn't encode immediate to logical op");
     and_(s, rd, rn, n, immr, imms);
   }
@@ -185,7 +185,7 @@ public:
 
   void ands(ARMEmitter::Size s, ARMEmitter::Register rd, ARMEmitter::Register rn, uint64_t Imm) {
     uint32_t n, immr, imms;
-    [[maybe_unused]] const auto IsImm = IsImmLogical(Imm, RegSizeInBits(s), &n, &imms, &immr);
+    const auto IsImm = IsImmLogical(Imm, RegSizeInBits(s), &n, &imms, &immr);
     LOGMAN_THROW_A_FMT(IsImm, "Couldn't encode immediate to logical op");
     ands(s, rd, rn, n, immr, imms);
   }
@@ -196,14 +196,14 @@ public:
 
   void orr(ARMEmitter::Size s, ARMEmitter::Register rd, ARMEmitter::Register rn, uint64_t Imm) {
     uint32_t n, immr, imms;
-    [[maybe_unused]] const auto IsImm = IsImmLogical(Imm, RegSizeInBits(s), &n, &imms, &immr);
+    const auto IsImm = IsImmLogical(Imm, RegSizeInBits(s), &n, &imms, &immr);
     LOGMAN_THROW_A_FMT(IsImm, "Couldn't encode immediate to logical op");
     orr(s, rd, rn, n, immr, imms);
   }
 
   void eor(ARMEmitter::Size s, ARMEmitter::Register rd, ARMEmitter::Register rn, uint64_t Imm) {
     uint32_t n, immr, imms;
-    [[maybe_unused]] const auto IsImm = IsImmLogical(Imm, RegSizeInBits(s), &n, &imms, &immr);
+    const auto IsImm = IsImmLogical(Imm, RegSizeInBits(s), &n, &imms, &immr);
     LOGMAN_THROW_A_FMT(IsImm, "Couldn't encode immediate to logical op");
     eor(s, rd, rn, n, immr, imms);
   }
@@ -333,7 +333,7 @@ public:
     bfi(s, rd, Reg::zr, lsb, width);
   }
   void bfxil(ARMEmitter::Size s, Register rd, Register rn, uint32_t lsb, uint32_t width) {
-    [[maybe_unused]] const auto reg_size_bits = RegSizeInBits(s);
+    const auto reg_size_bits = RegSizeInBits(s);
     const auto lsb_p_width = lsb + width;
 
     LOGMAN_THROW_A_FMT(width >= 1, "bfxil needs width >= 1");
@@ -977,7 +977,7 @@ private:
   }
 
   void xbfiz_helper(bool is_signed, ARMEmitter::Size s, Register rd, Register rn, uint32_t lsb, uint32_t width) {
-    [[maybe_unused]] const auto lsb_p_width = lsb + width;
+    const auto lsb_p_width = lsb + width;
     const auto reg_size_bits = RegSizeInBits(s);
 
     LOGMAN_THROW_A_FMT(lsb_p_width <= reg_size_bits, "lsb + width ({}) must be <= {}. lsb={}, width={}", lsb_p_width, reg_size_bits, lsb, width);

CodeEmitter/CodeEmitter/Emitter.h

@@ -12,6 +12,7 @@
 #include <CodeEmitter/Registers.h>
 
 #include <array>
+#include <bit>
 #include <cstdint>
 #include <utility>
 #include <type_traits>

CodeEmitter/CodeEmitter/SVEOps.inl

@@ -1541,7 +1541,7 @@ public:
   void sqincp(SubRegSize size, XRegister rdn, PRegister pm) {
     SVEIncDecPredicateCountScalar(0, 1, 0b10, 0b00, size, rdn, pm);
   }
-  void sqincp(SubRegSize size, XRegister rdn, PRegister pm, [[maybe_unused]] WRegister wn) {
+  void sqincp(SubRegSize size, XRegister rdn, PRegister pm, WRegister wn) {
     LOGMAN_THROW_A_FMT(rdn.Idx() == wn.Idx(), "rdn and wn must be the same");
     SVEIncDecPredicateCountScalar(0, 1, 0b00, 0b00, size, rdn, pm);
   }
@@ -1554,7 +1554,7 @@ public:
   void sqdecp(SubRegSize size, XRegister rdn, PRegister pm) {
     SVEIncDecPredicateCountScalar(0, 1, 0b10, 0b10, size, rdn, pm);
   }
-  void sqdecp(SubRegSize size, XRegister rdn, PRegister pm, [[maybe_unused]] WRegister wn) {
+  void sqdecp(SubRegSize size, XRegister rdn, PRegister pm, WRegister wn) {
     LOGMAN_THROW_A_FMT(rdn.Idx() == wn.Idx(), "rdn and wn must be the same");
     SVEIncDecPredicateCountScalar(0, 1, 0b00, 0b10, size, rdn, pm);
   }
@@ -3296,7 +3296,7 @@ private:
     const auto log2_size_bytes = FEXCore::ilog2(size_bytes);
 
     // We can index up to 512-bit registers with dup
-    [[maybe_unused]] const auto max_index = (64U >> log2_size_bytes) - 1;
+    const auto max_index = (64U >> log2_size_bytes) - 1;
     LOGMAN_THROW_A_FMT(Index <= max_index, "dup index ({}) too large. Must be within [0, {}].", Index, max_index);
 
     // imm2:tsz make up a 7 bit wide field, with each increasing element size
@@ -3326,7 +3326,7 @@ private:
 
     uint32_t shift = 0;
     if (!is_uint8_imm) {
-      [[maybe_unused]] const bool is_uint16_imm = (imm >> 16) == 0;
+      const bool is_uint16_imm = (imm >> 16) == 0;
 
       LOGMAN_THROW_A_FMT(is_uint16_imm, "Immediate ({}) must be a 16-bit value within [256, 65280]", imm);
       LOGMAN_THROW_A_FMT((imm % 256) == 0, "Immediate ({}) must be a multiple of 256", imm);
@@ -4152,7 +4152,7 @@ private:
 
     const auto& op_data = mem_op.MetaType.ScalarVectorType;
     const bool is_scaled = op_data.scale != 0;
-    [[maybe_unused]] const auto msize_value = FEXCore::ToUnderlying(msize);
+    const auto msize_value = FEXCore::ToUnderlying(msize);
 
     LOGMAN_THROW_A_FMT(op_data.scale == 0 || op_data.scale == msize_value, "scale may only be 0 or {}", msize_value);
 
@@ -4266,7 +4266,7 @@ private:
     const auto msize_value = FEXCore::ToUnderlying(msize);
     const auto msize_bytes = 1U << msize_value;
 
-    [[maybe_unused]] const auto imm_limit = (32U << msize_value) - msize_bytes;
+    const auto imm_limit = (32U << msize_value) - msize_bytes;
     const auto imm = mem_op.MetaType.VectorImmType.Imm;
     const auto imm_to_encode = imm >> msize_value;
 
@@ -4332,8 +4332,8 @@ private:
     LOGMAN_THROW_A_FMT(pg <= PReg::p7, "Can only use p0-p7 as a governing predicate");
     LOGMAN_THROW_A_FMT((imm % num_regs) == 0, "Offset must be a multiple of {}", num_regs);
 
-    [[maybe_unused]] const auto min_offset = -8 * num_regs;
-    [[maybe_unused]] const auto max_offset = 7 * num_regs;
+    const auto min_offset = -8 * num_regs;
+    const auto max_offset = 7 * num_regs;
     LOGMAN_THROW_A_FMT(imm >= min_offset && imm <= max_offset,
                        "Invalid load/store offset ({}). Offset must be a multiple of {} and be within [{}, {}]", imm, num_regs, min_offset,
                        max_offset);
@@ -4440,8 +4440,8 @@ private:
     LOGMAN_THROW_A_FMT(pg <= PReg::p7, "Can only use p0-p7 as a governing predicate");
 
     const auto esize = static_cast<int>(16 << ssz);
-    [[maybe_unused]] const auto max_imm = (esize << 3) - esize;
-    [[maybe_unused]] const auto min_imm = -(max_imm + esize);
+    const auto max_imm = (esize << 3) - esize;
+    const auto min_imm = -(max_imm + esize);
 
     LOGMAN_THROW_A_FMT((imm % esize) == 0, "imm ({}) must be a multiple of {}", imm, esize);
     LOGMAN_THROW_A_FMT(imm >= min_imm && imm <= max_imm, "imm ({}) must be within [{}, {}]", imm, min_imm, max_imm);
@@ -4485,7 +4485,7 @@ private:
     const auto msize_value = FEXCore::ToUnderlying(msize);
 
     const auto data_size_bytes = 1U << msize_value;
-    [[maybe_unused]] const auto max_imm = (64U << msize_value) - data_size_bytes;
+    const auto max_imm = (64U << msize_value) - data_size_bytes;
     LOGMAN_THROW_A_FMT((imm % data_size_bytes) == 0 && imm <= max_imm, "imm must be a multiple of {} and be within [0, {}]",
                        data_size_bytes, max_imm);
 
@@ -4861,7 +4861,7 @@ private:
                        "64-bit variants may only use Zm between z0-z15");
 
     const auto Underlying = FEXCore::ToUnderlying(size);
-    [[maybe_unused]] const uint32_t IndexMax = (16 / (1U << Underlying)) - 1;
+    const uint32_t IndexMax = (16 / (1U << Underlying)) - 1;
     LOGMAN_THROW_A_FMT(index <= IndexMax, "Index must be within 0-{}", IndexMax);
 
     // Can be bit 20 or 19 depending on whether or not the element size is 64-bit.
@@ -5117,12 +5117,13 @@ private:
   requires (std::is_same_v<T, float> || std::is_same_v<T, double>)
   using FloatToEquivalentUInt = std::conditional_t<std::is_same_v<T, float>, uint32_t, uint64_t>;
 
+#if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
   // Determines if a floating-point value is capable of being converted
   // into an 8-bit immediate. See pseudocode definition of VFPExpandImm
   // in ARM A-profile reference manual for a general overview of how this was derived.
   template<typename T>
   requires (std::is_same_v<T, float> || std::is_same_v<T, double>)
-  [[nodiscard, maybe_unused]]
+  [[nodiscard]]
   static bool IsValidFPValueForImm8(T value) {
     const uint64_t bits = FEXCore::BitCast<FloatToEquivalentUInt<T>>(value);
     const uint64_t datasize_idx = FEXCore::ilog2(sizeof(T)) - 1;
@@ -5162,10 +5163,13 @@ private:
 
     return true;
   }
+#endif
 
 protected:
   static uint32_t FP32ToImm8(float value) {
+#if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
     LOGMAN_THROW_A_FMT(IsValidFPValueForImm8(value), "Value ({}) cannot be encoded into an 8-bit immediate", value);
+#endif
 
     const auto bits = FEXCore::BitCast<uint32_t>(value);
     const auto sign = (bits & 0x80000000) >> 24;
@@ -5176,7 +5180,9 @@ protected:
   }
 
   static uint32_t FP64ToImm8(double value) {
+#if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
     LOGMAN_THROW_A_FMT(IsValidFPValueForImm8(value), "Value ({}) cannot be encoded into an 8-bit immediate", value);
+#endif
 
     const auto bits = FEXCore::BitCast<uint64_t>(value);
     const auto sign = (bits & 0x80000000'00000000) >> 56;
@@ -5202,7 +5208,7 @@ private:
     uint32_t shift = 0;
     if (!is_int8_imm) {
       const int32_t imm16_limit = 32768;
-      [[maybe_unused]] const bool is_int16_imm = -imm16_limit <= imm && imm < imm16_limit;
+      const bool is_int16_imm = -imm16_limit <= imm && imm < imm16_limit;
 
       LOGMAN_THROW_A_FMT(is_int16_imm, "Immediate ({}) must be a 16-bit value within [-32768, 32512]", imm);
       LOGMAN_THROW_A_FMT((imm % 256) == 0, "Immediate ({}) must be a multiple of 256", imm);

CodeEmitter/CodeEmitter/ScalarOps.inl

@@ -30,7 +30,7 @@ public:
     const uint32_t SizeImm = FEXCore::ToUnderlying(size);
     const uint32_t IndexShift = SizeImm + 1;
     const uint32_t ElementSize = 1U << SizeImm;
-    [[maybe_unused]] const uint32_t MaxIndex = 128U / (ElementSize * 8);
+    const uint32_t MaxIndex = 128U / (ElementSize * 8);
 
     LOGMAN_THROW_A_FMT(Index < MaxIndex, "Index too large. Index={}, Max Index: {}", Index, MaxIndex);
 
@@ -1381,7 +1381,7 @@ private:
   void ASIMDScalarXIndexedElement(uint32_t U, ScalarRegSize size, uint32_t opcode, VRegister rm, VRegister rn, VRegister rd, uint32_t index) {
     LOGMAN_THROW_A_FMT(size != ScalarRegSize::i8Bit, "Scalar size must not be 8-bit");
 
-    [[maybe_unused]] const auto invalid_bound = 16U >> FEXCore::ToUnderlying(size);
+    const auto invalid_bound = 16U >> FEXCore::ToUnderlying(size);
     LOGMAN_THROW_A_FMT(index < invalid_bound, "Index ({}) must be within [0-{}]", index, invalid_bound - 1);
 
     uint32_t Instr = 0b0101'1111'0000'0000'0000'0000'0000'0000;

CodeEmitter/CodeEmitter/VixlUtils.inl

@@ -41,7 +41,6 @@ static bool IsImmLogical(uint64_t value, unsigned width, unsigned* n = nullptr,
   [[maybe_unused]] constexpr auto kDRegSize = 64;
 
   constexpr auto kWRegSize = 32;
-  constexpr auto kXRegSize = 64;
 
   LOGMAN_THROW_A_FMT((width == kBRegSize) || (width == kHRegSize) || (width == kSRegSize) || (width == kDRegSize), "Unexpected imm size");
 
@@ -129,8 +128,8 @@ static bool IsImmLogical(uint64_t value, unsigned width, unsigned* n = nullptr,
     // Compute the repeat distance d, and set up a bitmask covering the basic
     // unit of repetition (i.e. a word with the bottom d bits set). Also, in all
     // of these cases the N bit of the output will be zero.
-    clz_a = CountLeadingZeros(a, kXRegSize);
-    int clz_c = CountLeadingZeros(c, kXRegSize);
+    clz_a = std::countl_zero(a);
+    int clz_c = std::countl_zero(c);
     d = clz_a - clz_c;
     mask = ((UINT64_C(1) << d) - 1);
     out_n = 0;
@@ -151,7 +150,7 @@ static bool IsImmLogical(uint64_t value, unsigned width, unsigned* n = nullptr,
       // of set bits in our word, meaning that we have the trivial case of
       // d == 64 and only one 'repetition'. Set up all the same variables as in
       // the general case above, and set the N bit in the output.
-      clz_a = CountLeadingZeros(a, kXRegSize);
+      clz_a = std::countl_zero(a);
       d = 64;
       mask = ~UINT64_C(0);
       out_n = 1;
@@ -159,7 +158,7 @@ static bool IsImmLogical(uint64_t value, unsigned width, unsigned* n = nullptr,
   }
 
   // If the repeat period d is not a power of two, it can't be encoded.
-  if (!IsPowerOf2(d)) {
+  if (!std::has_single_bit(uint32_t(d))) {
     return false;
   }
 
@@ -179,7 +178,7 @@ static bool IsImmLogical(uint64_t value, unsigned width, unsigned* n = nullptr,
   static const uint64_t multipliers[] = {
     0x0000000000000001UL, 0x0000000100000001UL, 0x0001000100010001UL, 0x0101010101010101UL, 0x1111111111111111UL, 0x5555555555555555UL,
   };
-  uint64_t multiplier = multipliers[CountLeadingZeros(d, kXRegSize) - 57];
+  uint64_t multiplier = multipliers[std::countl_zero(uint64_t(d)) - 57];
   uint64_t candidate = (b - a) * multiplier;
 
   if (value != candidate) {
@@ -194,7 +193,7 @@ static bool IsImmLogical(uint64_t value, unsigned width, unsigned* n = nullptr,
   // Count the set bits in our basic stretch. The special case of clz(0) == -1
   // makes the answer come out right for stretches that reach the very top of
   // the word (e.g. numbers like 0xffffc00000000000).
-  int clz_b = (b == 0) ? -1 : CountLeadingZeros(b, kXRegSize);
+  int clz_b = (b == 0) ? -1 : std::countl_zero(b);
   int s = clz_a - clz_b;
 
   // Decide how many bits to rotate right by, to put the low bit of that basic
@@ -285,11 +284,6 @@ INT_1_TO_63_LIST(DECLARE_IS_UINT_N)
 
 private:
 
-template<typename V>
-static inline bool IsPowerOf2(V value) {
-  return (value != 0) && ((value & (value - 1)) == 0);
-}
-
 // Some compilers dislike negating unsigned integers,
 // so we provide an equivalent.
 template<typename T>
@@ -302,50 +296,4 @@ static inline uint64_t LowestSetBit(uint64_t value) {
   return value & UnsignedNegate(value);
 }
 
-template<typename V>
-static inline int CountLeadingZeros(V value, int width = (sizeof(V) * 8)) {
-#if COMPILER_HAS_BUILTIN_CLZ
-  if (width == 32) {
-    return (value == 0) ? 32 : __builtin_clz(static_cast<unsigned>(value));
-  } else if (width == 64) {
-    return (value == 0) ? 64 : __builtin_clzll(value);
-  }
-#endif
-  return CountLeadingZerosFallBack(value, width);
-}
-
-static inline int CountLeadingZerosFallBack(uint64_t value, int width) {
-  LOGMAN_THROW_A_FMT(IsPowerOf2(width) && (width <= 64), "Invalid width");
-  if (value == 0) {
-    return width;
-  }
-  int count = 0;
-  value = value << (64 - width);
-  if ((value & UINT64_C(0xffffffff00000000)) == 0) {
-    count += 32;
-    value = value << 32;
-  }
-  if ((value & UINT64_C(0xffff000000000000)) == 0) {
-    count += 16;
-    value = value << 16;
-  }
-  if ((value & UINT64_C(0xff00000000000000)) == 0) {
-    count += 8;
-    value = value << 8;
-  }
-  if ((value & UINT64_C(0xf000000000000000)) == 0) {
-    count += 4;
-    value = value << 4;
-  }
-  if ((value & UINT64_C(0xc000000000000000)) == 0) {
-    count += 2;
-    value = value << 2;
-  }
-  if ((value & UINT64_C(0x8000000000000000)) == 0) {
-    count += 1;
-  }
-  count += (value == 0);
-  return count;
-}
-
 public:

External/code-format-helper/code-format-helper.py

@@ -214,7 +214,6 @@ class ClangFormatHelper(FormatHelper):
             self.clang_fmt_path,
             "--binary=clang-format-19",
             "--diff",
-            "--diff_from_common_commit",
         ]
 
         if args.start_rev and args.end_rev:

FEXCore/Scripts/config_generator.py

@@ -418,7 +418,7 @@ def print_parse_argloader_options(options):
 
             if (value_type == "strenum"):
                 output_argloader.write("\tfextl::string UserValue = Options[\"{0}\"];\n".format(op_key))
-                output_argloader.write("\tSet(FEXCore::Config::ConfigOption::CONFIG_{}, FEXCore::Config::EnumParser<FEXCore::Config::{}ConfigPair>(FEXCore::Config::{}_EnumPairs, UserValue));\n".format(op_key.upper(), op_key, op_key, op_key))
+                output_argloader.write("\tSet(FEXCore::Config::ConfigOption::CONFIG_{}, FEXCore::Config::EnumParser<FEXCore::Config::{}ConfigPair>(FEXCore::Config::{}_EnumPairs, UserValue));\n".format(op_key.upper(), op_key, op_key))
             elif (value_type == "strarray"):
                 # these need a bit more help
                 output_argloader.write("\tauto Array = Options.all(\"{0}\");\n".format(op_key))
@@ -447,13 +447,13 @@ def print_parse_envloader_options(options):
             value_type = op_vals["Type"]
             if (value_type == "strenum"):
                 output_argloader.write("else if (Key == \"FEX_{0}\") {{\n".format(op_key.upper()))
-                output_argloader.write("Value = FEXCore::Config::EnumParser<FEXCore::Config::{}ConfigPair>(FEXCore::Config::{}_EnumPairs, Value_View);\n".format(op_key, op_key, op_key))
+                output_argloader.write("\tValue = FEXCore::Config::EnumParser<FEXCore::Config::{}ConfigPair>(FEXCore::Config::{}_EnumPairs, Value_View);\n".format(op_key, op_key))
                 output_argloader.write("}\n")
 
             if ("ArgumentHandler" in op_vals):
                 conversion_func = "FEXCore::Config::Handler::{0}".format(op_vals["ArgumentHandler"])
                 output_argloader.write("else if (Key == \"FEX_{0}\") {{\n".format(op_key.upper()))
-                output_argloader.write("Value = {0}(Value_View);\n".format(conversion_func))
+                output_argloader.write("\tValue = {0}(Value_View);\n".format(conversion_func))
                 output_argloader.write("}\n")
     output_argloader.write("#endif\n")
 
@@ -467,15 +467,15 @@ def print_parse_jsonloader_options(options):
             value_type = op_vals["Type"]
             if (value_type == "strenum"):
                 output_argloader.write("else if (KeyName == \"{0}\") {{\n".format(op_key))
-                output_argloader.write("\tSet(KeyOption, FEXCore::Config::EnumParser<FEXCore::Config::{}ConfigPair>(FEXCore::Config::{}_EnumPairs, Value_View));\n".format(op_key, op_key, op_key))
+                output_argloader.write("\tSet(KeyOption, FEXCore::Config::EnumParser<FEXCore::Config::{}ConfigPair>(FEXCore::Config::{}_EnumPairs, Value_View));\n".format(op_key, op_key))
                 output_argloader.write("}\n")
             elif (value_type == "strarray"):
                 output_argloader.write("else if (KeyName == \"{0}\") {{\n".format(op_key))
                 output_argloader.write("\tAppendStrArrayValue(KeyOption, ConfigString);\n")
                 output_argloader.write("}\n")
     assert op_key is not None, "No options found in JSONLOADER"
-    output_argloader.write("else {{\n".format(op_key))
-    output_argloader.write("Set(KeyOption, ConfigString);\n")
+    output_argloader.write("else {\n")
+    output_argloader.write("\tSet(KeyOption, ConfigString);\n")
     output_argloader.write("}\n")
 
     output_argloader.write("#endif\n")

FEXCore/Scripts/json_ir_generator.py

@@ -58,6 +58,7 @@ class OpDefinition:
     JITDispatch: bool
     JITDispatchOverride: str
     TiedSource: int
+    Inline: list
     Arguments: list
     EmitValidation: list
     Desc: list
@@ -278,6 +279,12 @@ def parse_ops(ops):
             if "TiedSource" in op_val:
                 OpDef.TiedSource = op_val["TiedSource"]
 
+            # Pad Inline out to the argument count
+            OpDef.Inline = [''] * len(OpDef.Arguments)
+            if "Inline" in op_val:
+                Value = op_val["Inline"]
+                OpDef.Inline[0:len(Value)] = Value
+
             # Do some fixups of the data here
             if len(OpDef.EmitValidation) != 0:
                 for i in range(len(OpDef.EmitValidation)):
@@ -397,16 +404,16 @@ def print_ir_sizes():
     // Make sure our array maps directly to the IROps enum
     static_assert(IRSizes[IROps::OP_LAST] == -1ULL);
 
-    [[maybe_unused, nodiscard]] static size_t GetSize(IROps Op) { return IRSizes[Op]; }
-    [[nodiscard, gnu::const, gnu::visibility("default")]] std::string_view const& GetName(IROps Op);
-    [[nodiscard, gnu::const, gnu::visibility("default")]] uint8_t GetArgs(IROps Op);
-    [[nodiscard, gnu::const, gnu::visibility("default")]] uint8_t GetRAArgs(IROps Op);
-    [[nodiscard, gnu::const, gnu::visibility("default")]] FEXCore::IR::RegisterClassType GetRegClass(IROps Op);
-    [[nodiscard, gnu::const, gnu::visibility("default")]] bool HasSideEffects(IROps Op);
-    [[nodiscard, gnu::const, gnu::visibility("default")]] bool ImplicitFlagClobber(IROps Op);
-    [[nodiscard, gnu::const, gnu::visibility("default")]] bool GetHasDest(IROps Op);
-    [[nodiscard, gnu::const, gnu::visibility("default")]] bool LoweredX87(IROps Op);
-    [[nodiscard, gnu::const, gnu::visibility("default")]] int8_t TiedSource(IROps Op);
+    [[nodiscard]] inline size_t GetSize(IROps Op) { return IRSizes[Op]; }
+    [[nodiscard, gnu::const]] std::string_view const& GetName(IROps Op);
+    [[nodiscard, gnu::const]] uint8_t GetArgs(IROps Op);
+    [[nodiscard, gnu::const]] uint8_t GetRAArgs(IROps Op);
+    [[nodiscard, gnu::const]] FEXCore::IR::RegisterClassType GetRegClass(IROps Op);
+    [[nodiscard, gnu::const]] bool HasSideEffects(IROps Op);
+    [[nodiscard, gnu::const]] bool ImplicitFlagClobber(IROps Op);
+    [[nodiscard, gnu::const]] bool GetHasDest(IROps Op);
+    [[nodiscard, gnu::const]] bool LoweredX87(IROps Op);
+    [[nodiscard, gnu::const]] int8_t TiedSource(IROps Op);
 
     #undef IROP_SIZES
     #endif
@@ -588,13 +595,13 @@ def print_ir_arg_printer():
     output_file.write("#endif\n")
 
 def print_validation(op):
-    if op.EmitValidation != None:
-        output_file.write("\t\t#if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED\n")
+    if len(op.EmitValidation) != 0:
+        output_file.write("#if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED\n")
 
         for Validation in op.EmitValidation:
             Sanitized = Validation.replace("\"", "\\\"")
-            output_file.write("\tLOGMAN_THROW_A_FMT({}, \"{}\");\n".format(Validation, Sanitized))
-        output_file.write("\t\t#endif\n")
+            output_file.write("\t\tLOGMAN_THROW_A_FMT({}, \"{}\");\n".format(Validation, Sanitized))
+        output_file.write("#endif\n")
 
 # Print out IR allocator helpers
 def print_ir_allocator_helpers():
@@ -773,9 +780,29 @@ def print_ir_allocator_helpers():
                 output_file.write(") {\n")
                 output_file.write("\t\tauto ListDataBegin = DualListData.ListBegin();\n")
 
+                idx = 0
                 for arg in op.Arguments:
                     if arg.IsSSA:
-                        output_file.write("\t\t{}->AddUse();\n".format(arg.Name))
+                        # Inline an immediate if we can
+                        inline = op.Inline[idx]
+                        idx += 1
+
+                        if inline != '':
+                            Sized = "Size" in [x.Name for x in op.Arguments]
+                            P = ["Size" if Sized else "OpSize::i64Bit", arg.Name]
+
+                            # A few cases need extra info plumbed.
+                            if inline == "SubtractZero":
+                                P += ["Src2"]
+                            elif inline == "Mem":
+                                P += ["OffsetType", "OffsetScale"]
+                            elif inline == "Memtso":
+                                P += ["OffsetType", "OffsetScale", "true /* TSO */"]
+                                inline = "Mem"
+
+                            output_file.write(f"\t\t{arg.Name} = Inline{inline}({', '.join(P)});\n")
+
+                        output_file.write(f"\t\t{arg.Name}->AddUse();\n")
 
                 # Insert validation here. This is skipped for the
                 # OrderedNodeWrapper version because validation can depend on

FEXCore/Source/CMakeLists.txt

@@ -23,9 +23,6 @@ set (SRCS
   Interface/Core/Addressing.cpp
   Interface/Core/CPUID.cpp
   Interface/Core/Frontend.cpp
-  Interface/Core/ObjectCache/JobHandling.cpp
-  Interface/Core/ObjectCache/NamedRegionObjectHandler.cpp
-  Interface/Core/ObjectCache/ObjectCacheService.cpp
   Interface/Core/OpcodeDispatcher/AVX_128.cpp
   Interface/Core/OpcodeDispatcher/Crypto.cpp
   Interface/Core/OpcodeDispatcher/Flags.cpp
@@ -33,7 +30,6 @@ set (SRCS
   Interface/Core/OpcodeDispatcher/X87.cpp
   Interface/Core/OpcodeDispatcher/X87F64.cpp
   Interface/Core/OpcodeDispatcher.cpp
-  Interface/Core/X86Tables.cpp
   Interface/Core/X86HelperGen.cpp
   Interface/Core/ArchHelpers/Arm64Emitter.cpp
   Interface/Core/Dispatcher/Dispatcher.cpp
@@ -65,7 +61,6 @@ set (SRCS
   Interface/IR/IRDumper.cpp
   Interface/IR/IREmitter.cpp
   Interface/IR/PassManager.cpp
-  Interface/IR/Passes/ConstProp.cpp
   Interface/IR/Passes/IRDumperPass.cpp
   Interface/IR/Passes/IRValidation.cpp
   Interface/IR/Passes/RedundantFlagCalculationElimination.cpp

FEXCore/Source/Common/JitSymbols.h

@@ -2,12 +2,10 @@
 #pragma once
 
 #include <FEXCore/fextl/memory.h>
-#include <FEXCore/Debug/InternalThreadState.h>
 
 #include <chrono>
+#include <cstddef>
 #include <cstdint>
-#include <cstdio>
-#include <memory>
 #include <string_view>
 
 namespace FEXCore {

FEXCore/Source/Common/SoftFloat.h

@@ -294,7 +294,11 @@ struct FEX_PACKED X80SoftFloat {
   FCMP(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs, bool* eq, bool* lt, bool* nan) {
     *eq = extF80_eq(state, lhs, rhs);
     *lt = extF80_lt(state, lhs, rhs);
-    *nan = IsNan(lhs) || IsNan(rhs);
+
+    // Use IEEE 754 semantics: unordered if neither <, =, nor > is true
+    // This is more reliable than custom NaN detection
+    bool gt = !(*eq) && !(*lt) && extF80_le(state, rhs, lhs);
+    *nan = !(*eq) && !(*lt) && !gt;
   }
 
   FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FSCALE(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {

FEXCore/Source/Common/StringConv.h

@@ -7,69 +7,58 @@
 #include <optional>
 
 namespace FEXCore::StrConv {
-[[maybe_unused]]
-static bool Conv(std::string_view Value, bool* Result) {
+inline bool Conv(std::string_view Value, bool* Result) {
   *Result = std::strtoull(Value.data(), nullptr, 0);
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, uint8_t* Result) {
+inline bool Conv(std::string_view Value, uint8_t* Result) {
   *Result = std::strtoul(Value.data(), nullptr, 0);
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, int8_t* Result) {
+inline bool Conv(std::string_view Value, int8_t* Result) {
   *Result = std::strtol(Value.data(), nullptr, 0);
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, uint16_t* Result) {
+inline bool Conv(std::string_view Value, uint16_t* Result) {
   *Result = std::strtoul(Value.data(), nullptr, 0);
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, int16_t* Result) {
+inline bool Conv(std::string_view Value, int16_t* Result) {
   *Result = std::strtol(Value.data(), nullptr, 0);
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, uint32_t* Result) {
+inline bool Conv(std::string_view Value, uint32_t* Result) {
   *Result = std::strtoul(Value.data(), nullptr, 0);
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, int32_t* Result) {
+inline bool Conv(std::string_view Value, int32_t* Result) {
   *Result = std::strtol(Value.data(), nullptr, 0);
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, uint64_t* Result) {
+inline bool Conv(std::string_view Value, uint64_t* Result) {
   *Result = std::strtoull(Value.data(), nullptr, 0);
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, int64_t* Result) {
+inline bool Conv(std::string_view Value, int64_t* Result) {
   *Result = std::strtoll(Value.data(), nullptr, 0);
   return true;
 }
 
 template<typename T, typename = std::enable_if<std::is_enum<T>::value, T>>
-[[maybe_unused]]
-static bool Conv(std::string_view Value, T* Result) {
+inline bool Conv(std::string_view Value, T* Result) {
   *Result = static_cast<T>(std::stoull(Value.data(), nullptr, 0));
   return true;
 }
 
-[[maybe_unused]]
-static bool Conv(std::string_view Value, fextl::string* Result) {
+inline bool Conv(std::string_view Value, fextl::string* Result) {
   *Result = Value;
   return true;
 }

FEXCore/Source/Interface/Config/Config.json.in

@@ -18,17 +18,6 @@
           "Maximum number of instruction to store in a block"
         ]
       },
-      "CacheObjectCodeCompilation": {
-        "Type": "uint32",
-        "Default": "FEXCore::Config::ConfigObjectCodeHandler::CONFIG_NONE",
-        "TextDefault": "none",
-        "Choices": [ "none", "read", "readwrite" ],
-        "ArgumentHandler": "CacheObjectCodeHandler",
-        "Desc": [
-          "Cache JIT object code to drive.",
-          "Allows JIT code to be shared between applications"
-        ]
-      },
       "HostFeatures": {
         "Type": "strenum",
         "Default": "FEXCore::Config::HostFeatures::OFF",
@@ -70,7 +59,9 @@
           "ENABLEPRESERVEALLABI": "enablepreserveallabi",
           "DISABLEPRESERVEALLABI": "disablepreserveallabi",
           "ENABLEWFXT": "enablewfxt",
-          "DISABLEWFXT": "disablewfxt"
+          "DISABLEWFXT": "disablewfxt",
+          "ENABLE3DNOW": "enable3dnow",
+          "DISABLE3DNOW": "disable3dnow"
         },
         "Desc": [
           "Allows controlling of the CPU features in the JIT.",
@@ -92,7 +83,8 @@
           "\t{enable,disable}rpres: Will force enable or disable rpres even if the host doesn't support it",
           "\t{enable,disable}svebitperm: Will force enable or disable svebitperm even if the host doesn't support it",
           "\t{enable,disable}preserveallabi: Will force enable or disable preserve_all abi even if the host doesn't support it",
-          "\t{enable,disable}wfxt: Will force enable or disable wfxt even if the host doesn't support it"
+          "\t{enable,disable}wfxt: Will force enable or disable wfxt even if the host doesn't support it",
+          "\t{enable,disable}3dnow: Will force enable or disable 3DNow even if the host doesn't support it"
         ]
       },
       "SmallTSCScale": {
@@ -473,32 +465,16 @@
         "Desc": [
           "Contrains the startup sleep to only apply to processes that match this name."
         ]
+      },
+      "MonoHacks": {
+        "Type": "bool",
+        "Default": "true",
+        "Desc": [
+          "Permits a hook-based SMC approach and smaller JIT blocks when mono is detected."
+        ]
       }
     },
     "Misc": {
-      "AOTIRCapture": {
-        "Type": "bool",
-        "Default": "false",
-        "Desc": [
-          "Captures IR and generates an AOT IR cache.",
-          "Captures both the loaded executable and libraries it loads."
-        ]
-      },
-      "AOTIRGenerate": {
-        "Type": "bool",
-        "Default": "false",
-        "Desc": [
-          "Scans file for executable code and generates an AOT IR cache.",
-          "Does not run the executable."
-        ]
-      },
-      "AOTIRLoad": {
-        "Type": "bool",
-        "Default": "false",
-        "Desc": [
-          "Loads an AOT IR cache for the loaded executable."
-        ]
-      },
       "ServerSocketPath": {
         "Type": "str",
         "Default": "",
@@ -512,6 +488,27 @@
         "Desc": [
           "Disables inline syscalls in order to support seccomp handling"
         ]
+      },
+      "ExtendedVolatileMetadata": {
+        "Type": "str",
+        "Default": "",
+        "Desc": [
+          "Configuration provided volatile metadata. Only implemented for WoW64/arm64ec.",
+          "Limited in its use but can be handy.",
+          "Extends on top of what Microsoft has for volatile metadata, but also supported for WoW64.",
+          "Colon delimited modules, then semi-colon delimited instructions, then comma delimited ranges",
+          "Default disables TSO in the module, unless instructions overlap the range",
+          "<module>;<offset begin>-<offset-end>,...;<instruction offset to force TSO>,...:<another>",
+          "examples:",
+          "  * Disable TSO for a full module: Just provide the module name:",
+          "      `hl2_linux`",
+          "  * Disable TSO for a part of the module:",
+          "      `hl2_linux;<offset begin>-<offset-end>`",
+          "  * Disable TSO for a part of the module, but enable TSO for some instructions within the module",
+          "      `hl2_linux;<offset begin>-<offset-end>;<instruction offset>,<instruction offset>`",
+          "  * Disable TSO for multiple modules",
+          "      `hl2_linux:libsdl2.so`"
+        ]
       }
     }
   },

FEXCore/Source/Interface/Context/Context.cpp

@@ -8,18 +8,12 @@
 #include <FEXCore/Core/CPUID.h>
 #include <FEXCore/Core/HostFeatures.h>
 #include <FEXCore/Core/SignalDelegator.h>
+#include <FEXCore/HLE/SyscallHandler.h>
+
 #include <FEXCore/Core/Thunks.h>
 #include "FEXCore/Debug/InternalThreadState.h"
 
-#include <string.h>
-#include <utility>
-
 namespace FEXCore::Context {
-void InitializeStaticTables(OperatingMode Mode) {
-  X86Tables::InitializeInfoTables(Mode);
-  IR::InstallOpcodeHandlers(Mode);
-}
-
 fextl::unique_ptr<FEXCore::Context::Context> FEXCore::Context::Context::CreateNewContext(const FEXCore::HostFeatures& Features) {
   return fextl::make_unique<FEXCore::Context::ContextImpl>(Features);
 }

FEXCore/Source/Interface/Context/Context.h

@@ -2,11 +2,12 @@
 #pragma once
 
 #include "Common/JitSymbols.h"
+#include "Interface/Core/CPUBackend.h"
 #include "Interface/Core/CPUID.h"
 #include "Interface/Core/X86HelperGen.h"
-#include "Interface/Core/ObjectCache/ObjectCacheService.h"
 #include "Interface/Core/Dispatcher/Dispatcher.h"
 #include "Interface/IR/AOTIR.h"
+#include <Interface/IR/IntrusiveIRList.h>
 #include <FEXCore/Config/Config.h>
 #include <FEXCore/Core/Context.h>
 #include <FEXCore/Core/CoreState.h>
@@ -33,10 +34,6 @@ namespace FEXCore {
 class CodeLoader;
 class ThunkHandler;
 
-namespace CodeSerialize {
-  class CodeObjectSerializeService;
-}
-
 namespace CPU {
   class Arm64JITCore;
   class Dispatcher;
@@ -50,8 +47,6 @@ namespace HLE {
 } // namespace FEXCore
 
 namespace FEXCore::IR {
-struct IRListCopy;
-class IRListView;
 namespace Validation {
   class IRValidation;
 }
@@ -90,6 +85,7 @@ public:
 
   bool IsAddressInCurrentBlock(FEXCore::Core::InternalThreadState* Thread, uint64_t Address, uint64_t Size) override;
   bool IsCurrentBlockSingleInst(FEXCore::Core::InternalThreadState* Thread) override;
+  uint64_t GetGuestBlockEntry(FEXCore::Core::InternalThreadState* Thread) override;
 
   uint64_t RestoreRIPFromHostPC(FEXCore::Core::InternalThreadState* Thread, uint64_t HostPC) override;
   uint32_t ReconstructCompactedEFLAGS(FEXCore::Core::InternalThreadState* Thread, bool WasInJIT, const uint64_t* HostGPRs, uint64_t PSTATE) override;
@@ -148,22 +144,7 @@ public:
   FEXCore::IR::AOTIRCacheEntry* LoadAOTIRCacheEntry(const fextl::string& Name) override;
   void UnloadAOTIRCacheEntry(FEXCore::IR::AOTIRCacheEntry* Entry) override;
 
-  void SetAOTIRLoader(AOTIRLoaderCBFn CacheReader) override {
-    IRCaptureCache.SetAOTIRLoader(std::move(CacheReader));
-  }
-  void SetAOTIRWriter(AOTIRWriterCBFn CacheWriter) override {
-    IRCaptureCache.SetAOTIRWriter(std::move(CacheWriter));
-  }
-  void SetAOTIRRenamer(AOTIRRenamerCBFn CacheRenamer) override {
-    IRCaptureCache.SetAOTIRRenamer(std::move(CacheRenamer));
-  }
-
-  void FinalizeAOTIRCache() override {
-    IRCaptureCache.FinalizeAOTIRCache();
-  }
-  void WriteFilesWithCode(AOTIRCodeFileWriterFn Writer) override {
-    IRCaptureCache.WriteFilesWithCode(Writer);
-  }
+  void FinalizeAOTIRCache() override {}
 
   void OnCodeBufferAllocated(CPU::CodeBuffer&) override;
   void ClearCodeCache(FEXCore::Core::InternalThreadState* Thread, bool NewCodeBuffer = true) override;
@@ -189,6 +170,12 @@ public:
 
   void RemoveForceTSOInformation(uint64_t Address, uint64_t Size) override;
 
+  void MarkMonoDetected() override {
+    MonoDetected = true;
+  }
+
+  void MarkMonoBackpatcherBlock(uint64_t BlockEntry) override;
+
 public:
   friend class FEXCore::HLE::SyscallHandler;
 #ifdef JIT_ARM64
@@ -213,9 +200,6 @@ public:
     FEX_CONFIG_OPT(VectorTSOEnabled, VECTORTSOENABLED);
     FEX_CONFIG_OPT(MemcpySetTSOEnabled, MEMCPYSETTSOENABLED);
     FEX_CONFIG_OPT(ABILocalFlags, ABILOCALFLAGS);
-    FEX_CONFIG_OPT(AOTIRCapture, AOTIRCAPTURE);
-    FEX_CONFIG_OPT(AOTIRGenerate, AOTIRGENERATE);
-    FEX_CONFIG_OPT(AOTIRLoad, AOTIRLOAD);
     FEX_CONFIG_OPT(SMCChecks, SMCCHECKS);
     FEX_CONFIG_OPT(MaxInstPerBlock, MAXINST);
     FEX_CONFIG_OPT(RootFSPath, ROOTFS);
@@ -224,12 +208,12 @@ public:
     FEX_CONFIG_OPT(BlockJITNaming, BLOCKJITNAMING);
     FEX_CONFIG_OPT(GDBSymbols, GDBSYMBOLS);
     FEX_CONFIG_OPT(ParanoidTSO, PARANOIDTSO);
-    FEX_CONFIG_OPT(CacheObjectCodeCompilation, CACHEOBJECTCODECOMPILATION);
     FEX_CONFIG_OPT(x87ReducedPrecision, X87REDUCEDPRECISION);
     FEX_CONFIG_OPT(DisableTelemetry, DISABLETELEMETRY);
     FEX_CONFIG_OPT(DisableVixlIndirectCalls, DISABLE_VIXL_INDIRECT_RUNTIME_CALLS);
     FEX_CONFIG_OPT(SmallTSCScale, SMALLTSCSCALE);
     FEX_CONFIG_OPT(StrictInProcessSplitLocks, STRICTINPROCESSSPLITLOCKS);
+    FEX_CONFIG_OPT(MonoHacks, MONOHACKS);
   } Config;
 
   FEXCore::ForkableSharedMutex CodeInvalidationMutex;
@@ -248,24 +232,17 @@ public:
   X86GeneratedCode X86CodeGen;
 
   ContextImpl(const FEXCore::HostFeatures& Features);
-  ~ContextImpl();
 
   static bool ThreadRemoveCodeEntry(FEXCore::Core::InternalThreadState* Thread, uint64_t GuestRIP);
 
-  // Wrapper which takes CpuStateFrame instead of InternalThreadState and unique_locks CodeInvalidationMutex
-  // Must be called from owning thread
-  static void ThreadRemoveCodeEntryFromJit(FEXCore::Core::CpuStateFrame* Frame, uint64_t GuestRIP) {
-    auto Thread = Frame->Thread;
-    auto lk = GuardSignalDeferringSection(static_cast<ContextImpl*>(Thread->CTX)->CodeInvalidationMutex, Thread);
+  static void ThreadRemoveCodeEntryFromJit(FEXCore::Core::CpuStateFrame* Frame, uint64_t GuestRIP);
 
-    // NOTE: Other threads sharing the same CodeBuffer may reference
-    //       invalidated data ranges through their L1/L2 caches. This is
-    //       not currently a problem since FEX does not repurpose the
-    //       invalidated CodeBuffer memory range currently.
-    ThreadRemoveCodeEntry(Thread, GuestRIP);
-  }
+  // This is used as a replacement for the SMC writes in the mono callsite backpatcher that avoids atomic operations
+  // (safe as the invalidation mutex is locked) and manually invalidates the modified range. Allowing SMC to be detected
+  // even if faulting is disabled.
+  static void MonoBackpatcherWrite(FEXCore::Core::CpuStateFrame* Frame, uint8_t Size, uint64_t Address, uint64_t Value);
 
-  void RemoveCustomIREntrypoint(uintptr_t Entrypoint);
+  void RemoveCustomIREntrypoint(FEXCore::Core::InternalThreadState* Thread, uintptr_t Entrypoint);
 
   struct GenerateIRResult {
     std::optional<IR::IRListView> IRView;
@@ -324,6 +301,10 @@ public:
     return ExitOnHLT;
   }
 
+  bool AreMonoHacksActive() const {
+    return Config.MonoHacks && MonoDetected;
+  }
+
 protected:
   void UpdateAtomicTSOEmulationConfig() {
     if (SupportsHardwareTSO) {
@@ -356,7 +337,6 @@ private:
   void InitializeCompiler(FEXCore::Core::InternalThreadState* Thread);
 
   IR::AOTIRCaptureCache IRCaptureCache;
-  fextl::unique_ptr<FEXCore::CodeSerialize::CodeObjectSerializeService> CodeObjectCacheService;
 
   bool IsMemoryShared = false;
   bool SupportsHardwareTSO = false;
@@ -377,5 +357,8 @@ private:
   fextl::unordered_map<uint64_t, CustomIRHandlerEntry> CustomIRHandlers;
   IntervalList<uint64_t> ForceTSOValidRanges; // The ranges for which ForceTSOInstructions has populated data
   fextl::set<uint64_t> ForceTSOInstructions;
+
+  bool MonoDetected = false;
+  std::atomic<uint64_t> MonoBackpatcherBlock;
 };
 } // namespace FEXCore::Context

FEXCore/Source/Interface/Core/Addressing.cpp

@@ -11,8 +11,7 @@ Ref LoadEffectiveAddress(IREmitter* IREmit, AddressMode A, IR::OpSize GPRSize, b
   Ref Tmp = A.Base;
 
   if (A.Offset) {
-    Ref Offset = IREmit->Constant(A.Offset);
-    Tmp = Tmp ? IREmit->_Add(GPRSize, Tmp, Offset) : Offset;
+    Tmp = Tmp ? IREmit->Add(GPRSize, Tmp, A.Offset) : IREmit->Constant(A.Offset);
   }
 
   if (A.Index) {
@@ -25,7 +24,7 @@ Ref LoadEffectiveAddress(IREmitter* IREmit, AddressMode A, IR::OpSize GPRSize, b
         Tmp = IREmit->_Lshl(GPRSize, A.Index, IREmit->Constant(Log2));
       }
     } else {
-      Tmp = Tmp ? IREmit->_Add(GPRSize, Tmp, A.Index) : A.Index;
+      Tmp = Tmp ? IREmit->Add(GPRSize, Tmp, A.Index) : A.Index;
     }
   }
 
@@ -46,7 +45,7 @@ Ref LoadEffectiveAddress(IREmitter* IREmit, AddressMode A, IR::OpSize GPRSize, b
   }
 
   if (A.Segment && AddSegmentBase) {
-    Tmp = Tmp ? IREmit->_Add(GPRSize, Tmp, A.Segment) : A.Segment;
+    Tmp = Tmp ? IREmit->Add(GPRSize, Tmp, A.Segment) : A.Segment;
   }
 
   return Tmp ?: IREmit->Constant(0);
@@ -54,19 +53,15 @@ Ref LoadEffectiveAddress(IREmitter* IREmit, AddressMode A, IR::OpSize GPRSize, b
 
 AddressMode SelectAddressMode(IREmitter* IREmit, AddressMode A, IR::OpSize GPRSize, bool HostSupportsTSOImm9, bool AtomicTSO, bool Vector,
                               IR::OpSize AccessSize) {
-  auto SoftwareAddressCalculation = [IREmit, &A, GPRSize]() -> AddressMode {
-    return {
-      .Base = LoadEffectiveAddress(IREmit, A, GPRSize, true),
-      .Index = IREmit->Invalid(),
-    };
-  };
-
   const auto Is32Bit = GPRSize == OpSize::i32Bit;
   const auto GPRSizeMatchesAddrSize = A.AddrSize == GPRSize;
   const auto OffsetIndexToLargeFor32Bit = Is32Bit && (A.Offset <= -16384 || A.Offset >= 16384);
   if (!GPRSizeMatchesAddrSize || OffsetIndexToLargeFor32Bit) {
     // If address size doesn't match GPR size then no optimizations can occur.
-    return SoftwareAddressCalculation();
+    return {
+      .Base = LoadEffectiveAddress(IREmit, A, GPRSize, true),
+      .Index = IREmit->Invalid(),
+    };
   }
 
   // Loadstore rules:
@@ -100,7 +95,7 @@ AddressMode SelectAddressMode(IREmitter* IREmit, AddressMode A, IR::OpSize GPRSi
   const bool OffsetIsSIMM9 = A.Offset && A.Offset >= -256 && A.Offset <= 255;
   const bool OffsetIsUnsignedScaled = A.Offset > 0 && (A.Offset & (AccessSizeAsImm - 1)) == 0 && (A.Offset / AccessSizeAsImm) <= 4095;
 
-  auto InlineImmOffsetLoadstore = [IREmit, &GPRSize](AddressMode A) -> AddressMode {
+  if ((AtomicTSO && !Vector && HostSupportsTSOImm9 && OffsetIsSIMM9) || (!AtomicTSO && (OffsetIsSIMM9 || OffsetIsUnsignedScaled))) {
     // Peel off the offset
     AddressMode B = A;
     B.Offset = 0;
@@ -111,32 +106,20 @@ AddressMode SelectAddressMode(IREmitter* IREmit, AddressMode A, IR::OpSize GPRSi
       .IndexType = MEM_OFFSET_SXTX,
       .IndexScale = 1,
     };
-  };
+  }
 
-  auto ScaledRegisterLoadstore = [IREmit, GPRSize](AddressMode A) -> AddressMode {
+  if (AtomicTSO) {
+    // TODO: LRCPC3 support for vector Imm9.
+  } else if (!Is32Bit && A.Base && (A.Index || A.Segment) && !A.Offset && (A.IndexScale == 1 || A.IndexScale == AccessSizeAsImm)) {
+    // ScaledRegisterLoadstore
     if (A.Index && A.Segment) {
-      A.Base = IREmit->_Add(GPRSize, A.Base, A.Segment);
+      A.Base = IREmit->Add(GPRSize, A.Base, A.Segment);
     } else if (A.Segment) {
       A.Index = A.Segment;
       A.IndexScale = 1;
     }
-    return A;
-  };
 
-  if (AtomicTSO) {
-    if (!Vector) {
-      if (HostSupportsTSOImm9 && OffsetIsSIMM9) {
-        return InlineImmOffsetLoadstore(A);
-      }
-    } else {
-      // TODO: LRCPC3 support for vector Imm9.
-    }
-  } else {
-    if (OffsetIsSIMM9 || OffsetIsUnsignedScaled) {
-      return InlineImmOffsetLoadstore(A);
-    } else if (!Is32Bit && A.Base && (A.Index || A.Segment) && !A.Offset && (A.IndexScale == 1 || A.IndexScale == AccessSizeAsImm)) {
-      return ScaledRegisterLoadstore(A);
-    }
+    return A;
   }
 
   if (Vector || !AtomicTSO) {
@@ -159,7 +142,10 @@ AddressMode SelectAddressMode(IREmitter* IREmit, AddressMode A, IR::OpSize GPRSi
   }
 
   // Fallback on software address calculation
-  return SoftwareAddressCalculation();
+  return {
+    .Base = LoadEffectiveAddress(IREmit, A, GPRSize, true),
+    .Index = IREmit->Invalid(),
+  };
 }
 
 

FEXCore/Source/Interface/Core/ArchHelpers/Arm64Emitter.cpp

@@ -712,7 +712,7 @@ void Arm64Emitter::SpillStaticRegs(ARMEmitter::Register TmpReg, bool FPRs, uint3
   // Now handle PF/AF
   if (PFAFSpillMask) {
     auto PFOffset = offsetof(FEXCore::Core::CpuStateFrame, State.pf_raw);
-    [[maybe_unused]] auto AFOffset = offsetof(FEXCore::Core::CpuStateFrame, State.af_raw);
+    auto AFOffset = offsetof(FEXCore::Core::CpuStateFrame, State.af_raw);
     LOGMAN_THROW_A_FMT(PFAFSpillMask == PFAFMask, "PF/AF not spilled together");
     LOGMAN_THROW_A_FMT(AFOffset == PFOffset + 4, "PF/AF are together");
 

FEXCore/Source/Interface/Core/ArchHelpers/Arm64Emitter.h

@@ -2,7 +2,7 @@
 #pragma once
 
 #include "FEXCore/Utils/EnumUtils.h"
-#include "Interface/Core/ObjectCache/Relocations.h"
+#include "Interface/Core/JIT/Relocations.h"
 
 #ifdef VIXL_DISASSEMBLER
 #include <aarch64/disasm-aarch64.h>

FEXCore/Source/Interface/Core/CPUBackend.cpp

@@ -317,7 +317,7 @@ namespace CPU {
     // Resize the code buffer and reallocate our code size
     CurrentCodeBuffer = CodeBuffers.StartLargerCodeBuffer();
 
-    RegisterForSignalHandler(PrevCodeBuffer);
+    RegisterForSignalHandler(std::move(PrevCodeBuffer));
     return CurrentCodeBuffer.get();
   }
 
@@ -326,14 +326,13 @@ namespace CPU {
       // We have signal handlers that have generated code
       // This means that we can not safely clear the code at this point in time
       // Keep a reference to the old code buffer to delay deallocation
-      SignalHandlerCodeBuffers.push_back(CodeBuffer);
+      SignalHandlerCodeBuffers.push_back(std::move(CodeBuffer));
     } else {
       SignalHandlerCodeBuffers.clear();
     }
   }
 
   fextl::shared_ptr<CodeBuffer> CPUBackend::CheckCodeBufferUpdate() {
-    fextl::shared_ptr<CodeBuffer> OldCodeBuffer;
     auto NewCodeBuffer = CodeBuffers.GetLatest();
     if (CurrentCodeBuffer != NewCodeBuffer) {
       RegisterForSignalHandler(CurrentCodeBuffer);

FEXCore/Source/Interface/Core/CPUID.cpp

@@ -925,38 +925,38 @@ FEXCore::CPUID::FunctionResults CPUIDEmu::Function_8000_0001h(uint32_t Leaf) con
             (0 << 30) | // Reserved
             (0 << 31);  // Reserved
 
-  Res.edx = (1 << 0) |                // FPU
-            (1 << 1) |                // Virtual mode extensions
-            (1 << 2) |                // Debugging extensions
-            (1 << 3) |                // Page size extensions
-            (1 << 4) |                // TSC
-            (1 << 5) |                // MSR support
-            (1 << 6) |                // PAE
-            (1 << 7) |                // Machine Check Exception
-            (1 << 8) |                // CMPXCHG8B
-            (1 << 9) |                // APIC
-            (0 << 10) |               // Reserved
-            (1 << 11) |               // SYSCALL/SYSRET
-            (1 << 12) |               // MTRR
-            (1 << 13) |               // Page global extension
-            (1 << 14) |               // Machine Check architecture
-            (1 << 15) |               // CMOV
-            (1 << 16) |               // Page attribute table
-            (1 << 17) |               // Page-size extensions
-            (0 << 18) |               // Reserved
-            (0 << 19) |               // Reserved
-            (1 << 20) |               // NX
-            (0 << 21) |               // Reserved
-            (1 << 22) |               // MMXExt
-            (1 << 23) |               // MMX
-            (1 << 24) |               // FXSAVE/FXRSTOR
-            (1 << 25) |               // FXSAVE/FXRSTOR Optimizations
-            (0 << 26) |               // 1 gigabit pages
-            (SUPPORTS_RDTSCP << 27) | // RDTSCP
-            (0 << 28) |               // Reserved
-            (1 << 29) |               // Long Mode
-            (1 << 30) |               // 3DNow! Extensions
-            (1 << 31);                // 3DNow!
+  Res.edx = (1 << 0) |                                // FPU
+            (1 << 1) |                                // Virtual mode extensions
+            (1 << 2) |                                // Debugging extensions
+            (1 << 3) |                                // Page size extensions
+            (1 << 4) |                                // TSC
+            (1 << 5) |                                // MSR support
+            (1 << 6) |                                // PAE
+            (1 << 7) |                                // Machine Check Exception
+            (1 << 8) |                                // CMPXCHG8B
+            (1 << 9) |                                // APIC
+            (0 << 10) |                               // Reserved
+            (1 << 11) |                               // SYSCALL/SYSRET
+            (1 << 12) |                               // MTRR
+            (1 << 13) |                               // Page global extension
+            (1 << 14) |                               // Machine Check architecture
+            (1 << 15) |                               // CMOV
+            (1 << 16) |                               // Page attribute table
+            (1 << 17) |                               // Page-size extensions
+            (0 << 18) |                               // Reserved
+            (0 << 19) |                               // Reserved
+            (1 << 20) |                               // NX
+            (0 << 21) |                               // Reserved
+            (1 << 22) |                               // MMXExt
+            (1 << 23) |                               // MMX
+            (1 << 24) |                               // FXSAVE/FXRSTOR
+            (1 << 25) |                               // FXSAVE/FXRSTOR Optimizations
+            (0 << 26) |                               // 1 gigabit pages
+            (SUPPORTS_RDTSCP << 27) |                 // RDTSCP
+            (0 << 28) |                               // Reserved
+            (1 << 29) |                               // Long Mode
+            (CTX->HostFeatures.Supports3DNow << 30) | // 3DNow! Extensions
+            (CTX->HostFeatures.Supports3DNow << 31);  // 3DNow!
   return Res;
 }
 

FEXCore/Source/Interface/Core/Core.cpp

@@ -14,7 +14,6 @@ $end_info$
 #include "Interface/Core/CPUBackend.h"
 #include "Interface/Core/CPUID.h"
 #include "Interface/Core/Frontend.h"
-#include "Interface/Core/ObjectCache/ObjectCacheService.h"
 #include "Interface/Core/OpcodeDispatcher.h"
 #include "Interface/Core/JIT/JITClass.h"
 #include "Interface/Core/Dispatcher/Dispatcher.h"
@@ -79,9 +78,6 @@ ContextImpl::ContextImpl(const FEXCore::HostFeatures& Features)
   : HostFeatures {Features}
   , CPUID {this}
   , IRCaptureCache {this} {
-  if (Config.CacheObjectCodeCompilation() != FEXCore::Config::ConfigObjectCodeHandler::CONFIG_NONE) {
-    CodeObjectCacheService = fextl::make_unique<FEXCore::CodeSerialize::CodeObjectSerializeService>(this);
-  }
   if (!Config.Is64BitMode()) {
     // When operating in 32-bit mode, the virtual memory we care about is only the lower 32-bits.
     Config.VirtualMemSize = 1ULL << 32;
@@ -105,14 +101,6 @@ ContextImpl::ContextImpl(const FEXCore::HostFeatures& Features)
   UpdateAtomicTSOEmulationConfig();
 }
 
-ContextImpl::~ContextImpl() {
-  {
-    if (CodeObjectCacheService) {
-      CodeObjectCacheService->Shutdown();
-    }
-  }
-}
-
 struct GetFrameBlockInfoResult {
   const CPU::CPUBackend::JITCodeHeader* InlineHeader;
   const CPU::CPUBackend::JITCodeTail* InlineTail;
@@ -139,6 +127,11 @@ bool ContextImpl::IsCurrentBlockSingleInst(FEXCore::Core::InternalThreadState* T
   return InlineTail && InlineTail->SingleInst;
 }
 
+uint64_t ContextImpl::GetGuestBlockEntry(FEXCore::Core::InternalThreadState* Thread) {
+  auto [_, InlineTail] = GetFrameBlockInfo(Thread->CurrentFrame);
+  return InlineTail ? InlineTail->RIP : 0;
+}
+
 uint64_t ContextImpl::RestoreRIPFromHostPC(FEXCore::Core::InternalThreadState* Thread, uint64_t HostPC) {
   const auto Frame = Thread->CurrentFrame;
   const uint64_t BlockBegin = Frame->State.InlineJITBlockHeader;
@@ -349,36 +342,9 @@ bool ContextImpl::InitCore() {
   Dispatcher = FEXCore::CPU::Dispatcher::Create(this);
 
   // Set up the SignalDelegator config since core is initialized.
-  FEXCore::SignalDelegator::SignalDelegatorConfig SignalConfig {
-    .DispatcherBegin = Dispatcher->Start,
-    .DispatcherEnd = Dispatcher->End,
+  SignalDelegation->SetConfig(Dispatcher->MakeSignalDelegatorConfig());
 
-    .AbsoluteLoopTopAddress = Dispatcher->AbsoluteLoopTopAddress,
-    .AbsoluteLoopTopAddressFillSRA = Dispatcher->AbsoluteLoopTopAddressFillSRA,
-    .SignalHandlerReturnAddress = Dispatcher->SignalHandlerReturnAddress,
-    .SignalHandlerReturnAddressRT = Dispatcher->SignalHandlerReturnAddressRT,
-
-    .PauseReturnInstruction = Dispatcher->PauseReturnInstruction,
-    .ThreadPauseHandlerAddressSpillSRA = Dispatcher->ThreadPauseHandlerAddressSpillSRA,
-    .ThreadPauseHandlerAddress = Dispatcher->ThreadPauseHandlerAddress,
-
-    // Stop handlers.
-    .ThreadStopHandlerAddressSpillSRA = Dispatcher->ThreadStopHandlerAddressSpillSRA,
-    .ThreadStopHandlerAddress = Dispatcher->ThreadStopHandlerAddress,
-
-    // SRA information.
-    .SRAGPRCount = Dispatcher->GetSRAGPRCount(),
-    .SRAFPRCount = Dispatcher->GetSRAFPRCount(),
-  };
-
-  Dispatcher->GetSRAGPRMapping(SignalConfig.SRAGPRMapping);
-  Dispatcher->GetSRAFPRMapping(SignalConfig.SRAFPRMapping);
-
-  // Give this configuration to the SignalDelegator.
-  SignalDelegation->SetConfig(SignalConfig);
-
-#ifndef _WIN32
-#elif !defined(_M_ARM_64EC)
+#if defined(_WIN32) && !defined(_M_ARM_64EC)
   // WOW64 always needs the interrupt fault check to be enabled.
   Config.NeedsPendingInterruptFaultCheck = true;
 #endif
@@ -398,12 +364,6 @@ void ContextImpl::HandleCallback(FEXCore::Core::InternalThreadState* Thread, uin
 void ContextImpl::ExecuteThread(FEXCore::Core::InternalThreadState* Thread) {
   Dispatcher->ExecuteDispatch(Thread->CurrentFrame);
 
-  if (CodeObjectCacheService) {
-    // Ensure the Code Object Serialization service has fully serialized this thread's data before clearing the cache
-    // Use the thread's object cache ref counter for this
-    CodeSerialize::CodeObjectSerializeService::WaitForEmptyJobQueue(&Thread->ObjectCacheRefCounter);
-  }
-
   // If it is the parent thread that died then just leave
   // TODO: This doesn't make sense when the parent thread doesn't outlive its children
 }
@@ -441,22 +401,6 @@ ContextImpl::CreateThread(uint64_t InitialRIP, uint64_t StackPointer, const FEXC
   Thread->CurrentFrame->State.gregs[X86State::REG_RSP] = StackPointer;
   Thread->CurrentFrame->State.rip = InitialRIP;
 
-  // Set up default code segment.
-  // Default code segment indexes match the numbers that the Linux kernel uses.
-  Thread->CurrentFrame->State.cs_idx = 6 << 3;
-  auto &GDT = Thread->CurrentFrame->State.gdt[Thread->CurrentFrame->State.cs_idx >> 3];
-  Thread->CurrentFrame->State.SetGDTBase(&GDT, 0);
-  Thread->CurrentFrame->State.SetGDTLimit(&GDT, 0xF'FFFFU);
-
-  if (Config.Is64BitMode) {
-    GDT.L = 1; // L = Long Mode = 64-bit
-    GDT.D = 0; // D = Default Operand SIze = Reserved
-  }
-  else {
-    GDT.L = 0; // L = Long Mode = 32-bit
-    GDT.D = 1; // D = Default Operand Size = 32-bit
-  }
-
   // Copy over the new thread state to the new object
   if (NewThreadState) {
     memcpy(&Thread->CurrentFrame->State, NewThreadState, sizeof(FEXCore::Core::CPUState));
@@ -520,12 +464,6 @@ void ContextImpl::OnCodeBufferAllocated(CPU::CodeBuffer& Buffer) {
 void ContextImpl::ClearCodeCache(FEXCore::Core::InternalThreadState* Thread, bool NewCodeBuffer) {
   FEXCORE_PROFILE_INSTANT("ClearCodeCache");
 
-  if (CodeObjectCacheService) {
-    // Ensure the Code Object Serialization service has fully serialized this thread's data before clearing the cache
-    // Use the thread's object cache ref counter for this
-    CodeSerialize::CodeObjectSerializeService::WaitForEmptyJobQueue(&Thread->ObjectCacheRefCounter);
-  }
-
   if (NewCodeBuffer) {
     // Allocate new CodeBuffer + L3 LookupCache and clear L1+L2 caches
     Thread->CPUBackend->ClearCache();
@@ -579,7 +517,8 @@ ContextImpl::GenerateIR(FEXCore::Core::InternalThreadState* Thread, uint64_t Gue
     auto BlockInfo = Thread->FrontendDecoder->GetDecodedBlockInfo();
     auto CodeBlocks = &BlockInfo->Blocks;
 
-    Thread->OpDispatcher->BeginFunction(GuestRIP, CodeBlocks, BlockInfo->TotalInstructionCount, BlockInfo->Is64BitMode);
+    Thread->OpDispatcher->BeginFunction(GuestRIP, CodeBlocks, BlockInfo->TotalInstructionCount, BlockInfo->Is64BitMode,
+                                        AreMonoHacksActive() && MonoBackpatcherBlock.load(std::memory_order_relaxed) == GuestRIP);
 
     const auto GPRSize = Thread->OpDispatcher->GetGPROpSize();
 
@@ -607,7 +546,7 @@ ContextImpl::GenerateIR(FEXCore::Core::InternalThreadState* Thread, uint64_t Gue
 
       if (InstsInBlock == 0) {
         // Special case for an empty instruction block.
-        Thread->OpDispatcher->ExitFunction(Thread->OpDispatcher->_EntrypointOffset(GPRSize, Block.Entry - GuestRIP));
+        Thread->OpDispatcher->ExitFunction(Thread->OpDispatcher->_InlineEntrypointOffset(GPRSize, Block.Entry - GuestRIP));
       }
 
       for (size_t i = 0; i < InstsInBlock; ++i) {
@@ -637,11 +576,12 @@ ContextImpl::GenerateIR(FEXCore::Core::InternalThreadState* Thread, uint64_t Gue
           Thread->OpDispatcher->_GuestOpcode(InstAddress - GuestRIP);
         }
 
-        if (Config.SMCChecks == FEXCore::Config::CONFIG_SMC_FULL) {
-          auto ExistingCodePtr = reinterpret_cast<uint64_t*>(Block.Entry + BlockInstructionsLength);
-
-          auto CodeChanged = Thread->OpDispatcher->_ValidateCode(ExistingCodePtr[0], ExistingCodePtr[1],
-                                                                 (uintptr_t)ExistingCodePtr - GuestRIP, DecodedInfo->InstSize);
+        if (Config.SMCChecks == FEXCore::Config::CONFIG_SMC_FULL || Block.ForceFullSMCDetection) {
+          auto ExistingCodePtr = reinterpret_cast<uint8_t*>(Block.Entry + BlockInstructionsLength);
+          auto InstAddressReg = Thread->OpDispatcher->_EntrypointOffset(GPRSize, InstAddress - GuestRIP);
+          std::array<uint8_t, 0x10> CodeOriginal;
+          memcpy(CodeOriginal.data(), ExistingCodePtr, DecodedInfo->InstSize);
+          auto CodeChanged = Thread->OpDispatcher->_ValidateCode(CodeOriginal, InstAddressReg, DecodedInfo->InstSize);
 
           auto InvalidateCodeCond = Thread->OpDispatcher->CondJump(CodeChanged);
 
@@ -651,7 +591,7 @@ ContextImpl::GenerateIR(FEXCore::Core::InternalThreadState* Thread, uint64_t Gue
 
           Thread->OpDispatcher->SetCurrentCodeBlock(CodeWasChangedBlock);
           Thread->OpDispatcher->_ThreadRemoveCodeEntry();
-          Thread->OpDispatcher->ExitFunction(Thread->OpDispatcher->_EntrypointOffset(GPRSize, InstAddress - GuestRIP));
+          Thread->OpDispatcher->ExitFunction(Thread->OpDispatcher->_InlineEntrypointOffset(GPRSize, InstAddress - GuestRIP));
 
           auto NextOpBlock = Thread->OpDispatcher->CreateNewCodeBlockAfter(CurrentBlock);
 
@@ -659,15 +599,21 @@ ContextImpl::GenerateIR(FEXCore::Core::InternalThreadState* Thread, uint64_t Gue
           Thread->OpDispatcher->SetCurrentCodeBlock(NextOpBlock);
         }
 
-        if (TableInfo && TableInfo->OpcodeDispatcher) {
-          auto Fn = TableInfo->OpcodeDispatcher;
+        if (TableInfo && TableInfo->OpcodeDispatcher.OpDispatch) {
+          auto Fn = TableInfo->OpcodeDispatcher.OpDispatch;
           Thread->OpDispatcher->ResetHandledLock();
           Thread->OpDispatcher->ResetDecodeFailure();
-          IR::ForceTSOMode ForceTSO =
-            BlockInForceTSOValidRange ?
-              (InstForceTSOIt != ForceTSOInstructions.end() && *InstForceTSOIt == InstAddress ? IR::ForceTSOMode::ForceEnabled :
-                                                                                                IR::ForceTSOMode::ForceDisabled) :
-              IR::ForceTSOMode::NoOverride;
+          IR::ForceTSOMode ForceTSO = IR::ForceTSOMode::NoOverride;
+          if (BlockInForceTSOValidRange) {
+            if (InstForceTSOIt != ForceTSOInstructions.end() && *InstForceTSOIt == InstAddress) {
+              ForceTSO = IR::ForceTSOMode::ForceEnabled;
+            } else {
+              ForceTSO = IR::ForceTSOMode::ForceDisabled;
+            }
+          } else if (DecodedInfo->ForceTSO) {
+            ForceTSO = IR::ForceTSOMode::ForceEnabled;
+          }
+
           Thread->OpDispatcher->SetForceTSO(ForceTSO);
           std::invoke(Fn, Thread->OpDispatcher, DecodedInfo);
           if (Thread->OpDispatcher->HadDecodeFailure()) {
@@ -717,7 +663,8 @@ ContextImpl::GenerateIR(FEXCore::Core::InternalThreadState* Thread, uint64_t Gue
 
         if (NeedsBlockEnd) {
           // We had some instructions. Early exit
-          Thread->OpDispatcher->ExitFunction(Thread->OpDispatcher->_EntrypointOffset(GPRSize, Block.Entry + BlockInstructionsLength - GuestRIP));
+          Thread->OpDispatcher->ExitFunction(
+            Thread->OpDispatcher->_InlineEntrypointOffset(GPRSize, Block.Entry + BlockInstructionsLength - GuestRIP));
           break;
         }
 
@@ -760,26 +707,9 @@ ContextImpl::GenerateIR(FEXCore::Core::InternalThreadState* Thread, uint64_t Gue
 }
 
 ContextImpl::CompileCodeResult ContextImpl::CompileCode(FEXCore::Core::InternalThreadState* Thread, uint64_t GuestRIP, uint64_t MaxInst) {
-  // JIT Code object cache lookup
-  if (CodeObjectCacheService) {
-    auto CodeCacheEntry = CodeObjectCacheService->FetchCodeObjectFromCache(GuestRIP);
-    if (CodeCacheEntry) {
-      auto CompiledCode = Thread->CPUBackend->RelocateJITObjectCode(GuestRIP, CodeCacheEntry);
-      if (CompiledCode) {
-        return {
-          .CompiledCode = {},
-          .DebugData = nullptr, // nullptr here ensures that code serialization doesn't occur on from cache read
-          .StartAddr = 0,       // Unused
-          .Length = 0,          // Unused
-          .NeedsAddGuestCodeRanges = false,
-        };
-      }
-    }
-  }
-
   if (SourcecodeResolver && Config.GDBSymbols()) {
     auto AOTIRCacheEntry = SyscallHandler->LookupAOTIRCacheEntry(Thread, GuestRIP);
-    if (AOTIRCacheEntry.Entry && !AOTIRCacheEntry.Entry->ContainsCode) {
+    if (AOTIRCacheEntry.Entry) {
       AOTIRCacheEntry.Entry->SourcecodeMap = SourcecodeResolver->GenerateMap(AOTIRCacheEntry.Entry->Filename, AOTIRCacheEntry.Entry->FileId);
     }
   }
@@ -879,25 +809,10 @@ uintptr_t ContextImpl::CompileBlock(FEXCore::Core::CpuStateFrame* Frame, uint64_
     }
   }
 
-  // Tell the object cache service to serialize the code if enabled
-  if (CodeObjectCacheService && Config.CacheObjectCodeCompilation == FEXCore::Config::ConfigObjectCodeHandler::CONFIG_READWRITE && DebugData) {
-    CodeObjectCacheService->AsyncAddSerializationJob(
-      fextl::make_unique<CodeSerialize::AsyncJobHandler::SerializationJobData>(CodeSerialize::AsyncJobHandler::SerializationJobData {
-        .GuestRIP = GuestRIP,
-        .GuestCodeLength = Length,
-        .GuestCodeHash = 0,
-        .HostCodeBegin = CompiledCode.BlockBegin,
-        .HostCodeLength = CompiledCode.Size,
-        .HostCodeHash = 0,
-        .ThreadJobRefCount = &Thread->ObjectCacheRefCounter,
-        .Relocations = std::move(*DebugData->Relocations),
-      }));
-  }
-
   // Clear any relocations that might have been generated
   Thread->CPUBackend->ClearRelocations();
 
-  if (IRCaptureCache.PostCompileCode(Thread, CompiledCode.BlockBegin, GuestRIP, StartAddr, Length, {}, DebugData.get(), false)) {
+  if (IRCaptureCache.PostCompileCode(Thread, CompiledCode.BlockBegin, GuestRIP, StartAddr, Length, DebugData.get())) {
     // Early exit
     return (uintptr_t)CodePtr;
   }
@@ -1002,6 +917,10 @@ bool ContextImpl::ThreadRemoveCodeEntry(FEXCore::Core::InternalThreadState* Thre
   return Thread->LookupCache->Erase(Thread->CurrentFrame, GuestRIP);
 }
 
+void ContextImpl::ThreadRemoveCodeEntryFromJit(FEXCore::Core::CpuStateFrame* Frame, uint64_t GuestRIP) {
+  static_cast<ContextImpl*>(Frame->Thread->CTX)->SyscallHandler->InvalidateGuestCodeRange(Frame->Thread, GuestRIP, 1);
+}
+
 std::optional<CustomIRResult>
 ContextImpl::AddCustomIREntrypoint(uintptr_t Entrypoint, CustomIREntrypointHandler Handler, void* Creator, void* Data) {
   LOGMAN_THROW_A_FMT(Config.Is64BitMode || !(Entrypoint >> 32), "64-bit Entrypoint in 32-bit mode {:x}", Entrypoint);
@@ -1075,16 +994,36 @@ void ContextImpl::RemoveForceTSOInformation(uint64_t Address, uint64_t Size) {
   ForceTSOInstructions.erase(ForceTSOInstructions.lower_bound(Address), ForceTSOInstructions.upper_bound(Address + Size));
 }
 
-void ContextImpl::RemoveCustomIREntrypoint(uintptr_t Entrypoint) {
+void ContextImpl::MarkMonoBackpatcherBlock(uint64_t BlockEntry) {
+  MonoBackpatcherBlock.store(BlockEntry, std::memory_order_relaxed);
+}
+
+void ContextImpl::RemoveCustomIREntrypoint(FEXCore::Core::InternalThreadState* Thread, uintptr_t Entrypoint) {
   LOGMAN_THROW_A_FMT(Config.Is64BitMode || !(Entrypoint >> 32), "64-bit Entrypoint in 32-bit mode {:x}", Entrypoint);
 
   std::scoped_lock lk(CustomIRMutex);
 
-  InvalidatedEntryAccumulator Accumulator;
-  InvalidateGuestCodeRange(nullptr, Accumulator, Entrypoint, 1);
   CustomIRHandlers.erase(Entrypoint);
-
   HasCustomIRHandlers = !CustomIRHandlers.empty();
+  SyscallHandler->InvalidateGuestCodeRange(Thread, Entrypoint, 1);
+}
+
+void ContextImpl::MonoBackpatcherWrite(FEXCore::Core::CpuStateFrame* Frame, uint8_t Size, uint64_t Address, uint64_t Value) {
+  auto Thread = Frame->Thread;
+  auto CTX = static_cast<ContextImpl*>(Thread->CTX);
+  {
+    auto lk = GuardSignalDeferringSection(CTX->CodeInvalidationMutex, Thread);
+
+    if (Size == 8) {
+      *reinterpret_cast<uint64_t *>(Address) = Value;
+    } else if (Size == 4) {
+      *reinterpret_cast<uint32_t *>(Address) = Value;
+    } else {
+      ERROR_AND_DIE_FMT("Unexpected write size for backpatcher: {}", Size);
+    }
+  }
+
+  CTX->SyscallHandler->InvalidateGuestCodeRange(Thread, Address, Size);
 }
 
 IR::AOTIRCacheEntry* ContextImpl::LoadAOTIRCacheEntry(const fextl::string& filename) {
@@ -1092,9 +1031,7 @@ IR::AOTIRCacheEntry* ContextImpl::LoadAOTIRCacheEntry(const fextl::string& filen
   return rv;
 }
 
-void ContextImpl::UnloadAOTIRCacheEntry(IR::AOTIRCacheEntry* Entry) {
-  IRCaptureCache.UnloadAOTIRCacheEntry(Entry);
-}
+void ContextImpl::UnloadAOTIRCacheEntry(IR::AOTIRCacheEntry* Entry) {}
 
 void ContextImpl::ConfigureAOTGen(FEXCore::Core::InternalThreadState* Thread, fextl::set<uint64_t>* ExternalBranches, uint64_t SectionMaxAddress) {
   Thread->FrontendDecoder->SetExternalBranches(ExternalBranches);

FEXCore/Source/Interface/Core/Dispatcher/Dispatcher.cpp

@@ -2,6 +2,7 @@
 
 #include "Common/SoftFloat.h"
 #include "Interface/Context/Context.h"
+#include "Interface/Core/CPUBackend.h"
 #include "Interface/Core/Dispatcher/Dispatcher.h"
 #include "Interface/Core/LookupCache.h"
 #include "Interface/Core/X86HelperGen.h"
@@ -19,11 +20,16 @@
 
 #include <CodeEmitter/Emitter.h>
 
+#ifdef VIXL_SIMULATOR
+#include <aarch64/simulator-aarch64.h>
+#endif
+
+#include <array>
 #include <atomic>
+#include <bit>
 #include <condition_variable>
 #include <csignal>
 #include <cstring>
-#include <signal.h>
 
 namespace FEXCore::CPU {
 
@@ -134,8 +140,6 @@ void Dispatcher::EmitDispatcher() {
 
   // We want to ensure that we are 16 byte aligned at the top of this loop
   Align16B();
-  ARMEmitter::BiDirectionalLabel FullLookup {};
-  ARMEmitter::BiDirectionalLabel CallBlock {};
 
   Bind(&LoopTop);
   AbsoluteLoopTopAddress = GetCursorAddress<uint64_t>();
@@ -143,23 +147,33 @@ void Dispatcher::EmitDispatcher() {
   // Load in our RIP
   ldr(RipReg, STATE_PTR(CpuStateFrame, State.rip));
 
+#ifdef _M_ARM_64EC
+  // Clobbers TMP1/2
+  // Check the EC code bitmap incase we need to exit the JIT to call into native code.
+  ARMEmitter::ForwardLabel l_NotECCode;
+  ldr(TMP1, ARMEmitter::XReg::x18, TEB_PEB_OFFSET);
+  ldr(TMP1, TMP1, PEB_EC_CODE_BITMAP_OFFSET);
+
+  lsr(ARMEmitter::Size::i64Bit, TMP2, RipReg, 15);
+  and_(ARMEmitter::Size::i64Bit, TMP2, TMP2, 0x1fffffffffff8);
+  ldr(TMP1, TMP1, TMP2, ARMEmitter::ExtendedType::LSL_64, 0);
+  lsr(ARMEmitter::Size::i64Bit, TMP2, RipReg, 12);
+  lsrv(ARMEmitter::Size::i64Bit, TMP1, TMP1, TMP2);
+  tbz(TMP1, 0, &l_NotECCode);
+
+  str(REG_CALLRET_SP, STATE_PTR(CpuStateFrame, State.callret_sp));
+
+  add(ARMEmitter::Size::i64Bit, ARMEmitter::Reg::rsp, StaticRegisters[X86State::REG_RSP], 0);
+  mov(EC_CALL_CHECKER_PC_REG, RipReg);
+  ldr(TMP2, STATE_PTR(CpuStateFrame, Pointers.Common.ExitFunctionEC));
+  br(TMP2);
+
+  Bind(&l_NotECCode);
+#endif
+
   ldrb(TMP1, STATE_PTR(CpuStateFrame, State.flags[X86State::RFLAG_TF_RAW_LOC]));
   cbnz(ARMEmitter::Size::i32Bit, TMP1, &CompileSingleStep);
 
-  // L1 Cache
-  ldr(TMP1, STATE_PTR(CpuStateFrame, Pointers.Common.L1Pointer));
-
-  and_(ARMEmitter::Size::i64Bit, TMP4, RipReg.R(), LookupCache::L1_ENTRIES_MASK);
-  add(ARMEmitter::Size::i64Bit, TMP1, TMP1, TMP4, ARMEmitter::ShiftType::LSL, 4);
-  ldp<ARMEmitter::IndexType::OFFSET>(TMP4, TMP1, TMP1, 0);
-  sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, RipReg);
-  cbnz(ARMEmitter::Size::i64Bit, TMP1, &FullLookup);
-
-  br(TMP4);
-
-  // L1C check failed, do a full lookup
-  Bind(&FullLookup);
-
   // This is the block cache lookup routine
   // It matches what is going on it LookupCache.h::FindBlock
   ldr(TMP1, STATE_PTR(CpuStateFrame, Pointers.Common.L2Pointer));
@@ -287,40 +301,10 @@ void Dispatcher::EmitDispatcher() {
     br(TMP1);
   }
 
-#ifdef _M_ARM_64EC
-  // Clobbers TMP1/2
-  auto EmitECExitCheck = [&]() {
-    // Check the EC code bitmap incase we need to exit the JIT to call into native code.
-    ARMEmitter::ForwardLabel l_NotECCode;
-    ldr(TMP1, ARMEmitter::XReg::x18, TEB_PEB_OFFSET);
-    ldr(TMP1, TMP1, PEB_EC_CODE_BITMAP_OFFSET);
-
-    lsr(ARMEmitter::Size::i64Bit, TMP2, RipReg, 15);
-    and_(ARMEmitter::Size::i64Bit, TMP2, TMP2, 0x1fffffffffff8);
-    ldr(TMP1, TMP1, TMP2, ARMEmitter::ExtendedType::LSL_64, 0);
-    lsr(ARMEmitter::Size::i64Bit, TMP2, RipReg, 12);
-    lsrv(ARMEmitter::Size::i64Bit, TMP1, TMP1, TMP2);
-    tbz(TMP1, 0, &l_NotECCode);
-
-    str(REG_CALLRET_SP, STATE_PTR(CpuStateFrame, State.callret_sp));
-
-    add(ARMEmitter::Size::i64Bit, ARMEmitter::Reg::rsp, StaticRegisters[X86State::REG_RSP], 0);
-    mov(EC_CALL_CHECKER_PC_REG, RipReg);
-    ldr(TMP2, STATE_PTR(CpuStateFrame, Pointers.Common.ExitFunctionEC));
-    br(TMP2);
-
-    Bind(&l_NotECCode);
-  };
-#endif
-
   // Need to create the block
   {
     Bind(&NoBlock);
 
-#ifdef _M_ARM_64EC
-    EmitECExitCheck();
-#endif
-
     EmitSignalGuardedRegion([&]() {
       SpillStaticRegs(TMP1);
 
@@ -355,10 +339,6 @@ void Dispatcher::EmitDispatcher() {
   {
     Bind(&CompileSingleStep);
 
-#ifdef _M_ARM_64EC
-    EmitECExitCheck();
-#endif
-
     EmitSignalGuardedRegion([&]() {
       SpillStaticRegs(TMP1);
 
@@ -569,8 +549,8 @@ void Dispatcher::EmitDispatcher() {
       FABI_F80_I16_I32_PTR,
       FABI_F32_I16_F80_PTR,
       FABI_F64_I16_F80_PTR,
-      FABI_F64_I16_F64_PTR,
-      FABI_F64_I16_F64_F64_PTR,
+      FABI_F64_F64_PTR,
+      FABI_F64_F64_F64_PTR,
       FABI_I16_I16_F80_PTR,
       FABI_I32_I16_F80_PTR,
       FABI_I64_I16_F80_PTR,
@@ -578,7 +558,7 @@ void Dispatcher::EmitDispatcher() {
       FABI_F80_I16_F80_PTR,
       FABI_F80_I16_F80_F80_PTR,
       FABI_F80x2_I16_F80_PTR,
-      FABI_F64x2_I16_F64_PTR,
+      FABI_F64x2_F64_PTR,
       FABI_I32_I64_I64_V128_V128_I16,
       FABI_I32_V128_V128_I16,
     }};
@@ -777,7 +757,7 @@ uint64_t Dispatcher::GenerateABICall(FallbackABI ABI) {
 
     ldrh(ARMEmitter::WReg::w0, STATE, offsetof(FEXCore::Core::CPUState, FCW));
     if (!TMP_ABIARGS) {
-      fmov(VABI1.D(), VTMP1.D());
+      mov(VABI1.Q(), VTMP1.Q());
     }
 
     mov(ARMEmitter::XReg::x1, STATE);
@@ -810,7 +790,7 @@ uint64_t Dispatcher::GenerateABICall(FallbackABI ABI) {
 
     FillF64Result();
   } break;
-  case FABI_F64_I16_F64_PTR: {
+  case FABI_F64_F64_PTR: {
     // Linux Reg/Win32 Reg:
     // tmp4 (x4/x13): FallbackHandler
     // x30: return
@@ -820,18 +800,17 @@ uint64_t Dispatcher::GenerateABICall(FallbackABI ABI) {
     if (!TMP_ABIARGS) {
       fmov(VABI1.D(), VTMP1.D());
     }
-    ldrh(ARMEmitter::WReg::w0, STATE, offsetof(FEXCore::Core::CPUState, FCW));
-    mov(ARMEmitter::XReg::x1, STATE);
+    mov(ARMEmitter::XReg::x0, STATE);
 
     if (!CTX->Config.DisableVixlIndirectCalls) [[unlikely]] {
-      GenerateIndirectRuntimeCall<double, uint16_t, double, uint64_t>(FallbackPointerReg);
+      GenerateIndirectRuntimeCall<double, double, uint64_t>(FallbackPointerReg);
     } else {
       blr(FallbackPointerReg);
     }
 
     FillF64Result();
   } break;
-  case FABI_F64_I16_F64_F64_PTR: {
+  case FABI_F64_F64_F64_PTR: {
     // Linux Reg/Win32 Reg:
     // tmp4 (x4/x13): FallbackHandler
     // x30: return
@@ -844,10 +823,9 @@ uint64_t Dispatcher::GenerateABICall(FallbackABI ABI) {
       fmov(VABI2.D(), VTMP2.D());
     }
 
-    ldrh(ARMEmitter::WReg::w0, STATE, offsetof(FEXCore::Core::CPUState, FCW));
-    mov(ARMEmitter::XReg::x1, STATE);
+    mov(ARMEmitter::XReg::x0, STATE);
     if (!CTX->Config.DisableVixlIndirectCalls) [[unlikely]] {
-      GenerateIndirectRuntimeCall<double, uint16_t, double, double, uint64_t>(FallbackPointerReg);
+      GenerateIndirectRuntimeCall<double, double, double, uint64_t>(FallbackPointerReg);
     } else {
       blr(FallbackPointerReg);
     }
@@ -1007,7 +985,7 @@ uint64_t Dispatcher::GenerateABICall(FallbackABI ABI) {
 
     FillF80x2Result();
   } break;
-  case FABI_F64x2_I16_F64_PTR: {
+  case FABI_F64x2_F64_PTR: {
     // Linux Reg/Win32 Reg:
     // tmp4 (x4/x13): FallbackHandler
     // x30: return
@@ -1016,14 +994,13 @@ uint64_t Dispatcher::GenerateABICall(FallbackABI ABI) {
 
     SpillForABICall(CTX->HostFeatures.SupportsPreserveAllABI, TMP3, true);
 
-    ldrh(ARMEmitter::WReg::w0, STATE, offsetof(FEXCore::Core::CPUState, FCW));
-    mov(ARMEmitter::XReg::x1, STATE);
+    mov(ARMEmitter::XReg::x0, STATE);
     if (!TMP_ABIARGS) {
       fmov(VABI1.D(), VTMP1.D());
     }
 
     if (!CTX->Config.DisableVixlIndirectCalls) [[unlikely]] {
-      // GenerateIndirectRuntimeCall<FEXCore::VectorScalarF64Pair, uint16_t, FEXCore::VectorRegType, uint64_t>(FallbackPointerReg);
+      // GenerateIndirectRuntimeCall<FEXCore::VectorScalarF64Pair, FEXCore::VectorRegType, uint64_t>(FallbackPointerReg);
     } else {
       blr(FallbackPointerReg);
     }
@@ -1125,6 +1102,53 @@ void Dispatcher::InitThreadPointers(FEXCore::Core::InternalThreadState* Thread)
   }
 }
 
+SignalDelegatorConfig Dispatcher::MakeSignalDelegatorConfig() const {
+  // PF/AF are the final two SRA registers. We only want GPRs
+  const auto GPRCount = uint16_t(StaticRegisters.size() - 2);
+  const auto FPRCount = uint16_t(StaticFPRegisters.size());
+
+  const auto GetSRAGPRMapping = [GPRCount, this] {
+    SignalDelegatorConfig::SRAIndexMapping Mapping {};
+    for (size_t i = 0; i < GPRCount; ++i) {
+      Mapping[i] = StaticRegisters[i].Idx();
+    }
+    return Mapping;
+  };
+
+  const auto GetSRAFPRMapping = [FPRCount, this] {
+    SignalDelegatorConfig::SRAIndexMapping Mapping {};
+    for (size_t i = 0; i < FPRCount; ++i) {
+      Mapping[i] = StaticFPRegisters[i].Idx();
+    }
+    return Mapping;
+  };
+
+  return FEXCore::SignalDelegatorConfig {
+    .DispatcherBegin = Start,
+    .DispatcherEnd = End,
+
+    .AbsoluteLoopTopAddress = AbsoluteLoopTopAddress,
+    .AbsoluteLoopTopAddressFillSRA = AbsoluteLoopTopAddressFillSRA,
+    .SignalHandlerReturnAddress = SignalHandlerReturnAddress,
+    .SignalHandlerReturnAddressRT = SignalHandlerReturnAddressRT,
+
+    .PauseReturnInstruction = PauseReturnInstruction,
+    .ThreadPauseHandlerAddressSpillSRA = ThreadPauseHandlerAddressSpillSRA,
+    .ThreadPauseHandlerAddress = ThreadPauseHandlerAddress,
+
+    // Stop handlers.
+    .ThreadStopHandlerAddressSpillSRA = ThreadStopHandlerAddressSpillSRA,
+    .ThreadStopHandlerAddress = ThreadStopHandlerAddress,
+
+    // SRA information.
+    .SRAGPRCount = GPRCount,
+    .SRAFPRCount = FPRCount,
+
+    .SRAGPRMapping = GetSRAGPRMapping(),
+    .SRAFPRMapping = GetSRAFPRMapping(),
+  };
+}
+
 fextl::unique_ptr<Dispatcher> Dispatcher::Create(FEXCore::Context::ContextImpl* CTX) {
   return fextl::make_unique<Dispatcher>(CTX);
 }

FEXCore/Source/Interface/Core/Dispatcher/Dispatcher.h

@@ -2,25 +2,18 @@
 #pragma once
 
 #include "Interface/Core/ArchHelpers/Arm64Emitter.h"
-#include "Interface/Core/CPUBackend.h"
 #include "Interface/Core/Interpreter/InterpreterOps.h"
 
-#include <FEXCore/Core/CoreState.h>
 #include <FEXCore/fextl/memory.h>
 
-#ifdef VIXL_SIMULATOR
-#include <aarch64/simulator-aarch64.h>
-#endif
-
+#include <array>
+#include <cstddef>
 #include <cstdint>
-#include <signal.h>
-#include <stddef.h>
-#include <stack>
-#include <tuple>
 
 namespace FEXCore {
 struct GuestSigAction;
-}
+struct SignalDelegatorConfig;
+} // namespace FEXCore
 
 namespace FEXCore::Core {
 struct CpuStateFrame;
@@ -42,6 +35,32 @@ public:
   Dispatcher(FEXCore::Context::ContextImpl* ctx);
   ~Dispatcher();
 
+  void InitThreadPointers(FEXCore::Core::InternalThreadState* Thread);
+
+#ifdef VIXL_SIMULATOR
+  void ExecuteDispatch(FEXCore::Core::CpuStateFrame* Frame);
+  void ExecuteJITCallback(FEXCore::Core::CpuStateFrame* Frame, uint64_t RIP);
+#else
+  void ExecuteDispatch(FEXCore::Core::CpuStateFrame* Frame) {
+    DispatchPtr(Frame, false);
+  }
+
+  void ExecuteJITCallback(FEXCore::Core::CpuStateFrame* Frame, uint64_t RIP) {
+    CallbackPtr(Frame, RIP);
+  }
+#endif
+
+  SignalDelegatorConfig MakeSignalDelegatorConfig() const;
+
+protected:
+  FEXCore::Context::ContextImpl* CTX;
+
+  using AsmDispatch = void (*)(FEXCore::Core::CpuStateFrame* Frame, bool SingleInst);
+  using JITCallback = void (*)(FEXCore::Core::CpuStateFrame* Frame, uint64_t RIP);
+
+  AsmDispatch DispatchPtr;
+  JITCallback CallbackPtr;
+private:
   /**
    * @name Dispatch Helper functions
    * @{ */
@@ -59,62 +78,14 @@ public:
   uint64_t GuestSignal_SIGILL {};
   uint64_t GuestSignal_SIGTRAP {};
   uint64_t GuestSignal_SIGSEGV {};
-  uint64_t IntCallbackReturnAddress {};
 
   uint64_t PauseReturnInstruction {};
   std::array<uint64_t, FallbackABI::FABI_UNKNOWN> ABIPointers {};
-
   /**  @} */
 
   uint64_t Start {};
   uint64_t End {};
 
-  void InitThreadPointers(FEXCore::Core::InternalThreadState* Thread);
-
-#ifdef VIXL_SIMULATOR
-  void ExecuteDispatch(FEXCore::Core::CpuStateFrame* Frame);
-  void ExecuteJITCallback(FEXCore::Core::CpuStateFrame* Frame, uint64_t RIP);
-#else
-  void ExecuteDispatch(FEXCore::Core::CpuStateFrame* Frame) {
-    DispatchPtr(Frame, false);
-  }
-
-  void ExecuteJITCallback(FEXCore::Core::CpuStateFrame* Frame, uint64_t RIP) {
-    CallbackPtr(Frame, RIP);
-  }
-#endif
-
-  uint16_t GetSRAGPRCount() const {
-    // PF/AF are the final two SRA registers.
-    // Only return the SRA for GPRs.
-    return StaticRegisters.size() - 2;
-  }
-
-  uint16_t GetSRAFPRCount() const {
-    return StaticFPRegisters.size();
-  }
-
-  void GetSRAGPRMapping(uint8_t Mapping[16]) const {
-    for (size_t i = 0; i < StaticRegisters.size() - 2; ++i) {
-      Mapping[i] = StaticRegisters[i].Idx();
-    }
-  }
-
-  void GetSRAFPRMapping(uint8_t Mapping[16]) const {
-    for (size_t i = 0; i < StaticFPRegisters.size(); ++i) {
-      Mapping[i] = StaticFPRegisters[i].Idx();
-    }
-  }
-
-protected:
-  FEXCore::Context::ContextImpl* CTX;
-
-  using AsmDispatch = void (*)(FEXCore::Core::CpuStateFrame* Frame, bool SingleInst);
-  using JITCallback = void (*)(FEXCore::Core::CpuStateFrame* Frame, uint64_t RIP);
-
-  AsmDispatch DispatchPtr;
-  JITCallback CallbackPtr;
-private:
   // Long division helpers
   uint64_t LUDIVHandlerAddress {};
   uint64_t LDIVHandlerAddress {};

FEXCore/Source/Interface/Core/Frontend.cpp

@@ -71,7 +71,23 @@ Decoder::Decoder(FEXCore::Core::InternalThreadState* Thread)
   : Thread {Thread}
   , CTX {static_cast<FEXCore::Context::ContextImpl*>(Thread->CTX)}
   , OSABI {CTX->SyscallHandler ? CTX->SyscallHandler->GetOSABI() : FEXCore::HLE::SyscallOSABI::OS_UNKNOWN}
-  , PoolObject {CTX->FrontendAllocator, sizeof(FEXCore::X86Tables::DecodedInst) * DefaultDecodedBufferSize} {}
+  , PoolObject {CTX->FrontendAllocator, sizeof(FEXCore::X86Tables::DecodedInst) * DefaultDecodedBufferSize} {
+
+  FEX_CONFIG_OPT(ReducedPrecision, X87REDUCEDPRECISION);
+  if (ReducedPrecision) {
+    X87Table = &FEXCore::X86Tables::X87F64Ops;
+  } else {
+    X87Table = &FEXCore::X86Tables::X87F80Ops;
+  }
+
+  if (CTX->HostFeatures.SupportsAVX && CTX->HostFeatures.SupportsSVE256) {
+    VEXTable = &FEXCore::X86Tables::VEXTableOps;
+    VEXTableGroup = &FEXCore::X86Tables::VEXTableGroupOps;
+  } else if (CTX->HostFeatures.SupportsAVX) {
+    VEXTable = &FEXCore::X86Tables::VEXTableOps_AVX128;
+    VEXTableGroup = &FEXCore::X86Tables::VEXTableGroupOps_AVX128;
+  }
+}
 
 bool Decoder::CheckRangeExecutable(uint64_t Address, uint64_t Size) {
   // Treat FEX-internal X86 callbacks as always executable
@@ -83,6 +99,7 @@ bool Decoder::CheckRangeExecutable(uint64_t Address, uint64_t Size) {
     auto RangeInfo = CTX->SyscallHandler->QueryGuestExecutableRange(Thread, Address);
     ExecutableRangeBase = RangeInfo.Base;
     ExecutableRangeEnd = RangeInfo.Base + RangeInfo.Size;
+    ExecutableRangeWritable = RangeInfo.Writable;
 
     if (RangeInfo.Size == 0) {
       return false;
@@ -314,6 +331,13 @@ void Decoder::DecodeModRM_64(X86Tables::DecodedOperand* Operand, X86Tables::ModR
 }
 
 bool Decoder::NormalOp(const FEXCore::X86Tables::X86InstInfo* Info, uint16_t Op, DecodedHeader Options) {
+  if (Info->Type == FEXCore::X86Tables::TYPE_ARCH_DISPATCHER) [[unlikely]] {
+    // Dispatcher Op.
+    // TODO: Move this in to `NormalOpHeader`, Dispatch tables have a bug currently where some subtables don't inherit flags correctly.
+    // Can be seen by running FEX asm tests if this is removed.
+    return NormalOp(&Info->OpcodeDispatcher.Indirect[BlockInfo.Is64BitMode ? 1 : 0], Op);
+  }
+
   DecodeInst->OP = Op;
   DecodeInst->TableInfo = Info;
 
@@ -626,7 +650,7 @@ bool Decoder::NormalOp(const FEXCore::X86Tables::X86InstInfo* Info, uint16_t Op,
 }
 
 bool Decoder::NormalOpHeader(const FEXCore::X86Tables::X86InstInfo* Info, uint16_t Op) {
-  DecodeInst->OP = Op;
+  DecodeInst->OPRaw = DecodeInst->OP = Op;
   DecodeInst->TableInfo = Info;
 
   if (Info->Type == FEXCore::X86Tables::TYPE_UNKNOWN) {
@@ -642,6 +666,9 @@ bool Decoder::NormalOpHeader(const FEXCore::X86Tables::X86InstInfo* Info, uint16
   // A normal instruction is the most likely.
   if (Info->Type == FEXCore::X86Tables::TYPE_INST) [[likely]] {
     return NormalOp(Info, Op);
+  } else if (Info->Type == FEXCore::X86Tables::TYPE_ARCH_DISPATCHER) [[unlikely]] {
+    // Dispatcher Op.
+    return NormalOp(&Info->OpcodeDispatcher.Indirect[BlockInfo.Is64BitMode ? 1 : 0], Op);
   } else if (Info->Type >= FEXCore::X86Tables::TYPE_GROUP_1 && Info->Type <= FEXCore::X86Tables::TYPE_GROUP_11) {
     uint8_t ModRMByte = ReadByte();
     DecodeInst->ModRM = ModRMByte;
@@ -679,7 +706,7 @@ bool Decoder::NormalOpHeader(const FEXCore::X86Tables::X86InstInfo* Info, uint16
     ModRM.Hex = DecodeInst->ModRM;
 
     uint16_t LocalOp = OPD(Info->Type, PrefixType, ModRM.reg);
-    FEXCore::X86Tables::X86InstInfo* LocalInfo = &SecondInstGroupOps[LocalOp];
+    const FEXCore::X86Tables::X86InstInfo* LocalInfo = &SecondInstGroupOps[LocalOp];
 #undef OPD
     if (LocalInfo->Type == FEXCore::X86Tables::TYPE_SECOND_GROUP_MODRM && ModRM.mod == 0b11) {
       // Everything in this group is privileged instructions aside from XGETBV
@@ -703,8 +730,13 @@ bool Decoder::NormalOpHeader(const FEXCore::X86Tables::X86InstInfo* Info, uint16
     DecodeInst->DecodedModRM = true;
 
     uint16_t X87Op = ((Op - 0xD8) << 8) | ModRMByte;
-    return NormalOp(&X87Ops[X87Op], X87Op);
+    return NormalOp(&(*X87Table)[X87Op], X87Op);
   } else if (Info->Type == FEXCore::X86Tables::TYPE_VEX_TABLE_PREFIX) {
+    if (!VEXTable) {
+      // AVX not enabled.
+      return false;
+    }
+
     uint16_t map_select = 1;
     uint16_t pp = 0;
     const uint8_t Byte1 = ReadByte();
@@ -742,7 +774,6 @@ bool Decoder::NormalOpHeader(const FEXCore::X86Tables::X86InstInfo* Info, uint16
         DecodeInst->Flags |= DecodeFlags::FLAG_OPTION_AVX_W;
       }
       if (!(map_select >= 1 && map_select <= 3)) {
-        LogMan::Msg::EFmt("We don't understand a map_select of: {}", map_select);
         return false;
       }
     }
@@ -752,7 +783,7 @@ bool Decoder::NormalOpHeader(const FEXCore::X86Tables::X86InstInfo* Info, uint16
     Op = OPD(map_select, pp, VEXOp);
 #undef OPD
 
-    FEXCore::X86Tables::X86InstInfo* LocalInfo = &VEXTableOps[Op];
+    const FEXCore::X86Tables::X86InstInfo* LocalInfo = &(*VEXTable)[Op];
 
     if (LocalInfo->Type >= FEXCore::X86Tables::TYPE_VEX_GROUP_12 && LocalInfo->Type <= FEXCore::X86Tables::TYPE_VEX_GROUP_17) {
       // We have ModRM
@@ -766,7 +797,7 @@ bool Decoder::NormalOpHeader(const FEXCore::X86Tables::X86InstInfo* Info, uint16
 #define OPD(group, pp, opcode) (((group - TYPE_VEX_GROUP_12) << 4) | (pp << 3) | (opcode))
       Op = OPD(LocalInfo->Type, pp, ModRM.reg);
 #undef OPD
-      return NormalOp(&VEXTableGroupOps[Op], Op, options);
+      return NormalOp(&(*VEXTableGroup)[Op], Op, options);
     } else {
       return NormalOp(LocalInfo, Op, options);
     }
@@ -955,7 +986,10 @@ bool Decoder::DecodeInstructionImpl(uint64_t PC) {
       break;
     default:
       [[likely]] { // Default base table
-        auto Info = &FEXCore::X86Tables::BaseOps[Op];
+        const X86InstInfo* Info = &FEXCore::X86Tables::BaseOps[Op];
+        if (Info->Type == FEXCore::X86Tables::TYPE_ARCH_DISPATCHER) {
+          Info = &Info->OpcodeDispatcher.Indirect[BlockInfo.Is64BitMode ? 1 : 0];
+        }
 
         if (Info->Type == FEXCore::X86Tables::TYPE_REX_PREFIX) {
           DecodeInst->Flags |= DecodeFlags::FLAG_REX_PREFIX;
@@ -1007,11 +1041,27 @@ Decoder::DecodedBlockStatus Decoder::DecodeInstruction(uint64_t PC) {
     DecodeInst->TableInfo = nullptr;
     DecodeInst->InstSize = 0;
     return ErrorDuringDecoding ? DecodedBlockStatus::INVALID_INST : DecodedBlockStatus::NOEXEC_INST;
-  } else if (!DecodeInst->TableInfo || !DecodeInst->TableInfo->OpcodeDispatcher) {
+  } else if (!DecodeInst->TableInfo || (DecodeInst->TableInfo->Type == TYPE_INST && !DecodeInst->TableInfo->OpcodeDispatcher.OpDispatch)) {
     // If there wasn't an error during decoding but we have no dispatcher for the instruction then claim invalid instruction.
     return DecodedBlockStatus::INVALID_INST;
   }
 
+  if (CTX->AreMonoHacksActive()) {
+    // Unity uses a standard SPSC ringbuffer with cached read/write pointers and thread waiting flags at the following
+    // offsets, which are consistent between 32-bit and 64-bit Unity versions from 2015 onwards.
+    auto IsKnownAtomicDisplacement = [](uint64_t Displacement) {
+      return Displacement == 0x80 || Displacement == 0x84 || Displacement == 0xC0 || Displacement == 0xC4;
+    };
+
+    if (DecodeInst->OP == 0x8b && DecodeInst->Src[0].IsGPRIndirect() &&
+        IsKnownAtomicDisplacement(DecodeInst->Src[0].Data.GPRIndirect.Displacement)) {
+      DecodeInst->ForceTSO = true;
+    }
+    if (DecodeInst->OP == 0x89 && DecodeInst->Dest.IsGPRIndirect() && IsKnownAtomicDisplacement(DecodeInst->Dest.Data.GPRIndirect.Displacement)) {
+      DecodeInst->ForceTSO = true;
+    }
+  }
+
   return DecodedBlockStatus::SUCCESS;
 }
 
@@ -1027,6 +1077,13 @@ void Decoder::BranchTargetInMultiblockRange() {
   const auto InstEnd = DecodeInst->PC + DecodeInst->InstSize;
 
   if (DecodeInst->TableInfo->Flags & FEXCore::X86Tables::InstFlags::FLAGS_CALL) {
+    if (ExecutableRangeWritable && CTX->AreMonoHacksActive()) {
+      // Mono generated code often contains noreturn calls with garbage following them, and calls are always backpatched
+      // after CIL compilation leading to n recompiles for a multiblock with n calls. Choose to minimize stutters over
+      // raw performance and disable tracking past calls for mono generated code.
+      return;
+    }
+
     AddBranchTarget(InstEnd);
     BlockInfo.EntryPoints.emplace(InstEnd);
     return;
@@ -1058,10 +1115,16 @@ void Decoder::BranchTargetInMultiblockRange() {
     TargetRIP &= 0xFFFFFFFFU;
   }
 
+  if (Conditional) {
+    // If we are conditional then a target can be the instruction past the conditional instruction
+    AddBranchTarget(InstEnd);
+  }
+
   // If the target RIP is x86 code within the symbol ranges then we are golden
-  // Forbid cross-page branches to both avoid massive (range-wise) code blocks in highly fragmented code and trying to decode unmapped branch targets
-  bool ValidMultiblockMember =
-    TargetRIP >= SymbolMinAddress && TargetRIP < std::min(FEXCore::AlignUp(InstEnd, FEXCore::Utils::FEX_PAGE_SIZE), SymbolMaxAddress);
+  // Forbid distant branches to have the cost code better match the guest code layout, avoiding massive (range-wise) code
+  // blocks in highly fragmented guest code. Such branches are often not-taken branches to garbage in obfuscated code.
+  constexpr uint64_t MAX_FORWARD_BRANCH_DIST = FEXCore::Utils::FEX_PAGE_SIZE * 4;
+  bool ValidMultiblockMember = TargetRIP >= SymbolMinAddress && TargetRIP < std::min(InstEnd + MAX_FORWARD_BRANCH_DIST, SymbolMaxAddress);
 
 #ifdef _M_ARM_64EC
   ValidMultiblockMember = ValidMultiblockMember && !RtlIsEcCode(TargetRIP);
@@ -1072,9 +1135,6 @@ void Decoder::BranchTargetInMultiblockRange() {
     if (Conditional) {
       MaxCondBranchForward = std::max(MaxCondBranchForward, TargetRIP);
       MaxCondBranchBackwards = std::min(MaxCondBranchBackwards, TargetRIP);
-
-      // If we are conditional then a target can be the instruction past the conditional instruction
-      AddBranchTarget(InstEnd);
     }
 
     AddBranchTarget(TargetRIP);
@@ -1085,6 +1145,60 @@ void Decoder::BranchTargetInMultiblockRange() {
   }
 }
 
+bool Decoder::IsBranchMonoTailcall(uint64_t NumInstructions) const {
+  // While the mono call backpatching block can easily be detected due it being the only one to contain SMC-faulting
+  // atomics, that can't be said for the tailcall jump backpatcher which has changed several times across versions and
+  // can be partially inlined. To work around this, instead detect the tailcall site itself and force full non-signal-based
+  // SMC detection for that single block.
+  if (!ExecutableRangeWritable) {
+    // We only care about jitted code
+    return false;
+  }
+
+  // See mini-{amd64,x86}.c in the mono codebase, specifically where METHOD_JUMP patches are emitted.
+  if (GetGPROpSize() == IR::OpSize::i32Bit) {
+    // Matches:
+    // LEAVE
+    // <none> / NOP / MOV EAX, EAX / LEA EBP, [EBP+0]
+    // JMP imm32
+    if (DecodeInst->OP != 0xE9 || NumInstructions < 2) {
+      return false;
+    }
+
+    auto PrevInst = std::prev(DecodeInst);
+    if (PrevInst->OP == 0xC9) {
+      return true;
+    }
+
+    if (NumInstructions < 3 || std::prev(PrevInst)->OP != 0xC9) {
+      return false;
+    }
+
+    return PrevInst->OP == 0x90 || (PrevInst->OP == 0x8B && PrevInst->ModRM == 0xC0) ||
+           (PrevInst->OP == 0x8D && PrevInst->ModRM == 0x6D && PrevInst->Src[1].IsLiteral() && PrevInst->Src[1].Literal() == 0);
+  } else {
+    FEXCore::X86Tables::ModRMDecoded ModRM;
+    ModRM.Hex = DecodeInst->ModRM;
+    if (DecodeInst->OPRaw == 0xFF && ModRM.reg == 4 && DecodeInst->Src[0].IsGPR()) {
+      if (DecodeInst->Src[0].Data.GPR.GPR == FEXCore::X86State::REG_RAX) {
+        // Found in versions of mono from 2024 onwards - matches:
+        // REX.W JMP rax
+        return (DecodeInst->Flags & (DecodeFlags::FLAG_REX_PREFIX | DecodeFlags::FLAG_REX_WIDENING | DecodeFlags::FLAG_REX_XGPR_B |
+                                     DecodeFlags::FLAG_REX_XGPR_X | DecodeFlags::FLAG_REX_XGPR_R)) ==
+               (DecodeFlags::FLAG_REX_PREFIX | DecodeFlags::FLAG_REX_WIDENING);
+      } else if (NumInstructions > 1 && DecodeInst->Src[0].Data.GPR.GPR == FEXCore::X86State::REG_R11) {
+        // Found in older versions of mono - match:
+        // MOV r11, imm64
+        // JMP r11
+        auto PrevInst = std::prev(DecodeInst);
+        return PrevInst->OP == 0xBB && PrevInst->Dest.IsGPR() && PrevInst->Dest.Data.GPR.GPR == FEXCore::X86State::REG_R11;
+      }
+    }
+  }
+
+  return false;
+}
+
 bool Decoder::InstCanContinue() const {
   if (DecodeInst->PC + DecodeInst->InstSize == NextBlockStartAddress) {
     return false;
@@ -1199,8 +1313,8 @@ void Decoder::DecodeInstructionsAtEntry(FEXCore::Core::InternalThreadState *Thre
   DecodedBuffer = PoolObject.ReownOrClaimBuffer();
 
   // Decode operating mode from thread's CS segment.
-  const auto CSSegment = Thread->CurrentFrame->State.gdt[Thread->CurrentFrame->State.cs_idx >> 3];
-  BlockInfo.Is64BitMode = CSSegment.L == 1;
+  const auto CSSegment = Core::CPUState::GetSegmentFromIndex(Thread->CurrentFrame->State, Thread->CurrentFrame->State.cs_idx);
+  BlockInfo.Is64BitMode = CSSegment->L == 1;
   LOGMAN_THROW_A_FMT(BlockInfo.Is64BitMode == CTX->Config.Is64BitMode, "Expected operating mode to not change at runtime!");
 
   // XXX: Load symbol data
@@ -1345,6 +1459,7 @@ void Decoder::DecodeInstructionsAtEntry(FEXCore::Core::InternalThreadState *Thre
           // If the branch target is within our multiblock range then we can keep going on
           // We don't want to short circuit this since we want to calculate our ranges still
           // NOTE: This will invalidate BlockIt, this is fine as we immediately break from the loop and EraseBlock cannot be true
+          BlockIt->ForceFullSMCDetection = CTX->AreMonoHacksActive() && IsBranchMonoTailcall(BlockIt->NumInstructions);
           BranchTargetInMultiblockRange();
         }
 

FEXCore/Source/Interface/Core/Frontend.h

@@ -4,18 +4,21 @@
 #include "Interface/Core/X86Tables/X86Tables.h"
 #include "Interface/IR/IR.h"
 
-#include <FEXCore/HLE/SyscallHandler.h>
-#include <FEXCore/Utils/Telemetry.h>
+#include <FEXCore/Utils/ThreadPoolAllocator.h>
 #include <FEXCore/fextl/set.h>
 #include <FEXCore/fextl/vector.h>
 
 #include <array>
+#include <cstddef>
 #include <cstdint>
-#include <stddef.h>
+#include <optional>
 
 namespace FEXCore::Context {
 class ContextImpl;
 }
+namespace FEXCore::HLE {
+enum class SyscallOSABI;
+}
 
 namespace FEXCore::Frontend {
 class Decoder final {
@@ -34,6 +37,7 @@ public:
     FEXCore::X86Tables::DecodedInst* DecodedInstructions;
     DecodedBlockStatus BlockStatus;
     bool IsEntryPoint {};
+    bool ForceFullSMCDetection {};
   };
 
   struct DecodedBlockInformation final {
@@ -86,6 +90,7 @@ private:
   DecodedBlockStatus DecodeInstruction(uint64_t PC);
 
   void BranchTargetInMultiblockRange();
+  bool IsBranchMonoTailcall(uint64_t NumInstructions) const;
   bool InstCanContinue() const;
 
   void AddBranchTarget(uint64_t Target);
@@ -109,6 +114,7 @@ private:
 
   uint64_t ExecutableRangeBase {};
   uint64_t ExecutableRangeEnd {};
+  bool ExecutableRangeWritable {};
   bool HitNonExecutableRange {};
 
   const uint8_t* InstStream {};
@@ -147,6 +153,11 @@ private:
     &FEXCore::Frontend::Decoder::DecodeModRM_16,
   };
 
+  const std::array<X86Tables::X86InstInfo, X86Tables::MAX_X87_TABLE_SIZE>* X87Table;
+
+  const std::array<X86Tables::X86InstInfo, X86Tables::MAX_VEX_TABLE_SIZE>* VEXTable {};
+  const std::array<X86Tables::X86InstInfo, X86Tables::MAX_VEX_GROUP_TABLE_SIZE>* VEXTableGroup {};
+
   const uint8_t* AdjustAddrForSpecialRegion(const uint8_t* _InstStream, uint64_t EntryPoint, uint64_t RIP);
 };
 } // namespace FEXCore::Frontend

FEXCore/Source/Interface/Core/Interpreter/Fallbacks/F80Fallbacks.h

@@ -340,7 +340,7 @@ struct OpHandlers<IR::OP_F80SCALE> {
 
 template<>
 struct OpHandlers<IR::OP_F64SIN> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     return sin(src);
   }
@@ -348,7 +348,7 @@ struct OpHandlers<IR::OP_F64SIN> {
 
 template<>
 struct OpHandlers<IR::OP_F64COS> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     return cos(src);
   }
@@ -356,7 +356,7 @@ struct OpHandlers<IR::OP_F64COS> {
 
 template<>
 struct OpHandlers<IR::OP_F64SINCOS> {
-  FEXCORE_PRESERVE_ALL_ATTR static VectorScalarF64Pair handle(uint16_t FCW, double src, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static VectorScalarF64Pair handle(double src, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     double sin, cos;
 #ifdef _WIN32
@@ -371,7 +371,7 @@ struct OpHandlers<IR::OP_F64SINCOS> {
 
 template<>
 struct OpHandlers<IR::OP_F64TAN> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     return tan(src);
   }
@@ -379,7 +379,7 @@ struct OpHandlers<IR::OP_F64TAN> {
 
 template<>
 struct OpHandlers<IR::OP_F64F2XM1> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     return exp2(src) - 1.0;
   }
@@ -387,7 +387,7 @@ struct OpHandlers<IR::OP_F64F2XM1> {
 
 template<>
 struct OpHandlers<IR::OP_F64ATAN> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     return atan2(src1, src2);
   }
@@ -395,7 +395,7 @@ struct OpHandlers<IR::OP_F64ATAN> {
 
 template<>
 struct OpHandlers<IR::OP_F64FPREM> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     return fmod(src1, src2);
   }
@@ -403,7 +403,7 @@ struct OpHandlers<IR::OP_F64FPREM> {
 
 template<>
 struct OpHandlers<IR::OP_F64FPREM1> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     return remainder(src1, src2);
   }
@@ -411,7 +411,7 @@ struct OpHandlers<IR::OP_F64FPREM1> {
 
 template<>
 struct OpHandlers<IR::OP_F64FYL2X> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     return src2 * log2(src1);
   }
@@ -419,7 +419,7 @@ struct OpHandlers<IR::OP_F64FYL2X> {
 
 template<>
 struct OpHandlers<IR::OP_F64SCALE> {
-  FEXCORE_PRESERVE_ALL_ATTR static double handle(uint16_t FCW, double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
+  FEXCORE_PRESERVE_ALL_ATTR static double handle(double src1, double src2, FEXCore::Core::CpuStateFrame* Frame) {
     FEXCORE_PROFILE_INSTANT_INCREMENT(Frame->Thread, AccumulatedFloatFallbackCount, 1);
     if (src1 == 0.0) { // src1 might be +/- zero
       return src1;     // this will return negative or positive zero if when appropriate
@@ -445,7 +445,6 @@ struct OpHandlers<IR::OP_F80BCDSTORE> {
     uint64_t Tmp = Src1.ToI64(&State.State);
     X80SoftFloat Rv;
     uint8_t* BCD = reinterpret_cast<uint8_t*>(&Rv);
-    memset(BCD, 0, 10);
 
     for (size_t i = 0; i < 9; ++i) {
       if (Tmp == 0) {

FEXCore/Source/Interface/Core/Interpreter/Fallbacks/InterpreterFallbacks.cpp

@@ -82,24 +82,24 @@ void InterpreterOps::FillFallbackIndexPointers(Core::FallbackABIInfo* Info, uint
                                   reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F80SCALE>::handle)};
 
   // Double Precision Unary
-  Info[Core::OPINDEX_F64SIN] = {ABIHandlers[FABI_F64_I16_F64_PTR], reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64SIN>::handle)};
-  Info[Core::OPINDEX_F64COS] = {ABIHandlers[FABI_F64_I16_F64_PTR], reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64COS>::handle)};
-  Info[Core::OPINDEX_F64SINCOS] = {ABIHandlers[FABI_F64x2_I16_F64_PTR],
+  Info[Core::OPINDEX_F64SIN] = {ABIHandlers[FABI_F64_F64_PTR], reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64SIN>::handle)};
+  Info[Core::OPINDEX_F64COS] = {ABIHandlers[FABI_F64_F64_PTR], reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64COS>::handle)};
+  Info[Core::OPINDEX_F64SINCOS] = {ABIHandlers[FABI_F64x2_F64_PTR],
                                    reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64SINCOS>::handle)};
-  Info[Core::OPINDEX_F64TAN] = {ABIHandlers[FABI_F64_I16_F64_PTR], reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64TAN>::handle)};
-  Info[Core::OPINDEX_F64F2XM1] = {ABIHandlers[FABI_F64_I16_F64_PTR],
+  Info[Core::OPINDEX_F64TAN] = {ABIHandlers[FABI_F64_F64_PTR], reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64TAN>::handle)};
+  Info[Core::OPINDEX_F64F2XM1] = {ABIHandlers[FABI_F64_F64_PTR],
                                   reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64F2XM1>::handle)};
 
   // Double Precision Binary
-  Info[Core::OPINDEX_F64ATAN] = {ABIHandlers[FABI_F64_I16_F64_F64_PTR],
+  Info[Core::OPINDEX_F64ATAN] = {ABIHandlers[FABI_F64_F64_F64_PTR],
                                  reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64ATAN>::handle)};
-  Info[Core::OPINDEX_F64FPREM] = {ABIHandlers[FABI_F64_I16_F64_F64_PTR],
+  Info[Core::OPINDEX_F64FPREM] = {ABIHandlers[FABI_F64_F64_F64_PTR],
                                   reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64FPREM>::handle)};
-  Info[Core::OPINDEX_F64FPREM1] = {ABIHandlers[FABI_F64_I16_F64_F64_PTR],
+  Info[Core::OPINDEX_F64FPREM1] = {ABIHandlers[FABI_F64_F64_F64_PTR],
                                    reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64FPREM1>::handle)};
-  Info[Core::OPINDEX_F64FYL2X] = {ABIHandlers[FABI_F64_I16_F64_F64_PTR],
+  Info[Core::OPINDEX_F64FYL2X] = {ABIHandlers[FABI_F64_F64_F64_PTR],
                                   reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64FYL2X>::handle)};
-  Info[Core::OPINDEX_F64SCALE] = {ABIHandlers[FABI_F64_I16_F64_F64_PTR],
+  Info[Core::OPINDEX_F64SCALE] = {ABIHandlers[FABI_F64_F64_F64_PTR],
                                   reinterpret_cast<uint64_t>(&FEXCore::CPU::OpHandlers<IR::OP_F64SCALE>::handle)};
 
   // SSE4.2 string instructions
@@ -222,18 +222,18 @@ bool InterpreterOps::GetFallbackHandler(const IR::IROp_Header* IROp, FallbackInf
 
 #define COMMON_UNARY_F64_OP(OP)                            \
   case IR::OP_F64##OP: {                                   \
-    *Info = {FABI_F64_I16_F64_PTR, Core::OPINDEX_F64##OP}; \
+    *Info = {FABI_F64_F64_PTR, Core::OPINDEX_F64##OP}; \
     return true;                                           \
   }
 #define COMMON_UNARYPAIR_F64_OP(OP)                          \
   case IR::OP_F64##OP: {                                     \
-    *Info = {FABI_F64x2_I16_F64_PTR, Core::OPINDEX_F64##OP}; \
+    *Info = {FABI_F64x2_F64_PTR, Core::OPINDEX_F64##OP}; \
     return true;                                             \
   }
 
 #define COMMON_BINARY_F64_OP(OP)                               \
   case IR::OP_F64##OP: {                                       \
-    *Info = {FABI_F64_I16_F64_F64_PTR, Core::OPINDEX_F64##OP}; \
+    *Info = {FABI_F64_F64_F64_PTR, Core::OPINDEX_F64##OP}; \
     return true;                                               \
   }
 

FEXCore/Source/Interface/Core/Interpreter/InterpreterOps.h

@@ -19,8 +19,8 @@ enum FallbackABI {
   FABI_F80_I16_I32_PTR,
   FABI_F32_I16_F80_PTR,
   FABI_F64_I16_F80_PTR,
-  FABI_F64_I16_F64_PTR,
-  FABI_F64_I16_F64_F64_PTR,
+  FABI_F64_F64_PTR,
+  FABI_F64_F64_F64_PTR,
   FABI_I16_I16_F80_PTR,
   FABI_I32_I16_F80_PTR,
   FABI_I64_I16_F80_PTR,
@@ -28,7 +28,7 @@ enum FallbackABI {
   FABI_F80_I16_F80_PTR,
   FABI_F80_I16_F80_F80_PTR,
   FABI_F80x2_I16_F80_PTR,
-  FABI_F64x2_I16_F64_PTR,
+  FABI_F64x2_F64_PTR,
   FABI_I32_I64_I64_V128_V128_I16,
   FABI_I32_V128_V128_I16,
   FABI_UNKNOWN,

FEXCore/Source/Interface/Core/JIT/ALUOps.cpp

@@ -515,6 +515,12 @@ DEF_OP(AndWithFlags) {
   }
 }
 
+DEF_OP(AndShift) {
+  auto Op = IROp->C<IR::IROp_XorShift>();
+
+  and_(ConvertSize48(IROp), GetReg(Node), GetReg(Op->Src1), GetReg(Op->Src2), ConvertIRShiftType(Op->Shift), Op->ShiftAmount);
+}
+
 DEF_OP(XorShift) {
   auto Op = IROp->C<IR::IROp_XorShift>();
 
@@ -1368,12 +1374,12 @@ DEF_OP(VExtractToGPR) {
   const auto Op = IROp->C<IR::IROp_VExtractToGPR>();
   const auto OpSize = IROp->Size;
 
-  [[maybe_unused]] constexpr auto AVXRegBitSize = Core::CPUState::XMM_AVX_REG_SIZE * 8;
+  constexpr auto AVXRegBitSize = Core::CPUState::XMM_AVX_REG_SIZE * 8;
   constexpr auto SSERegBitSize = Core::CPUState::XMM_SSE_REG_SIZE * 8;
   const auto ElementSizeBits = IR::OpSizeAsBits(Op->Header.ElementSize);
 
   const auto Offset = ElementSizeBits * Op->Index;
-  [[maybe_unused]] const auto Is256Bit = Offset >= SSERegBitSize;
+  const auto Is256Bit = Offset >= SSERegBitSize;
   LOGMAN_THROW_A_FMT(!Is256Bit || HostSupportsSVE256, "Need SVE256 support in order to use {} with 256-bit operation", __func__);
 
   const auto Dst = GetReg(Node);

FEXCore/Source/Interface/Core/JIT/Arm64Relocations.cpp

@@ -33,7 +33,7 @@ void Arm64JITCore::InsertNamedThunkRelocation(ARMEmitter::Register Reg, const IR
 
   uint64_t Pointer = reinterpret_cast<uint64_t>(EmitterCTX->ThunkHandler->LookupThunk(Sum));
 
-  LoadConstant(ARMEmitter::Size::i64Bit, Reg, Pointer, EmitterCTX->Config.CacheObjectCodeCompilation());
+  LoadConstant(ARMEmitter::Size::i64Bit, Reg, Pointer, false);
   Relocations.emplace_back(MoveABI);
 }
 
@@ -77,7 +77,7 @@ void Arm64JITCore::InsertGuestRIPMove(ARMEmitter::Register Reg, uint64_t Constan
   MoveABI.GuestRIPMove.GuestRIP = Constant;
   MoveABI.GuestRIPMove.RegisterIndex = Reg.Idx();
 
-  LoadConstant(ARMEmitter::Size::i64Bit, Reg, Constant, EmitterCTX->Config.CacheObjectCodeCompilation());
+  LoadConstant(ARMEmitter::Size::i64Bit, Reg, Constant, false);
   Relocations.emplace_back(MoveABI);
 }
 

FEXCore/Source/Interface/Core/JIT/AtomicOps.cpp

@@ -322,13 +322,26 @@ DEF_OP(AtomicFetchNeg) {
 
   auto MemSrc = GetReg(Op->Addr);
 
-  ARMEmitter::BackwardLabel LoopTop;
-  Bind(&LoopTop);
-  ldaxr(SubEmitSize, TMP2, MemSrc);
-  neg(EmitSize, TMP3, TMP2);
-  stlxr(SubEmitSize, TMP4, TMP3, MemSrc);
-  cbnz(EmitSize, TMP4, &LoopTop);
-  mov(EmitSize, GetReg(Node), TMP2.R());
+  if (CTX->HostFeatures.SupportsAtomics) {
+    // Use a CAS loop to avoid needing to emulate unaligned LLSC atomics
+    ldr(SubEmitSize, TMP2, MemSrc);
+    ARMEmitter::BackwardLabel LoopTop;
+    Bind(&LoopTop);
+    mov(EmitSize, TMP4, TMP2);
+    neg(EmitSize, TMP3, TMP2);
+    casal(SubEmitSize, TMP2, TMP3, MemSrc);
+    sub(EmitSize, TMP3, TMP2, TMP4);
+    cbnz(EmitSize, TMP3, &LoopTop);
+    mov(EmitSize, GetReg(Node), TMP2.R());
+  } else {
+    ARMEmitter::BackwardLabel LoopTop;
+    Bind(&LoopTop);
+    ldaxr(SubEmitSize, TMP2, MemSrc);
+    neg(EmitSize, TMP3, TMP2);
+    stlxr(SubEmitSize, TMP4, TMP3, MemSrc);
+    cbnz(EmitSize, TMP4, &LoopTop);
+    mov(EmitSize, GetReg(Node), TMP2.R());
+  }
 }
 
 DEF_OP(TelemetrySetValue) {

FEXCore/Source/Interface/Core/JIT/BranchOps.cpp

@@ -146,6 +146,15 @@ DEF_OP(ExitFunction) {
         } else {
           stp<ARMEmitter::IndexType::PRE>(ARMEmitter::XReg::zr, ARMEmitter::XReg::zr, REG_CALLRET_SP, -0x10);
         }
+      } else if (Op->Hint == IR::BranchHint::CheckTF) {
+        ARMEmitter::ForwardLabel TFUnset;
+        ldrb(TMP1, STATE_PTR(CpuStateFrame, State.flags[X86State::RFLAG_TF_RAW_LOC]));
+        cbz(ARMEmitter::Size::i32Bit, TMP1, &TFUnset);
+        LoadConstant(ARMEmitter::Size::i64Bit, TMP1, NewRIP);
+        str(TMP1, STATE, offsetof(FEXCore::Core::CpuStateFrame, State.rip));
+        ldr(TMP2, STATE, offsetof(FEXCore::Core::CpuStateFrame, Pointers.Common.DispatcherLoopTop));
+        blr(TMP2);
+        Bind(&TFUnset);
       }
 
       EmitLinkedBranch(NewRIP, Op->Hint == IR::BranchHint::Call);
@@ -205,21 +214,20 @@ DEF_OP(ExitFunction) {
 
 DEF_OP(Jump) {
   const auto Op = IROp->C<IR::IROp_Jump>();
-  const auto Target = Op->TargetBlock;
 
-  PendingTargetLabel = &JumpTargets.try_emplace(Target.ID()).first->second;
+  PendingTargetLabel = JumpTarget(Op->TargetBlock);
 }
 
 DEF_OP(CondJump) {
   auto Op = IROp->C<IR::IROp_CondJump>();
 
-  auto TrueTargetLabel = &JumpTargets.try_emplace(Op->TrueBlock.ID()).first->second;
+  auto TrueTargetLabel = JumpTarget(Op->TrueBlock);
 
   if (Op->FromNZCV) {
     b(MapCC(Op->Cond), TrueTargetLabel);
   } else {
-    [[maybe_unused]] uint64_t Const;
-    [[maybe_unused]] const bool isConst = IsInlineConstant(Op->Cmp2, &Const);
+    uint64_t Const;
+    const bool isConst = IsInlineConstant(Op->Cmp2, &Const);
 
     auto Reg = GetReg(Op->Cmp1);
     const auto Size = Op->CompareSize == IR::OpSize::i32Bit ? ARMEmitter::Size::i32Bit : ARMEmitter::Size::i64Bit;
@@ -244,7 +252,7 @@ DEF_OP(CondJump) {
     }
   }
 
-  PendingTargetLabel = &JumpTargets.try_emplace(Op->FalseBlock.ID()).first->second;
+  PendingTargetLabel = JumpTarget(Op->FalseBlock);
 }
 
 DEF_OP(Syscall) {
@@ -438,48 +446,50 @@ DEF_OP(Thunk) {
 
 DEF_OP(ValidateCode) {
   auto Op = IROp->C<IR::IROp_ValidateCode>();
-  const auto* OldCode = (const uint8_t*)&Op->CodeOriginalLow;
+  auto OldCode = Op->CodeOriginal.data();
+  auto Base = GetReg(Op->Header.Args[0]).X();
   int len = Op->CodeLength;
-  int idx = 0;
-
-  LoadConstant(ARMEmitter::Size::i64Bit, GetReg(Node), 0);
-  LoadConstant(ARMEmitter::Size::i64Bit, TMP1, Entry + Op->Offset);
-  LoadConstant(ARMEmitter::Size::i64Bit, TMP2, 1);
+  int Offset = 0;
+  ARMEmitter::ForwardLabel Fail;
 
   const auto Dst = GetReg(Node);
 
-  while (len >= 8) {
-    ldr(ARMEmitter::XReg::x2, TMP1, idx);
-    LoadConstant(ARMEmitter::Size::i64Bit, TMP4, *(const uint64_t*)(OldCode + idx));
-    cmp(ARMEmitter::Size::i64Bit, TMP3, TMP4);
-    csel(ARMEmitter::Size::i64Bit, Dst, Dst, TMP2, ARMEmitter::Condition::CC_EQ);
-    len -= 8;
-    idx += 8;
-  }
-  while (len >= 4) {
-    ldr(ARMEmitter::WReg::w2, TMP1, idx);
-    LoadConstant(ARMEmitter::Size::i64Bit, TMP4, *(const uint32_t*)(OldCode + idx));
-    cmp(ARMEmitter::Size::i32Bit, TMP3, TMP4);
-    csel(ARMEmitter::Size::i64Bit, Dst, Dst, TMP2, ARMEmitter::Condition::CC_EQ);
-    len -= 4;
-    idx += 4;
-  }
-  while (len >= 2) {
-    ldrh(TMP3, TMP1, idx);
-    LoadConstant(ARMEmitter::Size::i64Bit, TMP4, *(const uint16_t*)(OldCode + idx));
-    cmp(ARMEmitter::Size::i32Bit, TMP3, TMP4);
-    csel(ARMEmitter::Size::i64Bit, Dst, Dst, TMP2, ARMEmitter::Condition::CC_EQ);
-    len -= 2;
-    idx += 2;
-  }
-  while (len >= 1) {
-    ldrb(TMP3, TMP1, idx);
-    LoadConstant(ARMEmitter::Size::i64Bit, TMP4, *(const uint8_t*)(OldCode + idx));
-    cmp(ARMEmitter::Size::i32Bit, TMP3, TMP4);
-    csel(ARMEmitter::Size::i64Bit, Dst, Dst, TMP2, ARMEmitter::Condition::CC_EQ);
-    len -= 1;
-    idx += 1;
-  }
+  auto EmitCheck = [&](size_t Size, auto&& LoadData) {
+    while (len >= Size) {
+      LoadData();
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, TMP2);
+      cbnz(ARMEmitter::Size::i64Bit, TMP1, &Fail);
+      len -= Size;
+      Offset += Size;
+    }
+  };
+
+  EmitCheck(8, [&]() {
+    ldr(TMP1, Base, Offset);
+    LoadConstant(ARMEmitter::Size::i64Bit, TMP2, *(const uint64_t*)(OldCode + Offset));
+  });
+
+  EmitCheck(4, [&]() {
+    ldr(TMP1.W(), Base, Offset);
+    LoadConstant(ARMEmitter::Size::i32Bit, TMP2, *(const uint32_t*)(OldCode + Offset));
+  });
+
+  EmitCheck(2, [&]() {
+    ldrh(TMP1.W(), Base, Offset);
+    LoadConstant(ARMEmitter::Size::i32Bit, TMP2, *(const uint16_t*)(OldCode + Offset));
+  });
+
+  EmitCheck(1, [&]() {
+    ldrb(TMP1.W(), Base, Offset);
+    LoadConstant(ARMEmitter::Size::i32Bit, TMP2, *(const uint8_t*)(OldCode + Offset));
+  });
+
+  ARMEmitter::ForwardLabel End;
+  LoadConstant(ARMEmitter::Size::i32Bit, Dst, 0);
+  b(&End);
+  Bind(&Fail);
+  LoadConstant(ARMEmitter::Size::i32Bit, Dst, 1);
+  Bind(&End);
 }
 
 DEF_OP(ThreadRemoveCodeEntry) {

FEXCore/Source/Interface/Core/JIT/EncryptionOps.cpp

@@ -16,7 +16,7 @@ DEF_OP(VAESImc) {
 
 DEF_OP(VAESEnc) {
   const auto Op = IROp->C<IR::IROp_VAESEnc>();
-  [[maybe_unused]] const auto OpSize = IROp->Size;
+  const auto OpSize = IROp->Size;
 
   const auto Dst = GetVReg(Node);
   const auto Key = GetVReg(Op->Key);
@@ -41,7 +41,7 @@ DEF_OP(VAESEnc) {
 
 DEF_OP(VAESEncLast) {
   const auto Op = IROp->C<IR::IROp_VAESEncLast>();
-  [[maybe_unused]] const auto OpSize = IROp->Size;
+  const auto OpSize = IROp->Size;
 
   const auto Dst = GetVReg(Node);
   const auto Key = GetVReg(Op->Key);
@@ -64,7 +64,7 @@ DEF_OP(VAESEncLast) {
 
 DEF_OP(VAESDec) {
   const auto Op = IROp->C<IR::IROp_VAESDec>();
-  [[maybe_unused]] const auto OpSize = IROp->Size;
+  const auto OpSize = IROp->Size;
 
   const auto Dst = GetVReg(Node);
   const auto Key = GetVReg(Op->Key);
@@ -89,7 +89,7 @@ DEF_OP(VAESDec) {
 
 DEF_OP(VAESDecLast) {
   const auto Op = IROp->C<IR::IROp_VAESDecLast>();
-  [[maybe_unused]] const auto OpSize = IROp->Size;
+  const auto OpSize = IROp->Size;
 
   const auto Dst = GetVReg(Node);
   const auto Key = GetVReg(Op->Key);
@@ -322,7 +322,7 @@ DEF_OP(VSha256U1) {
 
 DEF_OP(PCLMUL) {
   const auto Op = IROp->C<IR::IROp_PCLMUL>();
-  [[maybe_unused]] const auto OpSize = IROp->Size;
+  const auto OpSize = IROp->Size;
 
   const auto Dst = GetVReg(Node);
   const auto Src1 = GetVReg(Op->Src1);

FEXCore/Source/Interface/Core/JIT/JIT.cpp

@@ -222,7 +222,7 @@ void Arm64JITCore::Op_Unhandled(const IR::IROp_Header* IROp, IR::Ref Node) {
       FillF64Result();
     } break;
 
-    case FABI_F64_I16_F64_PTR: {
+    case FABI_F64_F64_PTR: {
       // Linux Reg/Win32 Reg:
       // tmp4 (x4/x13): FallbackHandler
       // x30: return
@@ -239,7 +239,7 @@ void Arm64JITCore::Op_Unhandled(const IR::IROp_Header* IROp, IR::Ref Node) {
       ldr<ARMEmitter::IndexType::POST>(ARMEmitter::XReg::lr, ARMEmitter::Reg::rsp, 16);
       FillF64Result();
     } break;
-    case FABI_F64x2_I16_F64_PTR: {
+    case FABI_F64x2_F64_PTR: {
       // Linux Reg/Win32 Reg:
       // tmp4 (x4/x13): FallbackHandler
       // x30: return
@@ -264,7 +264,7 @@ void Arm64JITCore::Op_Unhandled(const IR::IROp_Header* IROp, IR::Ref Node) {
       FillF64x2Result(DstLo, DstHi);
     } break;
 
-    case FABI_F64_I16_F64_F64_PTR: {
+    case FABI_F64_F64_F64_PTR: {
       // Linux Reg/Win32 Reg:
       // tmp4 (x4/x13): FallbackHandler
       // x30: return
@@ -639,6 +639,7 @@ Arm64JITCore::Arm64JITCore(FEXCore::Context::ContextImpl* ctx, FEXCore::Core::In
     Common.PrintValue = reinterpret_cast<uint64_t>(PrintValue);
     Common.PrintVectorValue = reinterpret_cast<uint64_t>(PrintVectorValue);
     Common.ThreadRemoveCodeEntryFromJIT = reinterpret_cast<uintptr_t>(&Context::ContextImpl::ThreadRemoveCodeEntryFromJit);
+    Common.MonoBackpatcherWrite = reinterpret_cast<uint64_t>(&Context::ContextImpl::MonoBackpatcherWrite);
     Common.CPUIDObj = reinterpret_cast<uint64_t>(&CTX->CPUID);
 
     {
@@ -827,6 +828,7 @@ CPUBackend::CompiledCode Arm64JITCore::CompileCode(uint64_t Entry, uint64_t Size
   CallReturnTargets.clear();
   PendingJumpThunks.clear();
   uint32_t SSACount = IR->GetSSACount();
+  JumpTargets.resize(IR->GetHeader()->BlockCount, {});
 
   this->Entry = Entry;
   this->DebugData = DebugData;
@@ -886,10 +888,10 @@ CPUBackend::CompiledCode Arm64JITCore::CompileCode(uint64_t Entry, uint64_t Size
     auto BlockStartHostCode = GetCursorAddress<uint8_t*>();
     {
       const auto Node = IR->GetID(BlockNode);
-      const auto IsTarget = JumpTargets.try_emplace(Node).first;
+      const auto Target = &JumpTargets[BlockIROp->ID];
 
       // if there's a pending branch, and it is not fall-through
-      if (PendingTargetLabel && PendingTargetLabel != &IsTarget->second) {
+      if (PendingTargetLabel && PendingTargetLabel != Target) {
         b(PendingTargetLabel);
         PendingTargetLabel = nullptr;
       }
@@ -900,7 +902,7 @@ CPUBackend::CompiledCode Arm64JITCore::CompileCode(uint64_t Entry, uint64_t Size
         const auto IsReturnTarget = CallReturnTargets.try_emplace(Node).first;
         if (PendingTargetLabel) {
           // If there is a fallthrough branch to this block, skip over the entrypoint code.
-          b(&IsTarget->second);
+          b(Target);
         } else if (PendingCallReturnTargetLabel && PendingCallReturnTargetLabel != &IsReturnTarget->second) {
           // If we just emitted a call, but the block we're now emitting is not the return block so don't fallthrough.
           b(PendingCallReturnTargetLabel);
@@ -921,7 +923,7 @@ CPUBackend::CompiledCode Arm64JITCore::CompileCode(uint64_t Entry, uint64_t Size
       }
       PendingTargetLabel = nullptr;
 
-      Bind(&IsTarget->second);
+      Bind(Target);
     }
 
     for (auto [CodeNode, IROp] : IR->GetCode(BlockNode)) {

FEXCore/Source/Interface/Core/JIT/JITClass.h

@@ -24,6 +24,7 @@ $end_info$
 
 #include <array>
 #include <cstdint>
+#include <functional>
 #include <utility>
 #include <variant>
 
@@ -67,7 +68,13 @@ private:
   uint64_t Entry {};
   CPUBackend::CompiledCode CodeData {};
 
-  fextl::map<IR::NodeID, ARMEmitter::BiDirectionalLabel> JumpTargets;
+  fextl::vector<ARMEmitter::BiDirectionalLabel> JumpTargets;
+
+  ARMEmitter::BiDirectionalLabel* JumpTarget(IR::OrderedNodeWrapper Node) {
+    auto Block = IR->GetOp<IR::IROp_CodeBlock>(Node);
+    return &JumpTargets[Block->ID];
+  }
+
   fextl::map<IR::NodeID, ARMEmitter::BiDirectionalLabel> CallReturnTargets;
 
   struct PendingJumpThunk {
@@ -238,10 +245,10 @@ private:
     case FEXCore::IR::COND_FGE: return ARMEmitter::Condition::CC_GE;
     case FEXCore::IR::COND_FLEU: return ARMEmitter::Condition::CC_LE;
     case FEXCore::IR::COND_FGT: return ARMEmitter::Condition::CC_GT;
-    case FEXCore::IR::COND_FU: return ARMEmitter::Condition::CC_VS;
-    case FEXCore::IR::COND_FNU: return ARMEmitter::Condition::CC_VC;
-    case FEXCore::IR::COND_VS:
-    case FEXCore::IR::COND_VC:
+    case FEXCore::IR::COND_FU:
+    case FEXCore::IR::COND_VS: return ARMEmitter::Condition::CC_VS;
+    case FEXCore::IR::COND_FNU:
+    case FEXCore::IR::COND_VC: return ARMEmitter::Condition::CC_VC;
     case FEXCore::IR::COND_MI: return ARMEmitter::Condition::CC_MI;
     case FEXCore::IR::COND_PL: return ARMEmitter::Condition::CC_PL;
     default: LOGMAN_MSG_A_FMT("Unsupported compare type"); return ARMEmitter::Condition::CC_NV;

FEXCore/Source/Interface/Core/JIT/MemoryOps.cpp

@@ -140,7 +140,7 @@ DEF_OP(LoadRegister) {
 
     mov(GetReg(Node).X(), StaticRegisters[Op->Reg].X());
   } else if (Op->Class == IR::FPRClass) {
-    [[maybe_unused]] const auto regSize = HostSupportsAVX256 ? IR::OpSize::i256Bit : IR::OpSize::i128Bit;
+    const auto regSize = HostSupportsAVX256 ? IR::OpSize::i256Bit : IR::OpSize::i128Bit;
     LOGMAN_THROW_A_FMT(Op->Reg < StaticFPRegisters.size(), "out of range reg");
     LOGMAN_THROW_A_FMT(IROp->Size == regSize, "expected sized");
 
@@ -181,7 +181,7 @@ DEF_OP(StoreRegister) {
     // Always use 64-bit, it's faster. Upper bits ignored for 32-bit mode.
     mov(ARMEmitter::Size::i64Bit, GetReg(Reg), GetReg(Op->Value));
   } else if (Reg.Class == IR::FPRFixedClass) {
-    [[maybe_unused]] const auto regSize = HostSupportsAVX256 ? IR::OpSize::i256Bit : IR::OpSize::i128Bit;
+    const auto regSize = HostSupportsAVX256 ? IR::OpSize::i256Bit : IR::OpSize::i128Bit;
     LOGMAN_THROW_A_FMT(IROp->Size == regSize, "expected sized");
 
     const auto guest = GetVReg(Reg);
@@ -742,7 +742,7 @@ DEF_OP(LoadMemTSO) {
     const auto Dst = GetReg(Node);
     uint64_t Offset = 0;
     if (!Op->Offset.IsInvalid()) {
-      [[maybe_unused]] bool IsInline = IsInlineConstant(Op->Offset, &Offset);
+      bool IsInline = IsInlineConstant(Op->Offset, &Offset);
       LOGMAN_THROW_A_FMT(IsInline, "expected immediate");
     }
 
@@ -1750,7 +1750,7 @@ DEF_OP(StoreMemTSO) {
     const auto Src = GetZeroableReg(Op->Value);
     uint64_t Offset = 0;
     if (!Op->Offset.IsInvalid()) {
-      [[maybe_unused]] bool IsInline = IsInlineConstant(Op->Offset, &Offset);
+      bool IsInline = IsInlineConstant(Op->Offset, &Offset);
       LOGMAN_THROW_A_FMT(IsInline, "expected immediate");
     }
 
@@ -2586,7 +2586,7 @@ DEF_OP(VStoreNonTemporalPair) {
   const auto Op = IROp->C<IR::IROp_VStoreNonTemporalPair>();
   const auto OpSize = IROp->Size;
 
-  [[maybe_unused]] const auto Is128Bit = OpSize == IR::OpSize::i128Bit;
+  const auto Is128Bit = OpSize == IR::OpSize::i128Bit;
   LOGMAN_THROW_A_FMT(Is128Bit, "This IR operation only operates at 128-bit wide");
 
   const auto ValueLow = GetVReg(Op->ValueLow);

FEXCore/Source/Interface/Core/JIT/MiscOps.cpp

@@ -286,4 +286,43 @@ DEF_OP(Yield) {
   yield();
 }
 
+DEF_OP(MonoBackpatcherWrite) {
+  auto Op = IROp->C<IR::IROp_MonoBackpatcherWrite>();
+
+  mov(ARMEmitter::Size::i64Bit, TMP3, GetReg(Op->Addr));
+  mov(ARMEmitter::Size::i64Bit, TMP4, GetReg(Op->Value));
+
+  PushDynamicRegs(TMP1);
+  SpillStaticRegs(TMP1);
+
+  mov(ARMEmitter::Size::i64Bit, ARMEmitter::Reg::r0, STATE.R());
+  mov(ARMEmitter::Size::i64Bit, ARMEmitter::Reg::r1, IR::OpSizeToSize(Op->Size));
+
+  if (!TMP_ABIARGS) {
+    mov(ARMEmitter::Size::i64Bit, ARMEmitter::Reg::r2, TMP3);
+    mov(ARMEmitter::Size::i64Bit, ARMEmitter::Reg::r3, TMP4);
+  }
+
+#ifdef _M_ARM_64EC
+  ldr(TMP2, ARMEmitter::XReg::x18, TEB_CPU_AREA_OFFSET);
+  LoadConstant(ARMEmitter::Size::i32Bit, TMP1, 1);
+  strb(TMP1.W(), TMP2, CPU_AREA_IN_SYSCALL_CALLBACK_OFFSET);
+#endif
+
+  ldr(ARMEmitter::XReg::x4, STATE, offsetof(FEXCore::Core::CpuStateFrame, Pointers.Common.MonoBackpatcherWrite));
+  if (!CTX->Config.DisableVixlIndirectCalls) [[unlikely]] {
+    GenerateIndirectRuntimeCall<void, void*, uint8_t, uint64_t, uint64_t>(ARMEmitter::Reg::r4);
+  } else {
+    blr(ARMEmitter::Reg::r4);
+  }
+
+#ifdef _M_ARM_64EC
+  ldr(TMP2, ARMEmitter::XReg::x18, TEB_CPU_AREA_OFFSET);
+  strb(ARMEmitter::WReg::zr, TMP2, CPU_AREA_IN_SYSCALL_CALLBACK_OFFSET);
+#endif
+
+  FillStaticRegs();
+  PopDynamicRegs();
+}
+
 } // namespace FEXCore::CPU

FEXCore/Source/Interface/Core/JIT/VectorOps.cpp

@@ -1352,7 +1352,7 @@ DEF_OP(VFMin) {
 
   const auto ElementSize = Op->Header.ElementSize;
   const auto SubRegSize = ConvertSubRegSize248(IROp);
-  [[maybe_unused]] const auto IsScalar = ElementSize == OpSize;
+  const auto IsScalar = ElementSize == OpSize;
   const auto Is256Bit = OpSize == IR::OpSize::i256Bit;
   LOGMAN_THROW_A_FMT(!Is256Bit || HostSupportsSVE256, "Need SVE256 support in order to use {} with 256-bit operation", __func__);
 
@@ -1425,7 +1425,7 @@ DEF_OP(VFMax) {
 
   const auto ElementSize = Op->Header.ElementSize;
   const auto SubRegSize = ConvertSubRegSize248(IROp);
-  [[maybe_unused]] const auto IsScalar = ElementSize == OpSize;
+  const auto IsScalar = ElementSize == OpSize;
   const auto Is256Bit = OpSize == IR::OpSize::i256Bit;
   LOGMAN_THROW_A_FMT(!Is256Bit || HostSupportsSVE256, "Need SVE256 support in order to use {} with 256-bit operation", __func__);
 

FEXCore/Source/Interface/Core/ObjectCache/CodeObjectSerializationConfig.h

@@ -1,86 +0,0 @@
-// SPDX-License-Identifier: MIT
-#pragma once
-
-#include <FEXCore/Utils/CompilerDefs.h>
-
-#include <cstdint>
-
-namespace FEXCore::CodeSerialize {
-// If any of the config options mismatch on load then the cache won't be used
-// Any of these will result in codegen changes
-struct FEX_PACKED CodeObjectSerializationConfig {
-  // Cookie in the header of the file, isn't part of the config hash
-  uint64_t Cookie {};
-
-  // Instructions per block configuration
-  int32_t MaxInstPerBlock {};
-
-  // Follows CPUID 4000_0001_EAX[3:0]
-  unsigned Arch : 4;
-
-  // Multiblock enabled
-  unsigned MultiBlock : 1;
-
-  // Hardware TSO enabled
-  unsigned HardwareTSOEnabled : 1;
-
-  // TSO enabled
-  unsigned TSOEnabled : 1;
-
-  // ABI local flag unsafe optimization
-  unsigned ABILocalFlags : 1;
-
-  // Paranoid TSO mode enabled
-  unsigned ParanoidTSO : 1;
-
-  // Guest code execution mode (We don't support live mode switch)
-  unsigned Is64BitMode : 1;
-
-  // SMC checks style
-  unsigned SMCChecks : 2;
-
-  // x87 reduced precision
-  unsigned x87ReducedPrecision : 1;
-
-  // Padding to remove uninitialized data warning from asan
-  // Shows remaining amount of bits available for config
-  unsigned _Pad : 19;
-
-  bool operator==(const CodeObjectSerializationConfig& other) const {
-    return Cookie == other.Cookie && MaxInstPerBlock == other.MaxInstPerBlock && Arch == other.Arch && MultiBlock == other.MultiBlock &&
-           HardwareTSOEnabled == other.HardwareTSOEnabled && TSOEnabled == other.TSOEnabled && ABILocalFlags == other.ABILocalFlags &&
-           ParanoidTSO == other.ParanoidTSO && Is64BitMode == other.Is64BitMode && SMCChecks == other.SMCChecks &&
-           x87ReducedPrecision == other.x87ReducedPrecision;
-  }
-  static uint64_t GetHash(const CodeObjectSerializationConfig& other) {
-    // For < 64-bits of data just pack directly
-    // Skip the cookie
-    uint64_t Hash {};
-    Hash <<= 32;
-    Hash |= other.MaxInstPerBlock;
-    Hash <<= 1;
-    Hash |= other.Arch;
-    Hash <<= 1;
-    Hash |= other.MultiBlock;
-    Hash <<= 1;
-    Hash |= other.HardwareTSOEnabled;
-    Hash <<= 1;
-    Hash |= other.TSOEnabled;
-    Hash <<= 1;
-    Hash |= other.ABILocalFlags;
-    Hash <<= 1;
-    Hash |= other.ParanoidTSO;
-    Hash <<= 1;
-    Hash |= other.Is64BitMode;
-    Hash <<= 2;
-    Hash |= other.SMCChecks;
-    Hash <<= 1;
-    Hash |= other.x87ReducedPrecision;
-    return Hash;
-  }
-};
-
-static_assert(sizeof(CodeObjectSerializationConfig) == 16, "Size changed");
-static_assert((sizeof(CodeObjectSerializationConfig) - sizeof(uint64_t)) == 8, "Config size exceeded 64its. Need to change how the hash is "
-                                                                               "generated!");
-} // namespace FEXCore::CodeSerialize

FEXCore/Source/Interface/Core/ObjectCache/JobHandling.cpp

@@ -1,121 +0,0 @@
-// SPDX-License-Identifier: MIT
-#include "Interface/Core/ObjectCache/ObjectCacheService.h"
-
-#include <FEXCore/Config/Config.h>
-#include <FEXCore/fextl/memory.h>
-#include <FEXCore/fextl/string.h>
-#include <FEXHeaderUtils/Filesystem.h>
-
-#include <fcntl.h>
-#include <xxhash.h>
-
-namespace FEXCore::CodeSerialize {
-void AsyncJobHandler::AsyncAddNamedRegionJob(uintptr_t Base, uintptr_t Size, uintptr_t Offset, const fextl::string& filename) {
-#ifndef _WIN32
-  // This function adds a named region *JOB* to our named region handler
-  // This needs to be as fast as possible to keep out of the way of the JIT
-
-  const fextl::string BaseFilename = FHU::Filesystem::GetFilename(filename);
-
-  if (!BaseFilename.empty()) {
-    // Create a new entry that once set up will be put in to our section object map
-    auto Entry = fextl::make_unique<CodeRegionEntry>(Base, Size, Offset, filename, NamedRegionHandler->DefaultCodeHeader(Base, Offset));
-
-    // Lock the job ref counter so we can block anything attempting to use the entry before it is loaded
-    Entry->NamedJobRefCountMutex.lock();
-
-    CodeRegionMapType::iterator EntryIterator;
-    {
-      std::unique_lock lk {CodeObjectCacheService->GetEntryMapMutex()};
-
-      auto& EntryMap = CodeObjectCacheService->GetEntryMap();
-
-      auto it = EntryMap.emplace(Base, std::move(Entry));
-      if (!it.second) {
-        // This happens when an application overwrites a previous region without unmapping what was there
-
-        // Lock this entry's Named job reference counter.
-        // Once this passes then we know that this section has been loaded.
-        it.first->second->NamedJobRefCountMutex.lock();
-
-        // Finalize anything the region needs to do first.
-        CodeObjectCacheService->DoCodeRegionClosure(it.first->second->Base, it.first->second.get());
-
-        // munmap the file that was mapped
-        FEXCore::Allocator::munmap(it.first->second->CodeData, it.first->second->FileSize);
-
-        // Remove this entry from the unrelocated map as well
-        {
-          std::unique_lock lk2 {CodeObjectCacheService->GetUnrelocatedEntryMapMutex()};
-          CodeObjectCacheService->GetUnrelocatedEntryMap().erase(it.first->second->EntryHeader.OriginalBase);
-        }
-
-        // Now overwrite the entry in the map
-        it = EntryMap.insert_or_assign(Base, std::move(Entry));
-        EntryIterator = it.first;
-      } else {
-        // No overwrite, just insert
-        EntryIterator = it.first;
-      }
-    }
-
-    // Now that this entry has been added to the map, we can insert a load job using the entry iterator.
-    // This allows us to quickly unblock the JIT thread when it is loading multiple regions and have the async thread
-    // do the loading for us.
-    //
-    // Create the async work queue job now so it can load
-    NamedRegionHandler->AsyncAddNamedRegionWorkItem(BaseFilename, filename, true, EntryIterator);
-
-    // Tell the async thread that it has work to do
-    CodeObjectCacheService->NotifyWork();
-  }
-#endif
-}
-
-void AsyncJobHandler::AsyncRemoveNamedRegionJob(uintptr_t Base, uintptr_t Size) {
-#ifndef _WIN32
-  // Removing a named region through the job system
-  // We need to find the entry that we are deleting first
-  fextl::unique_ptr<CodeRegionEntry> EntryPointer;
-  {
-    std::unique_lock lk {CodeObjectCacheService->GetEntryMapMutex()};
-
-    auto& EntryMap = CodeObjectCacheService->GetEntryMap();
-    auto it = EntryMap.find(Base);
-    if (it != EntryMap.end()) {
-      // Lock the job ref counter since we are erasing it
-      // Once this passes it will have been loaded
-      it->second->NamedJobRefCountMutex.lock();
-
-      // Take the pointer from the map
-      EntryPointer = std::move(it->second);
-
-      // We can now unmap the file data
-      FEXCore::Allocator::munmap(EntryPointer->CodeData, EntryPointer->FileSize);
-
-      // Remove this from the entry map
-      EntryMap.erase(it);
-
-      // Remove this entry from the unrelocated map as well
-      {
-        std::unique_lock lk2 {CodeObjectCacheService->GetUnrelocatedEntryMapMutex()};
-        CodeObjectCacheService->GetUnrelocatedEntryMap().erase(EntryPointer->EntryHeader.OriginalBase);
-      }
-    } else {
-      // Tried to remove something that wasn't in our code object tracking
-      return;
-    }
-
-    // Create the async work queue job now so it can finalize what it needs to do
-    NamedRegionHandler->AsyncRemoveNamedRegionWorkItem(Base, Size, std::move(EntryPointer));
-
-    // Tell the async thread that it has work to do
-    CodeObjectCacheService->NotifyWork();
-  }
-#endif
-}
-
-void AsyncJobHandler::AsyncAddSerializationJob(fextl::unique_ptr<SerializationJobData> Data) {
-  // XXX: Actually add serialization job
-}
-} // namespace FEXCore::CodeSerialize

FEXCore/Source/Interface/Core/ObjectCache/NamedRegionObjectHandler.cpp

@@ -1,71 +0,0 @@
-// SPDX-License-Identifier: MIT
-#include "Interface/Core/ObjectCache/ObjectCacheService.h"
-
-#include <FEXCore/fextl/memory.h>
-#include <FEXCore/fextl/string.h>
-
-namespace FEXCore::CodeSerialize {
-NamedRegionObjectHandler::NamedRegionObjectHandler(FEXCore::Context::ContextImpl* ctx) {
-  DefaultSerializationConfig.Cookie = CODE_COOKIE;
-
-  // Initialize the Arch from CPUID
-  uint32_t Arch = ctx->CPUID.RunFunction(0x4000'0001, 0).eax & 0xF;
-  DefaultSerializationConfig.Arch = Arch;
-
-  DefaultSerializationConfig.MaxInstPerBlock = ctx->Config.MaxInstPerBlock;
-  DefaultSerializationConfig.MultiBlock = ctx->Config.Multiblock;
-  DefaultSerializationConfig.TSOEnabled = ctx->Config.TSOEnabled;
-  DefaultSerializationConfig.ABILocalFlags = ctx->Config.ABILocalFlags;
-  DefaultSerializationConfig.ParanoidTSO = ctx->Config.ParanoidTSO;
-  DefaultSerializationConfig.Is64BitMode = ctx->Config.Is64BitMode;
-  DefaultSerializationConfig.SMCChecks = ctx->Config.SMCChecks;
-  DefaultSerializationConfig.x87ReducedPrecision = ctx->Config.x87ReducedPrecision;
-}
-
-void NamedRegionObjectHandler::AddNamedRegionObject(CodeRegionMapType::iterator Entry, const fextl::string& base_filename,
-                                                    const fextl::string& filename, bool Executable) {
-  // XXX: Add named region objects
-
-  // XXX: Until entry loading is complete just claim it is loaded
-  Entry->second->NamedJobRefCountMutex.unlock();
-}
-
-void NamedRegionObjectHandler::RemoveNamedRegionObject(uintptr_t Base, uintptr_t Size, fextl::unique_ptr<CodeRegionEntry> Entry) {
-  // XXX: Remove named region objects
-
-  // XXX: Until entry loading is complete just claim it is loaded
-  Entry->NamedJobRefCountMutex.unlock();
-}
-
-void NamedRegionObjectHandler::HandleNamedRegionObjectJobs() {
-  // Walk through all of our jobs sequentially until the work queue is empty
-  while (NamedWorkQueueJobs.load()) {
-    fextl::unique_ptr<AsyncJobHandler::NamedRegionWorkItem> WorkItem;
-
-    {
-      // Lock the work queue mutex for a short moment and grab an item from the list
-      std::unique_lock lk {NamedWorkQueueMutex};
-      size_t WorkItems = WorkQueue.size();
-      if (WorkItems != 0) {
-        WorkItem = std::move(WorkQueue.front());
-        WorkQueue.pop();
-      }
-
-      // Atomically update the number of jobs
-      --NamedWorkQueueJobs;
-    }
-
-    if (WorkItem) {
-      if (WorkItem->GetType() == AsyncJobHandler::NamedRegionJobType::JOB_ADD_NAMED_REGION) {
-        auto WorkAdd = static_cast<AsyncJobHandler::WorkItemAddNamedRegion*>(WorkItem.get());
-        AddNamedRegionObject(WorkAdd->Entry, WorkAdd->BaseFilename, WorkAdd->Filename, WorkAdd->Executable);
-      }
-
-      if (WorkItem->GetType() == AsyncJobHandler::NamedRegionJobType::JOB_REMOVE_NAMED_REGION) {
-        auto WorkRemove = static_cast<AsyncJobHandler::WorkItemRemoveNamedRegion*>(WorkItem.get());
-        RemoveNamedRegionObject(WorkRemove->Base, WorkRemove->Size, std::move(WorkRemove->Entry));
-      }
-    }
-  }
-}
-} // namespace FEXCore::CodeSerialize

FEXCore/Source/Interface/Core/ObjectCache/ObjectCacheService.cpp

@@ -1,85 +0,0 @@
-// SPDX-License-Identifier: MIT
-#include "Interface/Core/ObjectCache/ObjectCacheService.h"
-
-#include <FEXCore/Config/Config.h>
-#include <FEXCore/fextl/memory.h>
-#include <FEXCore/Utils/Threads.h>
-
-namespace {
-static void* ThreadHandler(void* Arg) {
-  FEXCore::CodeSerialize::CodeObjectSerializeService* This = reinterpret_cast<FEXCore::CodeSerialize::CodeObjectSerializeService*>(Arg);
-  This->ExecutionThread();
-  return nullptr;
-}
-} // namespace
-
-namespace FEXCore::CodeSerialize {
-CodeObjectSerializeService::CodeObjectSerializeService(FEXCore::Context::ContextImpl* ctx)
-  : CTX {ctx}
-  , AsyncHandler {&NamedRegionHandler, this}
-  , NamedRegionHandler {ctx} {
-  Initialize();
-}
-
-void CodeObjectSerializeService::Shutdown() {
-  if (CTX->Config.CacheObjectCodeCompilation() == FEXCore::Config::ConfigObjectCodeHandler::CONFIG_NONE) {
-    return;
-  }
-
-  WorkerThreadShuttingDown = true;
-
-  // Kick the working thread
-  WorkAvailable.NotifyAll();
-
-  if (WorkerThread->joinable()) {
-    // Wait for worker thread to close down
-    WorkerThread->join(nullptr);
-  }
-}
-
-void CodeObjectSerializeService::Initialize() {
-  // Add a canary so we don't crash on empty map iterator handling
-  auto it = AddressToEntryMap.insert_or_assign(~0ULL, fextl::make_unique<CodeRegionEntry>());
-  UnrelocatedAddressToEntryMap.insert_or_assign(~0ULL, it.first->second.get());
-
-  uint64_t OldMask = FEXCore::Threads::SetSignalMask(~0ULL);
-  WorkerThread = FEXCore::Threads::Thread::Create(ThreadHandler, this);
-  FEXCore::Threads::SetSignalMask(OldMask);
-}
-
-void CodeObjectSerializeService::DoCodeRegionClosure(uint64_t Base, CodeRegionEntry* it) {
-  if (Base == ~0ULL) {
-    // Don't do closure on canary
-    return;
-  }
-  // XXX: Do code region closure
-}
-
-const CodeObjectFileSection* CodeObjectSerializeService::FetchCodeObjectFromCache(uint64_t GuestRIP) {
-  // XXX: Actually fetch code objects from cache
-  return nullptr;
-}
-
-void CodeObjectSerializeService::ExecutionThread() {
-  // Set our thread name so we can see its relation
-  FEXCore::Threads::SetThreadName("ObjectCodeSeri\0");
-  while (WorkerThreadShuttingDown.load() != true) {
-    // Wait for work
-    WorkAvailable.Wait();
-
-    // Handle named region async jobs first. Highest priority
-    NamedRegionHandler.HandleNamedRegionObjectJobs();
-
-    // XXX: Handle code serialization jobs second.
-  }
-
-  // Do final code region closures on thread shutdown
-  for (auto& it : AddressToEntryMap) {
-    DoCodeRegionClosure(it.first, it.second.get());
-  }
-
-  // Safely clear our maps now
-  AddressToEntryMap.clear();
-  UnrelocatedAddressToEntryMap.clear();
-}
-} // namespace FEXCore::CodeSerialize

FEXCore/Source/Interface/Core/ObjectCache/ObjectCacheService.h

@@ -1,457 +0,0 @@
-// SPDX-License-Identifier: MIT
-#pragma once
-#include "Interface/Context/Context.h"
-#include "Interface/Core/ObjectCache/Relocations.h"
-#include "Interface/Core/ObjectCache/CodeObjectSerializationConfig.h"
-#include "Interface/IR/AOTIR.h"
-
-#include <FEXCore/Utils/Event.h>
-#include <FEXCore/Utils/Threads.h>
-#include <FEXCore/fextl/map.h>
-#include <FEXCore/fextl/memory.h>
-#include <FEXCore/fextl/queue.h>
-#include <FEXCore/fextl/robin_map.h>
-#include <FEXCore/fextl/string.h>
-#include <FEXCore/fextl/vector.h>
-
-#include <shared_mutex>
-
-namespace FEXCore::CodeSerialize {
-// XXX: Does this need to be signal safe?
-using CodeSerializationMutex = std::shared_mutex;
-struct CodeSerializationData {};
-
-struct CodeObjectFileSection {
-  bool Serialized;
-  bool Invalid;
-  const CodeSerializationData* Data;
-  const char* HostCode;
-  uint64_t NumRelocations;
-  const char* Relocations;
-};
-
-/**
- * @brief This is the file header that lives at the start of an object cache file
- *
- * This header is updated from multiple processes!
- * Care must be taken to use OS locks when updating the file backing including this header
- */
-struct CodeObjectSerializationHeader {
-  // The configuration that this file has
-  CodeObjectSerializationConfig Config;
-  // The original RIP that this object section was mapped at
-  uint64_t OriginalBase {};
-  // The original offset in to the file that this object section was loaded from
-  uint64_t OriginalOffset {};
-  // Total amount of code that should be in this file
-  uint64_t TotalCodeSize {};
-  // Used to reserve the TSL map
-  uint64_t NumCodeEntries {};
-  // The number of relocations that point to this section
-  uint64_t NumRelocationsTo {};
-  // Total relocations in this file
-  uint64_t TotalRelocationsCount {};
-};
-
-struct CodeRegionEntry {
-  /**
-   * @name Threaded initialization objects for the initial object creation
-   * @{ */
-  // Base address in memory where the code region is at
-  uint64_t Base {};
-
-  // Size of this code entry
-  uint64_t Size {};
-
-  // The offset inside the file that is mapped to Base
-  uint64_t Offset {};
-
-  // Filename of the object
-  fextl::string Filename {};
-
-  CodeObjectSerializationHeader EntryHeader {};
-  /**  @} */
-
-  // The filename of the object cache for this entry
-  fextl::string ObjectEntrySourceFilename {};
-
-  // In the case of file corruption that we can detect, we can disable serialization early for an entry
-  // We should be resiliant to corruption but things happen
-  bool StillSerializing {true};
-
-  // Long lived FD for serialization if we have multiple jobs to serialize
-  // Bursts of code entries are common and this reduces file lock overhead
-  //
-  // Especially useful over network mounts where file locks are very slow
-  int CurrentSerializedFD {-1};
-
-  /**
-   * @name Objects required to sync objects between threads
-   * @{ */
-  // Refcount for the number of outstanding code entries waiting to be written for this object section
-  CodeSerializationMutex ObjectJobRefCountMutex;
-
-  // Refcount for outstanding named object region entry loading itself
-  // Will block JIT code cache look up when this has a unique_lock held
-  CodeSerializationMutex NamedJobRefCountMutex;
-  /**  @} */
-
-  /**
-   * @name Object Entry data management
-   * @{ */
-
-  /**
-   * @name This is the raw file data that we loaded from the code region entry file
-   * @{ */
-  char* CodeData {};
-  size_t FileSize {};
-
-  fextl::vector<CodeObjectFileSection> FileCodeSections;
-  /**  @} */
-
-  // This per section map takes the most time to load and needs to be quick
-  // This is the map of all code segments for this entry
-  fextl::robin_map<uint64_t, CodeObjectFileSection*> SectionLookupMap {};
-  /**  @} */
-
-  // Default initialization
-  CodeRegionEntry() = default;
-
-  // Initializer specifically for threaded loading
-  CodeRegionEntry(uint64_t Base, uint64_t Size, uint64_t Offset, const fextl::string& Filename, const CodeObjectSerializationHeader& DefaultHeader)
-    : Base {Base}
-    , Size {Size}
-    , Offset {Offset}
-    , Filename {Filename}
-    , EntryHeader {DefaultHeader} {}
-};
-
-// Map type must use an interator that isn't invalidation on erase/insert
-using CodeRegionMapType = fextl::map<uint64_t, fextl::unique_ptr<CodeRegionEntry>>;
-using CodeRegionPtrMapType = fextl::map<uint64_t, CodeRegionEntry*>;
-
-class NamedRegionObjectHandler;
-class CodeObjectSerializeService;
-
-class AsyncJobHandler final {
-public:
-  /**
-   * @brief Structure containing all the data required to async serialize code objects
-   */
-  struct SerializationJobData {
-    uint64_t GuestRIP; ///< The RIP for the guest
-    // XXX: Support multiblock
-    uint64_t GuestCodeLength; ///< The Guest's code length
-    uint64_t GuestCodeHash;   ///< Hash of the guest code
-
-    void* HostCodeBegin;   ///< Host JIT code starting memory address
-    size_t HostCodeLength; ///< Host JIT code length
-    uint64_t HostCodeHash; ///< Host JIT code hash before any backpatching
-
-    // This is the thread specific ref counter for outstanding jobs.
-    // This shared mutex is incremented when the job is added, then decremented when the job is complete.
-    // If a thread is shutting down or clearing code cache then the thread will pull a unique lock on this mutex.
-    // This way it will wait until the async job handler is complete with it.
-    CodeSerializationMutex* ThreadJobRefCount;
-
-    // These are the reolocations for this serialization job
-    // Relatively small number of entries most of the time
-    fextl::vector<FEXCore::CPU::Relocation> Relocations;
-
-    /**
-     * @name Objects filled in from the Code Object Serialization service when a job is added
-     * @{ */
-    // This is the code region's ref counter for outstanding jobs.
-    // This shared mutex is incremented when the job is added, then decremented when the job is complete.
-    // If a named region is being removed then a unique lock will be pulled to wait for all jobs to complete and no new jobs to be added.
-    CodeSerializationMutex* ObjectJobRefCountMutexPtr;
-
-    // This is the code region iterator to reduce the number of map lookups
-    // This will remain valid while jobs are outstanding for this region
-    CodeRegionMapType::iterator CodeRegionIterator;
-    /**  @} */
-  };
-
-  AsyncJobHandler(NamedRegionObjectHandler* NamedRegionHandler, CodeObjectSerializeService* CodeObjectCacheService)
-    : NamedRegionHandler {NamedRegionHandler}
-    , CodeObjectCacheService {CodeObjectCacheService} {}
-
-protected:
-  friend class CodeObjectSerializeService;
-  friend class NamedRegionObjectHandler;
-  /**
-   * @name Async job submission functions
-   * @{ */
-  void AsyncAddNamedRegionJob(uintptr_t Base, uintptr_t Size, uintptr_t Offset, const fextl::string& filename);
-  void AsyncRemoveNamedRegionJob(uintptr_t Base, uintptr_t Size);
-  void AsyncAddSerializationJob(fextl::unique_ptr<SerializationJobData> Data);
-  /**  @} */
-
-  /**
-   * @name Async named region handling
-   * @{ */
-  /**
-   * @brief The async named region jobs to handle.
-   *
-   * Only two, Code serialization goes in to a different queue.
-   */
-  enum class NamedRegionJobType {
-    JOB_ADD_NAMED_REGION,
-    JOB_REMOVE_NAMED_REGION,
-  };
-
-  class NamedRegionWorkItem {
-  public:
-    NamedRegionJobType GetType() const {
-      return Type;
-    }
-
-  protected:
-    friend class WorkItemAddNamedRegion;
-    NamedRegionWorkItem(NamedRegionJobType type)
-      : Type {type} {}
-
-  private:
-    NamedRegionJobType Type;
-  };
-
-  class WorkItemAddNamedRegion : public NamedRegionWorkItem {
-  public:
-    WorkItemAddNamedRegion(const fextl::string& base, const fextl::string& filename, bool executable, CodeRegionMapType::iterator entry)
-      : NamedRegionWorkItem {NamedRegionJobType::JOB_ADD_NAMED_REGION}
-      , BaseFilename {base}
-      , Filename {filename}
-      , Executable {executable}
-      , Entry {entry} {}
-    const fextl::string BaseFilename;
-    const fextl::string Filename;
-    bool Executable;
-    CodeRegionMapType::iterator Entry;
-  };
-
-  class WorkItemRemoveNamedRegion : public NamedRegionWorkItem {
-  public:
-    WorkItemRemoveNamedRegion(uint64_t base, uint64_t size, fextl::unique_ptr<CodeRegionEntry> entry)
-      : NamedRegionWorkItem {NamedRegionJobType::JOB_REMOVE_NAMED_REGION}
-      , Base {base}
-      , Size {size}
-      , Entry {std::move(entry)} {}
-
-    uint64_t Base;
-    uint64_t Size;
-    fextl::unique_ptr<CodeRegionEntry> Entry;
-  };
-  /**  @} */
-
-private:
-  NamedRegionObjectHandler* NamedRegionHandler;
-  CodeObjectSerializeService* CodeObjectCacheService;
-};
-
-class NamedRegionObjectHandler final {
-public:
-  NamedRegionObjectHandler(FEXCore::Context::ContextImpl* ctx);
-
-  void HandleNamedRegionObjectJobs();
-
-  const CodeObjectSerializationConfig& GetDefaultSerializationConfig() const {
-    return DefaultSerializationConfig;
-  }
-
-protected:
-  friend class AsyncJobHandler;
-
-  // Return a default code header based off the default serialization config
-  CodeObjectSerializationHeader DefaultCodeHeader(uint64_t Base, uint64_t Offset) const {
-    return CodeObjectSerializationHeader {
-      .Config = DefaultSerializationConfig,
-      .OriginalBase = Base,
-      .OriginalOffset = Offset,
-      .NumCodeEntries = 0,
-      .NumRelocationsTo = 0,
-      .TotalRelocationsCount = 0,
-    };
-  }
-
-  /**
-   * @brief Adds an asynchronous add named region work item to the object queue
-   *
-   * This adds the job that will do the loading of file resources and data tracking.
-   */
-  void AsyncAddNamedRegionWorkItem(const fextl::string& base, const fextl::string& filename, bool executable, CodeRegionMapType::iterator entry) {
-    std::unique_lock lk {NamedWorkQueueMutex};
-    WorkQueue.emplace(fextl::make_unique<AsyncJobHandler::WorkItemAddNamedRegion>(base, filename, executable, entry));
-    ++NamedWorkQueueJobs;
-  }
-
-  void AsyncRemoveNamedRegionWorkItem(uint64_t Base, uint64_t Size, fextl::unique_ptr<CodeRegionEntry> Entry) {
-    std::unique_lock lk {NamedWorkQueueMutex};
-    WorkQueue.emplace(fextl::make_unique<AsyncJobHandler::WorkItemRemoveNamedRegion>(Base, Size, std::move(Entry)));
-    ++NamedWorkQueueJobs;
-  }
-
-private:
-  // Code version. If the code emission changes then this needs to increment
-  constexpr static uint32_t CODE_VERSION = 0x0;
-
-  // Default cookie header for the file header
-  constexpr static uint64_t CODE_COOKIE = FEXCore::IR::COOKIE_VERSION("FEXC", CODE_VERSION);
-
-  // Code serialization config for our current process configuration
-  CodeObjectSerializationConfig DefaultSerializationConfig;
-
-  // Atomic counter for number of jobs in the queue without needing to pull the mutex to check
-  std::atomic<uint64_t> NamedWorkQueueJobs {};
-
-  // Mutex for ading new jobs to the work queue
-  std::mutex NamedWorkQueueMutex {};
-
-  // The job queue itself
-  // Jobs get consumed as a FIFO
-  // Jobs always get appended to the end
-  fextl::queue<fextl::unique_ptr<AsyncJobHandler::NamedRegionWorkItem>> WorkQueue {};
-
-  /**
-   * @name Named Region object handling
-   * @{ */
-  void AddNamedRegionObject(CodeRegionMapType::iterator Entry, const fextl::string& base_filename, const fextl::string& filename, bool Executable);
-  void RemoveNamedRegionObject(uintptr_t Base, uintptr_t Size, fextl::unique_ptr<CodeRegionEntry> Entry);
-  /**  @} */
-};
-
-/**
- * @brief Context specific code object serialization class
- *
- * Contains everything required for FEXCore to serialize code objects
- */
-class CodeObjectSerializeService final {
-public:
-  CodeObjectSerializeService(FEXCore::Context::ContextImpl* ctx);
-
-  /**
-   * @brief Initialize the internal interface
-   *
-   * Is a public interface to allow the service to reinitialize after forking
-   */
-  void Initialize();
-
-  /**
-   * @brief Safely shut down the Code Object serialization service.
-   *
-   * This service needs to be resiliant to application crashes, but shutting down safely is still preferred.
-   */
-  void Shutdown();
-
-  /**
-   * @name Async interface
-   * @{ */
-  /**
-   * @brief Loads a named region in to the code serialization service. As async as possible.
-   *
-   * @param Base - Virtual address that this named region is loaded
-   * @param Size - The size of the region
-   * @param Offset - The offset from the file
-   * @param filename - The filename itself
-   */
-  void AsyncAddNamedRegionJob(uintptr_t Base, uintptr_t Size, uintptr_t Offset, const fextl::string& filename) {
-    AsyncHandler.AsyncAddNamedRegionJob(Base, Size, Offset, filename);
-  }
-
-  /**
-   * @brief Unloads a named region from the code serialization service. As async as possible.
-   *
-   * @param Base - Virtual address of the named region
-   * @param Size - The size of the region
-   */
-  void AsyncRemoveNamedRegionJob(uintptr_t Base, uintptr_t Size) {
-    AsyncHandler.AsyncRemoveNamedRegionJob(Base, Size);
-  }
-
-  /**
-   * @brief Adds a code object serialization job. As async as possible.
-   * Code hashing happens prior to async job serialization to catch invalidations due to backpatching.
-   *
-   * @param Data - A fully filled out struct containing all the code serialization
-   */
-  void AsyncAddSerializationJob(fextl::unique_ptr<AsyncJobHandler::SerializationJobData> Data) {
-    AsyncHandler.AsyncAddSerializationJob(std::move(Data));
-  }
-  /**  @} */
-
-  /**
-   * @name Synchronous interface
-   * @{ */
-  /**
-   * @brief Synchronously waits for this thread's job queue to become empty.
-   *
-   * This is necessary for when a thread is shutting down
-   *
-   * @param ThreadJobRefCount - The shared mutex to wait on until to be empty
-   */
-  static void WaitForEmptyJobQueue(CodeSerializationMutex* ThreadJobRefCount) {
-    // Once the shared mutex is empty this unique lock will be gained
-    std::unique_lock lk {*ThreadJobRefCount};
-  }
-
-  /**
-   * @brief Fetches object code from the Code Object Cache for JIT.
-   *
-   * @param GuestRIP - Which GuestRIP to search the cache for
-   *
-   * @return Data required for the JIT to relocate the Object code.
-   */
-  const CodeObjectFileSection* FetchCodeObjectFromCache(uint64_t GuestRIP);
-  /**  @} */
-
-  // Public for threading
-  void ExecutionThread();
-
-protected:
-  friend class AsyncJobHandler;
-
-  /**
-   * @brief Safely closes out code object regions from the map
-   *
-   * @param it - iterator to do a closure on
-   */
-  void DoCodeRegionClosure(uint64_t Base, CodeRegionEntry* it);
-
-  CodeSerializationMutex& GetEntryMapMutex() {
-    return EntryMapMutex;
-  }
-  CodeSerializationMutex& GetUnrelocatedEntryMapMutex() {
-    return EntryMapMutex;
-  }
-
-  CodeRegionMapType& GetEntryMap() {
-    return AddressToEntryMap;
-  }
-  CodeRegionPtrMapType& GetUnrelocatedEntryMap() {
-    return UnrelocatedAddressToEntryMap;
-  }
-
-  /**
-   * @brief Notify the async thread that it has work to do
-   */
-  void NotifyWork() {
-    WorkAvailable.NotifyOne();
-  }
-
-private:
-  FEXCore::Context::ContextImpl* CTX;
-
-  Event WorkAvailable {};
-  fextl::unique_ptr<FEXCore::Threads::Thread> WorkerThread;
-  std::atomic_bool WorkerThreadShuttingDown {false};
-  AsyncJobHandler AsyncHandler;
-  NamedRegionObjectHandler NamedRegionHandler;
-
-  // Mutex to hold when modifying the entry maps
-  CodeSerializationMutex EntryMapMutex;
-  CodeSerializationMutex UnrelocatedEntryMapMutex;
-
-  // Entry maps
-  CodeRegionMapType AddressToEntryMap;
-  CodeRegionPtrMapType UnrelocatedAddressToEntryMap;
-};
-} // namespace FEXCore::CodeSerialize

FEXCore/Source/Interface/Core/OpcodeDispatcher.h

@@ -27,8 +27,6 @@
 #include <xxhash.h>
 
 namespace FEXCore::IR {
-class Pass;
-class PassManager;
 
 enum class MemoryAccessType {
   // Choose TSO or Non-TSO depending on access type
@@ -87,9 +85,6 @@ struct DispatchTableEntry {
 };
 
 class OpDispatchBuilder final : public IREmitter {
-  friend class FEXCore::IR::Pass;
-  friend class FEXCore::IR::PassManager;
-
 public:
   Ref GetNewJumpBlock(uint64_t RIP) {
     auto it = JumpTargets.find(RIP);
@@ -201,8 +196,7 @@ public:
         const auto GPRSize = GetGPROpSize();
         // If we don't have a jump target to a new block then we have to leave
         // Set the RIP to the next instruction and leave
-        auto RelocatedNextRIP = _EntrypointOffset(GPRSize, NextRIP - Entry);
-        ExitFunction(RelocatedNextRIP);
+        ExitFunction(_InlineEntrypointOffset(GPRSize, NextRIP - Entry));
       } else if (it != JumpTargets.end()) {
         Jump(it->second.BlockEntry);
         return true;
@@ -266,7 +260,6 @@ public:
   }
 
   OpDispatchBuilder(FEXCore::Context::ContextImpl* ctx);
-  OpDispatchBuilder(FEXCore::Utils::IntrusivePooledAllocator& Allocator);
 
   void ResetWorkingList();
   void ResetDecodeFailure() {
@@ -300,7 +293,7 @@ public:
     return ShouldDump;
   }
 
-  void BeginFunction(uint64_t RIP, const fextl::vector<FEXCore::Frontend::Decoder::DecodedBlocks>* Blocks, uint32_t NumInstructions, bool Is64BitMode);
+  void BeginFunction(uint64_t RIP, const fextl::vector<FEXCore::Frontend::Decoder::DecodedBlocks>* Blocks, uint32_t NumInstructions, bool Is64BitMode, bool MonoBackpatcherBlock);
   void Finalize();
 
   // Dispatch builder functions
@@ -350,6 +343,9 @@ public:
   void LoopOp(OpcodeArgs);
   void JUMPOp(OpcodeArgs);
   void JUMPAbsoluteOp(OpcodeArgs);
+  void JUMPFARIndirectOp(OpcodeArgs);
+  void CALLFARIndirectOp(OpcodeArgs);
+  void RETFARIndirectOp(OpcodeArgs);
   void TESTOp(OpcodeArgs, uint32_t SrcIndex);
   void MOVSXDOp(OpcodeArgs);
   void MOVSXOp(OpcodeArgs);
@@ -735,6 +731,7 @@ public:
   void FADD(OpcodeArgs, IR::OpSize Width, bool Integer, OpResult ResInST0);
   void FDIV(OpcodeArgs, IR::OpSize Width, bool Integer, bool Reverse, OpResult ResInST0);
   void FMUL(OpcodeArgs, IR::OpSize Width, bool Integer, OpResult ResInST0);
+  void FNCLEX(OpcodeArgs);
   void FNINIT(OpcodeArgs);
   void FSUB(OpcodeArgs, IR::OpSize Width, bool Integer, bool Reverse, OpResult ResInST0);
   void FTST(OpcodeArgs);
@@ -849,8 +846,6 @@ public:
 
   void PSADBW(OpcodeArgs);
 
-  Ref BitwiseAtLeastTwo(Ref A, Ref B, Ref C);
-
   void SHA1NEXTEOp(OpcodeArgs);
   void SHA1MSG1Op(OpcodeArgs);
   void SHA1MSG2Op(OpcodeArgs);
@@ -940,7 +935,6 @@ public:
   RefVSIB AVX128_LoadVSIB(const X86Tables::DecodedOp& Op, const X86Tables::DecodedOperand& Operand, uint32_t Flags, bool NeedsHigh);
   void AVX128_StoreResult_WithOpSize(FEXCore::X86Tables::DecodedOp Op, const FEXCore::X86Tables::DecodedOperand& Operand, const RefPair Src,
                                      MemoryAccessType AccessType = MemoryAccessType::DEFAULT);
-  void InstallAVX128Handlers();
   void AVX128_VMOVScalarImpl(OpcodeArgs, IR::OpSize ElementSize);
   void AVX128_VectorALU(OpcodeArgs, IROps IROp, IR::OpSize ElementSize);
   void AVX128_VectorUnary(OpcodeArgs, IROps IROp, IR::OpSize ElementSize);
@@ -969,39 +963,26 @@ public:
   void AVX128_VMOVDDUP(OpcodeArgs);
   void AVX128_VMOVSLDUP(OpcodeArgs);
   void AVX128_VMOVSHDUP(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_VBROADCAST(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_VPUNPCKL(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_VPUNPCKH(OpcodeArgs);
+  void AVX128_VBROADCAST(OpcodeArgs, IR::OpSize ElementSize);
+  void AVX128_VPUNPCKL(OpcodeArgs, IR::OpSize ElementSize);
+  void AVX128_VPUNPCKH(OpcodeArgs, IR::OpSize ElementSize);
   void AVX128_MOVVectorUnaligned(OpcodeArgs);
-  template<IR::OpSize DstElementSize>
-  void AVX128_InsertCVTGPR_To_FPR(OpcodeArgs);
-  template<IR::OpSize SrcElementSize, bool HostRoundingMode>
-  void AVX128_CVTFPR_To_GPR(OpcodeArgs);
+  void AVX128_InsertCVTGPR_To_FPR(OpcodeArgs, IR::OpSize DstElementSize);
+  void AVX128_CVTFPR_To_GPR(OpcodeArgs, IR::OpSize SrcElementSize, bool HostRoundingMode);
   void AVX128_VANDN(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_VPACKSS(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_VPACKUS(OpcodeArgs);
+  void AVX128_VPACKSS(OpcodeArgs, IR::OpSize ElementSize);
+  void AVX128_VPACKUS(OpcodeArgs, IR::OpSize ElementSize);
   Ref AVX128_PSIGNImpl(IR::OpSize ElementSize, Ref Src1, Ref Src2);
-  template<IR::OpSize ElementSize>
-  void AVX128_VPSIGN(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_UCOMISx(OpcodeArgs);
+  void AVX128_VPSIGN(OpcodeArgs, IR::OpSize ElementSize);
+  void AVX128_UCOMISx(OpcodeArgs, IR::OpSize ElementSize);
   void AVX128_VectorScalarInsertALU(OpcodeArgs, FEXCore::IR::IROps IROp, IR::OpSize ElementSize);
   Ref AVX128_VFCMPImpl(IR::OpSize ElementSize, Ref Src1, Ref Src2, uint8_t CompType);
-  template<IR::OpSize ElementSize>
-  void AVX128_VFCMP(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_InsertScalarFCMP(OpcodeArgs);
+  void AVX128_VFCMP(OpcodeArgs, IR::OpSize ElementSize);
+  void AVX128_InsertScalarFCMP(OpcodeArgs, IR::OpSize ElementSize);
   void AVX128_MOVBetweenGPR_FPR(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_PExtr(OpcodeArgs);
+  void AVX128_PExtr(OpcodeArgs, IR::OpSize ElementSize);
   void AVX128_ExtendVectorElements(OpcodeArgs, IR::OpSize ElementSize, IR::OpSize DstElementSize, bool Signed);
-  template<IR::OpSize ElementSize>
-  void AVX128_MOVMSK(OpcodeArgs);
+  void AVX128_MOVMSK(OpcodeArgs, IR::OpSize ElementSize);
   void AVX128_MOVMSKB(OpcodeArgs);
   void AVX128_PINSRImpl(OpcodeArgs, IR::OpSize ElementSize, const X86Tables::DecodedOperand& Src1Op,
                         const X86Tables::DecodedOperand& Src2Op, const X86Tables::DecodedOperand& Imm);
@@ -1015,33 +996,25 @@ public:
   void AVX128_VINSERTPS(OpcodeArgs);
 
   Ref AVX128_PHSUBImpl(Ref Src1, Ref Src2, size_t ElementSize);
-  template<IR::OpSize ElementSize>
-  void AVX128_VPHSUB(OpcodeArgs);
+  void AVX128_VPHSUB(OpcodeArgs, IR::OpSize ElementSize);
 
   void AVX128_VPHSUBSW(OpcodeArgs);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VADDSUBP(OpcodeArgs);
+  void AVX128_VADDSUBP(OpcodeArgs, IR::OpSize ElementSize);
 
-  template<IR::OpSize ElementSize, bool Signed>
-  void AVX128_VPMULL(OpcodeArgs);
+  void AVX128_VPMULL(OpcodeArgs, IR::OpSize ElementSize, bool Signed);
 
   void AVX128_VPMULHRSW(OpcodeArgs);
 
-  template<bool Signed>
-  void AVX128_VPMULHW(OpcodeArgs);
+  void AVX128_VPMULHW(OpcodeArgs, bool Signed);
 
-  template<IR::OpSize DstElementSize, IR::OpSize SrcElementSize>
-  void AVX128_InsertScalar_CVT_Float_To_Float(OpcodeArgs);
+  void AVX128_InsertScalar_CVT_Float_To_Float(OpcodeArgs, IR::OpSize DstElementSize, IR::OpSize SrcElementSize);
 
-  template<IR::OpSize DstElementSize, IR::OpSize SrcElementSize>
-  void AVX128_Vector_CVT_Float_To_Float(OpcodeArgs);
+  void AVX128_Vector_CVT_Float_To_Float(OpcodeArgs, IR::OpSize DstElementSize, IR::OpSize SrcElementSize);
 
-  template<IR::OpSize SrcElementSize, bool HostRoundingMode>
-  void AVX128_Vector_CVT_Float_To_Int(OpcodeArgs);
+  void AVX128_Vector_CVT_Float_To_Int(OpcodeArgs, IR::OpSize SrcElementSize, bool HostRoundingMode);
 
-  template<IR::OpSize SrcElementSize, bool Widen>
-  void AVX128_Vector_CVT_Int_To_Float(OpcodeArgs);
+  void AVX128_Vector_CVT_Int_To_Float(OpcodeArgs, IR::OpSize SrcElementSize, bool Widen);
 
   void AVX128_VEXTRACT128(OpcodeArgs);
   void AVX128_VAESImc(OpcodeArgs);
@@ -1058,36 +1031,28 @@ public:
 
   void AVX128_PHMINPOSUW(OpcodeArgs);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VectorRound(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_InsertScalarRound(OpcodeArgs);
+  void AVX128_VectorRound(OpcodeArgs, IR::OpSize ElementSize);
+  void AVX128_InsertScalarRound(OpcodeArgs, IR::OpSize ElementSize);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VDPP(OpcodeArgs);
+  void AVX128_VDPP(OpcodeArgs, IR::OpSize ElementSize);
   void AVX128_VPERMQ(OpcodeArgs);
 
   void AVX128_VPSHUFW(OpcodeArgs, bool Low);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VSHUF(OpcodeArgs);
+  void AVX128_VSHUF(OpcodeArgs, IR::OpSize ElementSize);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VPERMILImm(OpcodeArgs);
+  void AVX128_VPERMILImm(OpcodeArgs, IR::OpSize ElementSize);
 
-  template<IROps IROp, IR::OpSize ElementSize>
-  void AVX128_VHADDP(OpcodeArgs);
+  void AVX128_VHADDP(OpcodeArgs, IROps IROp, IR::OpSize ElementSize);
 
   void AVX128_VPHADDSW(OpcodeArgs);
 
   void AVX128_VPMADDUBSW(OpcodeArgs);
   void AVX128_VPMADDWD(OpcodeArgs);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VBLEND(OpcodeArgs);
+  void AVX128_VBLEND(OpcodeArgs, IR::OpSize ElementSize);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VHSUBP(OpcodeArgs);
+  void AVX128_VHSUBP(OpcodeArgs, IR::OpSize ElementSize);
 
   void AVX128_VPSHUFB(OpcodeArgs);
   void AVX128_VPSADBW(OpcodeArgs);
@@ -1098,28 +1063,23 @@ public:
   void AVX128_VMASKMOVImpl(OpcodeArgs, IR::OpSize ElementSize, IR::OpSize DstSize, bool IsStore, const X86Tables::DecodedOperand& MaskOp,
                            const X86Tables::DecodedOperand& DataOp);
 
-  template<bool IsStore>
-  void AVX128_VPMASKMOV(OpcodeArgs);
+  void AVX128_VPMASKMOV(OpcodeArgs, bool IsStore);
 
-  template<IR::OpSize ElementSize, bool IsStore>
-  void AVX128_VMASKMOV(OpcodeArgs);
+  void AVX128_VMASKMOV(OpcodeArgs, IR::OpSize ElementSize, bool IsStore);
 
   void AVX128_MASKMOV(OpcodeArgs);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VectorVariableBlend(OpcodeArgs);
+  void AVX128_VectorVariableBlend(OpcodeArgs, IR::OpSize ElementSize);
 
   void AVX128_SaveAVXState(Ref MemBase);
   void AVX128_RestoreAVXState(Ref MemBase);
   void AVX128_DefaultAVXState();
 
   void AVX128_VPERM2(OpcodeArgs);
-  template<IR::OpSize ElementSize>
-  void AVX128_VTESTP(OpcodeArgs);
+  void AVX128_VTESTP(OpcodeArgs, IR::OpSize ElementSize);
   void AVX128_PTest(OpcodeArgs);
 
-  template<IR::OpSize ElementSize>
-  void AVX128_VPERMILReg(OpcodeArgs);
+  void AVX128_VPERMILReg(OpcodeArgs, IR::OpSize ElementSize);
 
   void AVX128_VPERMD(OpcodeArgs);
 
@@ -1132,8 +1092,7 @@ public:
   RefPair AVX128_VPGatherQPSImpl(Ref Dest, Ref Mask, RefVSIB VSIB);
   RefPair AVX128_VPGatherImpl(OpSize Size, OpSize ElementLoadSize, OpSize AddrElementSize, RefPair Dest, RefPair Mask, RefVSIB VSIB);
 
-  template<OpSize AddrElementSize>
-  void AVX128_VPGATHER(OpcodeArgs);
+  void AVX128_VPGATHER(OpcodeArgs, OpSize AddrElementSize);
 
   void AVX128_VCVTPH2PS(OpcodeArgs);
   void AVX128_VCVTPS2PH(OpcodeArgs);
@@ -1167,6 +1126,10 @@ public:
     StoreXMMRegister(XMM, Value);
   }
 
+  void AVXVectorALUOp(OpcodeArgs, IROps IROp, IR::OpSize ElementSize);
+  void AVXVectorUnaryOp(OpcodeArgs, IROps IROp, IR::OpSize ElementSize);
+  void AVXVectorVariableBlend(OpcodeArgs, IR::OpSize ElementSize);
+
   // End of AVX 256-bit implementation
 
   void InvalidOp(OpcodeArgs);
@@ -1189,13 +1152,42 @@ public:
     }
   }
 
-  void FlushRegisterCache(bool SRAOnly = false) {
+  void StoreContextHelper(IR::OpSize Size, RegisterClassType Class, Ref Value, uint32_t Offset) {
+    // For i128Bit, we won't see a normal Constant to inline, but as a special
+    // case we can replace with a 2x64-bit store which can use inline zeroes.
+    if (Size == OpSize::i128Bit) {
+      auto Header = GetOpHeader(WrapNode(Value));
+      const auto MAX_STP_OFFSET = (252 * 4);
+
+      if (Offset <= MAX_STP_OFFSET && Header->Op == OP_LOADNAMEDVECTORCONSTANT) {
+        auto Const = Header->C<IR::IROp_LoadNamedVectorConstant>();
+
+        if (Const->Constant == IR::NamedVectorConstant::NAMED_VECTOR_ZERO) {
+          Ref Zero = _Constant(0);
+          Ref STP = _StoreContextPair(IR::OpSize::i64Bit, GPRClass, Zero, Zero, Offset);
+
+          // XXX: This works around InlineConstant not having an associated
+          // register class, else we'd just do InlineConstant above.
+          Ref InlineZero = _InlineConstant(0);
+          ReplaceNodeArgument(STP, 0, InlineZero);
+          ReplaceNodeArgument(STP, 1, InlineZero);
+          return;
+        }
+      }
+    }
+
+    _StoreContext(Size, Class, Value, Offset);
+  }
+
+  void FlushRegisterCache(bool SRAOnly = false, bool MMXOnly = false) {
     // At block boundaries, fix up the carry flag.
     if (!SRAOnly) {
       RectifyCarryInvert(CFInvertedABI);
     }
 
-    CalculateDeferredFlags();
+    if (!MMXOnly) {
+      CalculateDeferredFlags();
+    }
 
     const auto GPRSize = GetGPROpSize();
     const auto VectorSize = GetGuestVectorLength();
@@ -1217,6 +1209,11 @@ public:
       Bits &= Mask;
     }
 
+    if (MMXOnly) {
+      Mask &= ((1ull << (MM7Index - MM0Index + 1)) - 1) << MM0Index;
+      Bits &= Mask;
+    }
+
     while (Bits != 0) {
       uint32_t Index = 63 - std::countl_zero(Bits);
       Ref Value = RegCache.Value[Index];
@@ -1254,7 +1251,7 @@ public:
           _StoreContextPair(Size, Class, ValueNext, Value, Offset - SizeInt);
           Bits &= ~NextBit;
         } else {
-          _StoreContext(Size, Class, Value, Offset);
+          StoreContextHelper(Size, Class, Value, Offset);
           // If Partial and MMX register, then we need to store all 1s in bits 64-80
           if (Partial && Index >= MM0Index && Index <= MM7Index) {
             _StoreContext(OpSize::i16Bit, IR::GPRClass, Constant(0xFFFF), Offset + 8);
@@ -1376,11 +1373,6 @@ private:
 
   Ref ADDSUBPOpImpl(OpSize Size, IR::OpSize ElementSize, Ref Src1, Ref Src2);
 
-  void AVXVectorALUOp(OpcodeArgs, IROps IROp, IR::OpSize ElementSize);
-  void AVXVectorUnaryOp(OpcodeArgs, IROps IROp, IR::OpSize ElementSize);
-
-  void AVXVectorVariableBlend(OpcodeArgs, IR::OpSize ElementSize);
-
   void AVXVariableShiftImpl(OpcodeArgs, IROps IROp);
 
   Ref AESKeyGenAssistImpl(OpcodeArgs);
@@ -1521,7 +1513,23 @@ private:
   void StoreGPRRegister(uint32_t GPR, const Ref Src, IR::OpSize Size = OpSize::iInvalid, uint8_t Offset = 0);
   void StoreXMMRegister(uint32_t XMM, const Ref Src);
 
-  Ref GetRelocatedPC(const FEXCore::X86Tables::DecodedOp& Op, int64_t Offset = 0);
+  Ref _GetRelocatedPC(const FEXCore::X86Tables::DecodedOp& Op, int64_t Offset, bool Inline) {
+    const auto GPRSize = GetGPROpSize();
+    const auto Offs = Op->PC + Op->InstSize + Offset - Entry;
+    return Inline ? _InlineEntrypointOffset(GPRSize, Offs) : _EntrypointOffset(GPRSize, Offs);
+  }
+
+  Ref GetRelocatedPC(const FEXCore::X86Tables::DecodedOp& Op, int64_t Offset = 0) {
+    return _GetRelocatedPC(Op, Offset, false);
+  }
+
+  void ExitRelocatedPC(const FEXCore::X86Tables::DecodedOp& Op, int64_t Offset = 0) {
+    ExitFunction(_GetRelocatedPC(Op, Offset, true /* Inline */));
+  }
+
+  void ExitRelocatedPC(const FEXCore::X86Tables::DecodedOp& Op, int64_t Offset, BranchHint Hint, Ref CallReturnAddress, Ref CallReturnBlock) {
+    ExitFunction(_GetRelocatedPC(Op, Offset, true /* Inline */), Hint, CallReturnAddress, CallReturnBlock);
+  }
 
   [[nodiscard]]
   static bool IsOperandMem(const X86Tables::DecodedOperand& Operand, bool Load) {
@@ -1650,7 +1658,7 @@ private:
     // This is currently worse for 8/16-bit, but that should be optimized. TODO
     if (SrcSize >= OpSize::i32Bit) {
       if (SetPF) {
-        CalculatePF(_SubWithFlags(SrcSize, Res, Constant(0)));
+        CalculatePF(SubWithFlags(SrcSize, Res, (uint64_t)0));
       } else {
         _SubNZCV(SrcSize, Res, Constant(0));
       }
@@ -2036,7 +2044,7 @@ private:
       } else {
         // Because we explicitly inverted for CF above, we use the unsafe
         // _NZCVSelect rather than the safe CF-aware version.
-        return _NZCVSelect(OpSize::i32Bit, CondForNZCVBit(BitOffset, Invert), Constant(1), Constant(0));
+        return _NZCVSelect01(CondForNZCVBit(BitOffset, Invert));
       }
     } else if (BitOffset == FEXCore::X86State::RFLAG_PF_RAW_LOC) {
       return LoadGPR(Core::CPUState::PF_AS_GREG);
@@ -2133,7 +2141,7 @@ private:
   void ConvertNZCVToX87() {
     LOGMAN_THROW_A_FMT(NZCVDirty && CachedNZCV, "NZCV must be saved");
 
-    Ref V = _NZCVSelect(OpSize::i32Bit, CondForNZCVBit(FEXCore::X86State::RFLAG_OF_RAW_LOC, false), Constant(1), Constant(0));
+    Ref V = _NZCVSelect01(CondForNZCVBit(FEXCore::X86State::RFLAG_OF_RAW_LOC, false));
 
     if (CTX->HostFeatures.SupportsFlagM2) {
       // Convert to x86 flags, saves us from or'ing after.
@@ -2141,8 +2149,8 @@ private:
     }
 
     // CF is inverted after FCMP
-    Ref C = _NZCVSelect(OpSize::i32Bit, CondForNZCVBit(FEXCore::X86State::RFLAG_CF_RAW_LOC, true), Constant(1), Constant(0));
-    Ref Z = _NZCVSelect(OpSize::i32Bit, CondForNZCVBit(FEXCore::X86State::RFLAG_ZF_RAW_LOC, false), Constant(1), Constant(0));
+    Ref C = _NZCVSelect01(CondForNZCVBit(FEXCore::X86State::RFLAG_CF_RAW_LOC, true));
+    Ref Z = _NZCVSelect01(CondForNZCVBit(FEXCore::X86State::RFLAG_ZF_RAW_LOC, false));
 
     if (!CTX->HostFeatures.SupportsFlagM2) {
       C = _Or(OpSize::i32Bit, C, V);
@@ -2238,7 +2246,7 @@ private:
 
   std::optional<CondClassType> DecodeNZCVCondition(uint8_t OP);
   Ref SelectBit(Ref Cmp, IR::OpSize ResultSize, Ref TrueValue, Ref FalseValue);
-  Ref SelectCC(uint8_t OP, IR::OpSize ResultSize, Ref TrueValue, Ref FalseValue);
+  Ref SelectCC0All1(uint8_t OP);
 
   /**
    * @brief Flushes NZCV. Mostly vestigial.
@@ -2313,7 +2321,6 @@ private:
    * @name These functions are used by the deferred flag handling while it is calculating and storing flags in to RFLAGs.
    * @{ */
   Ref LoadPFRaw(bool Mask, bool Invert);
-  Ref SelectPF(bool Invert, IR::OpSize ResultSize, Ref TrueValue, Ref FalseValue);
   Ref LoadAF();
   void FixupAF();
   void SetAFAndFixup(Ref AF);
@@ -2356,10 +2363,12 @@ private:
 
   void ChgStateMMX_X87() override {
     LOGMAN_THROW_A_FMT(MMXState == MMXState_MMX, "Expected state to be MMX");
+    // The opcode dispatcher register cache is used for MMX, but the x87 pass register cache is used for x87, spill to
+    // context to ensure coherence.
+    FlushRegisterCache(false, true);
     // We explicitly initialize to x87 state in StartNewBlock.
     // So if we ever change this to do something else, we need to
     // make sure that we consider if we need to explicitly set it there.
-    FlushRegisterCache();
     MMXState = MMXState_X87;
   }
 
@@ -2377,6 +2386,11 @@ private:
   bool Multiblock {};
   bool Is64BitMode {};
   uint64_t Entry {};
+
+  // Set if mono hacks are enabled and the current block is the mono callsite backpatcher, in which case the
+  // XCHG ops that would patch code are replaced with a hook that performs the write and manually invalidates
+  // the target address.
+  bool IsMonoBackpatcherBlock {false};
   IROp_IRHeader* CurrentHeader {};
 
   [[nodiscard]]
@@ -2546,7 +2560,7 @@ private:
     }
 
     ArithRef Presub(uint64_t K) {
-      return IsConstant ? ArithRef(E, K - C) : ArithRef(E, E->_Sub(OpSize::i64Bit, E->Constant(K), R));
+      return IsConstant ? ArithRef(E, K - C) : ArithRef(E, E->Sub(OpSize::i64Bit, E->Constant(K), R));
     }
 
     ArithRef Lshl(uint64_t Shift) {
@@ -2638,8 +2652,6 @@ private:
     return ArithRef(this, K);
   }
 
-  void InstallHostSpecificOpcodeHandlers();
-
   ///< Segment telemetry tracking
   uint32_t SegmentsNeedReadCheck {~0U};
   void CheckLegacySegmentWrite(Ref NewNode, uint32_t SegmentReg);
@@ -2653,16 +2665,14 @@ constexpr inline void InstallToTable(auto& FinalTable, const auto& LocalTable) {
     for (uint8_t i = 0; i < Op.Count; ++i) {
       auto& TableOp = FinalTable[OpNum + i];
 #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
-      if (TableOp.OpcodeDispatcher) {
+      if (TableOp.OpcodeDispatcher.OpDispatch) {
         ERROR_AND_DIE_FMT("Duplicate Entry {}", TableOp.Name);
       }
 #endif
 
-      TableOp.OpcodeDispatcher = Dispatcher;
+      TableOp.OpcodeDispatcher.OpDispatch = Dispatcher;
     }
   }
 }
 
-void InstallOpcodeHandlers(Context::OperatingMode Mode);
-
 } // namespace FEXCore::IR

FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp

@@ -21,466 +21,6 @@ class OrderedNode;
 
 #define OpcodeArgs [[maybe_unused]] FEXCore::X86Tables::DecodedOp Op
 
-void OpDispatchBuilder::InstallAVX128Handlers() {
-#define OPD(map_select, pp, opcode) (((map_select - 1) << 10) | (pp << 8) | (opcode))
-  static constexpr DispatchTableEntry AVX128Table[] = {
-    {OPD(1, 0b00, 0x10), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b01, 0x10), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b10, 0x10), 1, &OpDispatchBuilder::AVX128_VMOVSS},
-    {OPD(1, 0b11, 0x10), 1, &OpDispatchBuilder::AVX128_VMOVSD},
-    {OPD(1, 0b00, 0x11), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b01, 0x11), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b10, 0x11), 1, &OpDispatchBuilder::AVX128_VMOVSS},
-    {OPD(1, 0b11, 0x11), 1, &OpDispatchBuilder::AVX128_VMOVSD},
-
-    {OPD(1, 0b00, 0x12), 1, &OpDispatchBuilder::AVX128_VMOVLP},
-    {OPD(1, 0b01, 0x12), 1, &OpDispatchBuilder::AVX128_VMOVLP},
-    {OPD(1, 0b10, 0x12), 1, &OpDispatchBuilder::AVX128_VMOVSLDUP},
-    {OPD(1, 0b11, 0x12), 1, &OpDispatchBuilder::AVX128_VMOVDDUP},
-    {OPD(1, 0b00, 0x13), 1, &OpDispatchBuilder::AVX128_VMOVLP},
-    {OPD(1, 0b01, 0x13), 1, &OpDispatchBuilder::AVX128_VMOVLP},
-
-    {OPD(1, 0b00, 0x14), 1, &OpDispatchBuilder::AVX128_VPUNPCKL<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x14), 1, &OpDispatchBuilder::AVX128_VPUNPCKL<OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x15), 1, &OpDispatchBuilder::AVX128_VPUNPCKH<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x15), 1, &OpDispatchBuilder::AVX128_VPUNPCKH<OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x16), 1, &OpDispatchBuilder::AVX128_VMOVHP},
-    {OPD(1, 0b01, 0x16), 1, &OpDispatchBuilder::AVX128_VMOVHP},
-    {OPD(1, 0b10, 0x16), 1, &OpDispatchBuilder::AVX128_VMOVSHDUP},
-    {OPD(1, 0b00, 0x17), 1, &OpDispatchBuilder::AVX128_VMOVHP},
-    {OPD(1, 0b01, 0x17), 1, &OpDispatchBuilder::AVX128_VMOVHP},
-
-    {OPD(1, 0b00, 0x28), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b01, 0x28), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b00, 0x29), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b01, 0x29), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-
-    {OPD(1, 0b10, 0x2A), 1, &OpDispatchBuilder::AVX128_InsertCVTGPR_To_FPR<OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x2A), 1, &OpDispatchBuilder::AVX128_InsertCVTGPR_To_FPR<OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x2B), 1, &OpDispatchBuilder::AVX128_MOVVectorNT},
-    {OPD(1, 0b01, 0x2B), 1, &OpDispatchBuilder::AVX128_MOVVectorNT},
-
-    {OPD(1, 0b10, 0x2C), 1, &OpDispatchBuilder::AVX128_CVTFPR_To_GPR<OpSize::i32Bit, false>},
-    {OPD(1, 0b11, 0x2C), 1, &OpDispatchBuilder::AVX128_CVTFPR_To_GPR<OpSize::i64Bit, false>},
-
-    {OPD(1, 0b10, 0x2D), 1, &OpDispatchBuilder::AVX128_CVTFPR_To_GPR<OpSize::i32Bit, true>},
-    {OPD(1, 0b11, 0x2D), 1, &OpDispatchBuilder::AVX128_CVTFPR_To_GPR<OpSize::i64Bit, true>},
-
-    {OPD(1, 0b00, 0x2E), 1, &OpDispatchBuilder::AVX128_UCOMISx<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x2E), 1, &OpDispatchBuilder::AVX128_UCOMISx<OpSize::i64Bit>},
-    {OPD(1, 0b00, 0x2F), 1, &OpDispatchBuilder::AVX128_UCOMISx<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x2F), 1, &OpDispatchBuilder::AVX128_UCOMISx<OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x50), 1, &OpDispatchBuilder::AVX128_MOVMSK<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x50), 1, &OpDispatchBuilder::AVX128_MOVMSK<OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x51), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorUnary, IR::OP_VFSQRT, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x51), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorUnary, IR::OP_VFSQRT, OpSize::i64Bit>},
-    {OPD(1, 0b10, 0x51), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFSQRTSCALARINSERT, OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x51), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFSQRTSCALARINSERT, OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x52), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorUnary, IR::OP_VFRSQRT, OpSize::i32Bit>},
-    {OPD(1, 0b10, 0x52), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFRSQRTSCALARINSERT, OpSize::i32Bit>},
-
-    {OPD(1, 0b00, 0x53), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorUnary, IR::OP_VFRECP, OpSize::i32Bit>},
-    {OPD(1, 0b10, 0x53), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFRECPSCALARINSERT, OpSize::i32Bit>},
-
-    {OPD(1, 0b00, 0x54), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VAND, OpSize::i128Bit>},
-    {OPD(1, 0b01, 0x54), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VAND, OpSize::i128Bit>},
-
-    {OPD(1, 0b00, 0x55), 1, &OpDispatchBuilder::AVX128_VANDN},
-    {OPD(1, 0b01, 0x55), 1, &OpDispatchBuilder::AVX128_VANDN},
-
-    {OPD(1, 0b00, 0x56), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VOR, OpSize::i128Bit>},
-    {OPD(1, 0b01, 0x56), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VOR, OpSize::i128Bit>},
-
-    {OPD(1, 0b00, 0x57), 1, &OpDispatchBuilder::AVX128_VectorXOR},
-    {OPD(1, 0b01, 0x57), 1, &OpDispatchBuilder::AVX128_VectorXOR},
-
-    {OPD(1, 0b00, 0x58), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFADD, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x58), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFADD, OpSize::i64Bit>},
-    {OPD(1, 0b10, 0x58), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFADDSCALARINSERT, OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x58), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFADDSCALARINSERT, OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x59), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFMUL, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x59), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFMUL, OpSize::i64Bit>},
-    {OPD(1, 0b10, 0x59), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFMULSCALARINSERT, OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x59), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFMULSCALARINSERT, OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x5A), 1, &OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Float<OpSize::i64Bit, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x5A), 1, &OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Float<OpSize::i32Bit, OpSize::i64Bit>},
-    {OPD(1, 0b10, 0x5A), 1, &OpDispatchBuilder::AVX128_InsertScalar_CVT_Float_To_Float<OpSize::i64Bit, OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x5A), 1, &OpDispatchBuilder::AVX128_InsertScalar_CVT_Float_To_Float<OpSize::i32Bit, OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x5B), 1, &OpDispatchBuilder::AVX128_Vector_CVT_Int_To_Float<OpSize::i32Bit, false>},
-    {OPD(1, 0b01, 0x5B), 1, &OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Int<OpSize::i32Bit, true>},
-    {OPD(1, 0b10, 0x5B), 1, &OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Int<OpSize::i32Bit, false>},
-
-    {OPD(1, 0b00, 0x5C), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFSUB, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x5C), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFSUB, OpSize::i64Bit>},
-    {OPD(1, 0b10, 0x5C), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFSUBSCALARINSERT, OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x5C), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFSUBSCALARINSERT, OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x5D), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFMIN, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x5D), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFMIN, OpSize::i64Bit>},
-    {OPD(1, 0b10, 0x5D), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFMINSCALARINSERT, OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x5D), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFMINSCALARINSERT, OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x5E), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFDIV, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x5E), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFDIV, OpSize::i64Bit>},
-    {OPD(1, 0b10, 0x5E), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFDIVSCALARINSERT, OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x5E), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFDIVSCALARINSERT, OpSize::i64Bit>},
-
-    {OPD(1, 0b00, 0x5F), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFMAX, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x5F), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VFMAX, OpSize::i64Bit>},
-    {OPD(1, 0b10, 0x5F), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFMAXSCALARINSERT, OpSize::i32Bit>},
-    {OPD(1, 0b11, 0x5F), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorScalarInsertALU, IR::OP_VFMAXSCALARINSERT, OpSize::i64Bit>},
-
-    {OPD(1, 0b01, 0x60), 1, &OpDispatchBuilder::AVX128_VPUNPCKL<OpSize::i8Bit>},
-    {OPD(1, 0b01, 0x61), 1, &OpDispatchBuilder::AVX128_VPUNPCKL<OpSize::i16Bit>},
-    {OPD(1, 0b01, 0x62), 1, &OpDispatchBuilder::AVX128_VPUNPCKL<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x63), 1, &OpDispatchBuilder::AVX128_VPACKSS<OpSize::i16Bit>},
-    {OPD(1, 0b01, 0x64), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VCMPGT, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0x65), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VCMPGT, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0x66), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VCMPGT, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x67), 1, &OpDispatchBuilder::AVX128_VPACKUS<OpSize::i16Bit>},
-    {OPD(1, 0b01, 0x68), 1, &OpDispatchBuilder::AVX128_VPUNPCKH<OpSize::i8Bit>},
-    {OPD(1, 0b01, 0x69), 1, &OpDispatchBuilder::AVX128_VPUNPCKH<OpSize::i16Bit>},
-    {OPD(1, 0b01, 0x6A), 1, &OpDispatchBuilder::AVX128_VPUNPCKH<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x6B), 1, &OpDispatchBuilder::AVX128_VPACKSS<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x6C), 1, &OpDispatchBuilder::AVX128_VPUNPCKL<OpSize::i64Bit>},
-    {OPD(1, 0b01, 0x6D), 1, &OpDispatchBuilder::AVX128_VPUNPCKH<OpSize::i64Bit>},
-    {OPD(1, 0b01, 0x6E), 1, &OpDispatchBuilder::AVX128_MOVBetweenGPR_FPR},
-
-    {OPD(1, 0b01, 0x6F), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b10, 0x6F), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-
-    {OPD(1, 0b01, 0x70), 1, &OpDispatchBuilder::AVX128_VPERMILImm<OpSize::i32Bit>},
-    {OPD(1, 0b10, 0x70), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VPSHUFW, false>},
-    {OPD(1, 0b11, 0x70), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VPSHUFW, true>},
-
-    {OPD(1, 0b01, 0x74), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VCMPEQ, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0x75), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VCMPEQ, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0x76), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VCMPEQ, OpSize::i32Bit>},
-
-    {OPD(1, 0b00, 0x77), 1, &OpDispatchBuilder::AVX128_VZERO},
-
-    {OPD(1, 0b01, 0x7C), 1, &OpDispatchBuilder::AVX128_VHADDP<IR::OP_VFADDP, OpSize::i64Bit>},
-    {OPD(1, 0b11, 0x7C), 1, &OpDispatchBuilder::AVX128_VHADDP<IR::OP_VFADDP, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0x7D), 1, &OpDispatchBuilder::AVX128_VHSUBP<OpSize::i64Bit>},
-    {OPD(1, 0b11, 0x7D), 1, &OpDispatchBuilder::AVX128_VHSUBP<OpSize::i32Bit>},
-
-    {OPD(1, 0b01, 0x7E), 1, &OpDispatchBuilder::AVX128_MOVBetweenGPR_FPR},
-    {OPD(1, 0b10, 0x7E), 1, &OpDispatchBuilder::AVX128_MOVQ},
-
-    {OPD(1, 0b01, 0x7F), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-    {OPD(1, 0b10, 0x7F), 1, &OpDispatchBuilder::AVX128_VMOVAPS},
-
-    {OPD(1, 0b00, 0xC2), 1, &OpDispatchBuilder::AVX128_VFCMP<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0xC2), 1, &OpDispatchBuilder::AVX128_VFCMP<OpSize::i64Bit>},
-    {OPD(1, 0b10, 0xC2), 1, &OpDispatchBuilder::AVX128_InsertScalarFCMP<OpSize::i32Bit>},
-    {OPD(1, 0b11, 0xC2), 1, &OpDispatchBuilder::AVX128_InsertScalarFCMP<OpSize::i64Bit>},
-
-    {OPD(1, 0b01, 0xC4), 1, &OpDispatchBuilder::AVX128_VPINSRW},
-    {OPD(1, 0b01, 0xC5), 1, &OpDispatchBuilder::AVX128_PExtr<OpSize::i16Bit>},
-
-    {OPD(1, 0b00, 0xC6), 1, &OpDispatchBuilder::AVX128_VSHUF<OpSize::i32Bit>},
-    {OPD(1, 0b01, 0xC6), 1, &OpDispatchBuilder::AVX128_VSHUF<OpSize::i64Bit>},
-
-    {OPD(1, 0b01, 0xD0), 1, &OpDispatchBuilder::AVX128_VADDSUBP<OpSize::i64Bit>},
-    {OPD(1, 0b11, 0xD0), 1, &OpDispatchBuilder::AVX128_VADDSUBP<OpSize::i32Bit>},
-
-    {OPD(1, 0b01, 0xD1), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftWideImpl, OpSize::i16Bit, IROps::OP_VUSHRSWIDE>}, // VPSRL
-    {OPD(1, 0b01, 0xD2), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftWideImpl, OpSize::i32Bit, IROps::OP_VUSHRSWIDE>}, // VPSRL
-    {OPD(1, 0b01, 0xD3), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftWideImpl, OpSize::i64Bit, IROps::OP_VUSHRSWIDE>}, // VPSRL
-    {OPD(1, 0b01, 0xD4), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VADD, OpSize::i64Bit>},
-    {OPD(1, 0b01, 0xD5), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VMUL, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xD6), 1, &OpDispatchBuilder::AVX128_MOVQ},
-    {OPD(1, 0b01, 0xD7), 1, &OpDispatchBuilder::AVX128_MOVMSKB},
-
-    {OPD(1, 0b01, 0xD8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUQSUB, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xD9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUQSUB, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xDA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUMIN, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xDB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VAND, OpSize::i128Bit>},
-    {OPD(1, 0b01, 0xDC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUQADD, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xDD), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUQADD, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xDE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUMAX, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xDF), 1, &OpDispatchBuilder::AVX128_VANDN},
-
-    {OPD(1, 0b01, 0xE0), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VURAVG, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xE1), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftWideImpl, OpSize::i16Bit, IROps::OP_VSSHRSWIDE>}, // VPSRA
-    {OPD(1, 0b01, 0xE2), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftWideImpl, OpSize::i32Bit, IROps::OP_VSSHRSWIDE>}, // VPSRA
-    {OPD(1, 0b01, 0xE3), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VURAVG, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xE4), 1, &OpDispatchBuilder::AVX128_VPMULHW<false>},
-    {OPD(1, 0b01, 0xE5), 1, &OpDispatchBuilder::AVX128_VPMULHW<true>},
-
-    {OPD(1, 0b01, 0xE6), 1, &OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Int<OpSize::i64Bit, false>},
-    {OPD(1, 0b10, 0xE6), 1, &OpDispatchBuilder::AVX128_Vector_CVT_Int_To_Float<OpSize::i32Bit, true>},
-    {OPD(1, 0b11, 0xE6), 1, &OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Int<OpSize::i64Bit, true>},
-
-    {OPD(1, 0b01, 0xE7), 1, &OpDispatchBuilder::AVX128_MOVVectorNT},
-
-    {OPD(1, 0b01, 0xE8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSQSUB, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xE9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSQSUB, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xEA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSMIN, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xEB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VOR, OpSize::i128Bit>},
-    {OPD(1, 0b01, 0xEC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSQADD, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xED), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSQADD, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xEE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSMAX, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xEF), 1, &OpDispatchBuilder::AVX128_VectorXOR},
-
-    {OPD(1, 0b11, 0xF0), 1, &OpDispatchBuilder::AVX128_MOVVectorUnaligned},
-    {OPD(1, 0b01, 0xF1), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftWideImpl, OpSize::i16Bit, IROps::OP_VUSHLSWIDE>}, // VPSLL
-    {OPD(1, 0b01, 0xF2), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftWideImpl, OpSize::i32Bit, IROps::OP_VUSHLSWIDE>}, // VPSLL
-    {OPD(1, 0b01, 0xF3), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftWideImpl, OpSize::i64Bit, IROps::OP_VUSHLSWIDE>}, // VPSLL
-    {OPD(1, 0b01, 0xF4), 1, &OpDispatchBuilder::AVX128_VPMULL<OpSize::i32Bit, false>},
-    {OPD(1, 0b01, 0xF5), 1, &OpDispatchBuilder::AVX128_VPMADDWD},
-    {OPD(1, 0b01, 0xF6), 1, &OpDispatchBuilder::AVX128_VPSADBW},
-    {OPD(1, 0b01, 0xF7), 1, &OpDispatchBuilder::AVX128_MASKMOV},
-
-    {OPD(1, 0b01, 0xF8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSUB, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xF9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSUB, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xFA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSUB, OpSize::i32Bit>},
-    {OPD(1, 0b01, 0xFB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSUB, OpSize::i64Bit>},
-    {OPD(1, 0b01, 0xFC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VADD, OpSize::i8Bit>},
-    {OPD(1, 0b01, 0xFD), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VADD, OpSize::i16Bit>},
-    {OPD(1, 0b01, 0xFE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VADD, OpSize::i32Bit>},
-
-    {OPD(2, 0b01, 0x00), 1, &OpDispatchBuilder::AVX128_VPSHUFB},
-    {OPD(2, 0b01, 0x01), 1, &OpDispatchBuilder::AVX128_VHADDP<IR::OP_VADDP, OpSize::i16Bit>},
-    {OPD(2, 0b01, 0x02), 1, &OpDispatchBuilder::AVX128_VHADDP<IR::OP_VADDP, OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x03), 1, &OpDispatchBuilder::AVX128_VPHADDSW},
-    {OPD(2, 0b01, 0x04), 1, &OpDispatchBuilder::AVX128_VPMADDUBSW},
-
-    {OPD(2, 0b01, 0x05), 1, &OpDispatchBuilder::AVX128_VPHSUB<OpSize::i16Bit>},
-    {OPD(2, 0b01, 0x06), 1, &OpDispatchBuilder::AVX128_VPHSUB<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x07), 1, &OpDispatchBuilder::AVX128_VPHSUBSW},
-
-    {OPD(2, 0b01, 0x08), 1, &OpDispatchBuilder::AVX128_VPSIGN<OpSize::i8Bit>},
-    {OPD(2, 0b01, 0x09), 1, &OpDispatchBuilder::AVX128_VPSIGN<OpSize::i16Bit>},
-    {OPD(2, 0b01, 0x0A), 1, &OpDispatchBuilder::AVX128_VPSIGN<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x0B), 1, &OpDispatchBuilder::AVX128_VPMULHRSW},
-    {OPD(2, 0b01, 0x0C), 1, &OpDispatchBuilder::AVX128_VPERMILReg<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x0D), 1, &OpDispatchBuilder::AVX128_VPERMILReg<OpSize::i64Bit>},
-    {OPD(2, 0b01, 0x0E), 1, &OpDispatchBuilder::AVX128_VTESTP<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x0F), 1, &OpDispatchBuilder::AVX128_VTESTP<OpSize::i64Bit>},
-
-
-    {OPD(2, 0b01, 0x13), 1, &OpDispatchBuilder::AVX128_VCVTPH2PS},
-    {OPD(2, 0b01, 0x16), 1, &OpDispatchBuilder::AVX128_VPERMD},
-    {OPD(2, 0b01, 0x17), 1, &OpDispatchBuilder::AVX128_PTest},
-    {OPD(2, 0b01, 0x18), 1, &OpDispatchBuilder::AVX128_VBROADCAST<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x19), 1, &OpDispatchBuilder::AVX128_VBROADCAST<OpSize::i64Bit>},
-    {OPD(2, 0b01, 0x1A), 1, &OpDispatchBuilder::AVX128_VBROADCAST<OpSize::i128Bit>},
-    {OPD(2, 0b01, 0x1C), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorUnary, IR::OP_VABS, OpSize::i8Bit>},
-    {OPD(2, 0b01, 0x1D), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorUnary, IR::OP_VABS, OpSize::i16Bit>},
-    {OPD(2, 0b01, 0x1E), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorUnary, IR::OP_VABS, OpSize::i32Bit>},
-
-    {OPD(2, 0b01, 0x20), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i8Bit, OpSize::i16Bit, true>},
-    {OPD(2, 0b01, 0x21), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i8Bit, OpSize::i32Bit, true>},
-    {OPD(2, 0b01, 0x22), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i8Bit, OpSize::i64Bit, true>},
-    {OPD(2, 0b01, 0x23), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i16Bit, OpSize::i32Bit, true>},
-    {OPD(2, 0b01, 0x24), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i16Bit, OpSize::i64Bit, true>},
-    {OPD(2, 0b01, 0x25), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i32Bit, OpSize::i64Bit, true>},
-
-    {OPD(2, 0b01, 0x28), 1, &OpDispatchBuilder::AVX128_VPMULL<OpSize::i32Bit, true>},
-    {OPD(2, 0b01, 0x29), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VCMPEQ, OpSize::i64Bit>},
-    {OPD(2, 0b01, 0x2A), 1, &OpDispatchBuilder::AVX128_MOVVectorNT},
-    {OPD(2, 0b01, 0x2B), 1, &OpDispatchBuilder::AVX128_VPACKUS<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x2C), 1, &OpDispatchBuilder::AVX128_VMASKMOV<OpSize::i32Bit, false>},
-    {OPD(2, 0b01, 0x2D), 1, &OpDispatchBuilder::AVX128_VMASKMOV<OpSize::i64Bit, false>},
-    {OPD(2, 0b01, 0x2E), 1, &OpDispatchBuilder::AVX128_VMASKMOV<OpSize::i32Bit, true>},
-    {OPD(2, 0b01, 0x2F), 1, &OpDispatchBuilder::AVX128_VMASKMOV<OpSize::i64Bit, true>},
-
-    {OPD(2, 0b01, 0x30), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i8Bit, OpSize::i16Bit, false>},
-    {OPD(2, 0b01, 0x31), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i8Bit, OpSize::i32Bit, false>},
-    {OPD(2, 0b01, 0x32), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i8Bit, OpSize::i64Bit, false>},
-    {OPD(2, 0b01, 0x33), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i16Bit, OpSize::i32Bit, false>},
-    {OPD(2, 0b01, 0x34), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i16Bit, OpSize::i64Bit, false>},
-    {OPD(2, 0b01, 0x35), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ExtendVectorElements, OpSize::i32Bit, OpSize::i64Bit, false>},
-    {OPD(2, 0b01, 0x36), 1, &OpDispatchBuilder::AVX128_VPERMD},
-
-    {OPD(2, 0b01, 0x37), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VCMPGT, OpSize::i64Bit>},
-    {OPD(2, 0b01, 0x38), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSMIN, OpSize::i8Bit>},
-    {OPD(2, 0b01, 0x39), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSMIN, OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x3A), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUMIN, OpSize::i16Bit>},
-    {OPD(2, 0b01, 0x3B), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUMIN, OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x3C), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSMAX, OpSize::i8Bit>},
-    {OPD(2, 0b01, 0x3D), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VSMAX, OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x3E), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUMAX, OpSize::i16Bit>},
-    {OPD(2, 0b01, 0x3F), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VUMAX, OpSize::i32Bit>},
-
-    {OPD(2, 0b01, 0x40), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorALU, IR::OP_VMUL, OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x41), 1, &OpDispatchBuilder::AVX128_PHMINPOSUW},
-    {OPD(2, 0b01, 0x45), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VariableShiftImpl, IROps::OP_VUSHR>}, // VPSRLV
-    {OPD(2, 0b01, 0x46), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VariableShiftImpl, IROps::OP_VSSHR>}, // VPSRAVD
-    {OPD(2, 0b01, 0x47), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VariableShiftImpl, IROps::OP_VUSHL>}, // VPSLLV
-
-    {OPD(2, 0b01, 0x58), 1, &OpDispatchBuilder::AVX128_VBROADCAST<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x59), 1, &OpDispatchBuilder::AVX128_VBROADCAST<OpSize::i64Bit>},
-    {OPD(2, 0b01, 0x5A), 1, &OpDispatchBuilder::AVX128_VBROADCAST<OpSize::i128Bit>},
-
-    {OPD(2, 0b01, 0x78), 1, &OpDispatchBuilder::AVX128_VBROADCAST<OpSize::i8Bit>},
-    {OPD(2, 0b01, 0x79), 1, &OpDispatchBuilder::AVX128_VBROADCAST<OpSize::i16Bit>},
-
-    {OPD(2, 0b01, 0x8C), 1, &OpDispatchBuilder::AVX128_VPMASKMOV<false>},
-    {OPD(2, 0b01, 0x8E), 1, &OpDispatchBuilder::AVX128_VPMASKMOV<true>},
-
-    {OPD(2, 0b01, 0x90), 1, &OpDispatchBuilder::AVX128_VPGATHER<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x91), 1, &OpDispatchBuilder::AVX128_VPGATHER<OpSize::i64Bit>},
-    {OPD(2, 0b01, 0x92), 1, &OpDispatchBuilder::AVX128_VPGATHER<OpSize::i32Bit>},
-    {OPD(2, 0b01, 0x93), 1, &OpDispatchBuilder::AVX128_VPGATHER<OpSize::i64Bit>},
-
-    {OPD(2, 0b01, 0x96), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, true, 1, 3, 2>},  // VFMADDSUB
-    {OPD(2, 0b01, 0x97), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, false, 1, 3, 2>}, // VFMSUBADD
-
-    {OPD(2, 0b01, 0x98), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFMLA, 1, 3, 2>}, // VFMADD
-    {OPD(2, 0b01, 0x99), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFMLASCALARINSERT, 1, 3, 2>}, // VFMADD
-    {OPD(2, 0b01, 0x9A), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFMLS, 1, 3, 2>}, // VFMSUB
-    {OPD(2, 0b01, 0x9B), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFMLSSCALARINSERT, 1, 3, 2>}, // VFMSUB
-    {OPD(2, 0b01, 0x9C), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFNMLA, 1, 3, 2>}, // VFNMADD
-    {OPD(2, 0b01, 0x9D), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFNMLASCALARINSERT, 1, 3, 2>}, // VFNMADD
-    {OPD(2, 0b01, 0x9E), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFNMLS, 1, 3, 2>}, // VFNMSUB
-    {OPD(2, 0b01, 0x9F), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFNMLSSCALARINSERT, 1, 3, 2>}, // VFNMSUB
-
-    {OPD(2, 0b01, 0xA8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFMLA, 2, 1, 3>}, // VFMADD
-    {OPD(2, 0b01, 0xA9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFMLASCALARINSERT, 2, 1, 3>}, // VFMADD
-    {OPD(2, 0b01, 0xAA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFMLS, 2, 1, 3>}, // VFMSUB
-    {OPD(2, 0b01, 0xAB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFMLSSCALARINSERT, 2, 1, 3>}, // VFMSUB
-    {OPD(2, 0b01, 0xAC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFNMLA, 2, 1, 3>}, // VFNMADD
-    {OPD(2, 0b01, 0xAD), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFNMLASCALARINSERT, 2, 1, 3>}, // VFNMADD
-    {OPD(2, 0b01, 0xAE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFNMLS, 2, 1, 3>}, // VFNMSUB
-    {OPD(2, 0b01, 0xAF), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFNMLSSCALARINSERT, 2, 1, 3>}, // VFNMSUB
-
-    {OPD(2, 0b01, 0xB8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFMLA, 2, 3, 1>}, // VFMADD
-    {OPD(2, 0b01, 0xB9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFMLASCALARINSERT, 2, 3, 1>}, // VFMADD
-    {OPD(2, 0b01, 0xBA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFMLS, 2, 3, 1>}, // VFMSUB
-    {OPD(2, 0b01, 0xBB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFMLSSCALARINSERT, 2, 3, 1>}, // VFMSUB
-    {OPD(2, 0b01, 0xBC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFNMLA, 2, 3, 1>}, // VFNMADD
-    {OPD(2, 0b01, 0xBD), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFNMLASCALARINSERT, 2, 3, 1>}, // VFNMADD
-    {OPD(2, 0b01, 0xBE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAImpl, IR::OP_VFNMLS, 2, 3, 1>}, // VFNMSUB
-    {OPD(2, 0b01, 0xBF), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAScalarImpl, IR::OP_VFNMLSSCALARINSERT, 2, 3, 1>}, // VFNMSUB
-
-    {OPD(2, 0b01, 0xA6), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, true, 2, 1, 3>},  // VFMADDSUB
-    {OPD(2, 0b01, 0xA7), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, false, 2, 1, 3>}, // VFMSUBADD
-
-    {OPD(2, 0b01, 0xB6), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, true, 2, 3, 1>},  // VFMADDSUB
-    {OPD(2, 0b01, 0xB7), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, false, 2, 3, 1>}, // VFMSUBADD
-
-    {OPD(2, 0b01, 0xDB), 1, &OpDispatchBuilder::AVX128_VAESImc},
-    {OPD(2, 0b01, 0xDC), 1, &OpDispatchBuilder::AVX128_VAESEnc},
-    {OPD(2, 0b01, 0xDD), 1, &OpDispatchBuilder::AVX128_VAESEncLast},
-    {OPD(2, 0b01, 0xDE), 1, &OpDispatchBuilder::AVX128_VAESDec},
-    {OPD(2, 0b01, 0xDF), 1, &OpDispatchBuilder::AVX128_VAESDecLast},
-
-    {OPD(3, 0b01, 0x00), 1, &OpDispatchBuilder::AVX128_VPERMQ},
-    {OPD(3, 0b01, 0x01), 1, &OpDispatchBuilder::AVX128_VPERMQ},
-    {OPD(3, 0b01, 0x02), 1, &OpDispatchBuilder::AVX128_VBLEND<OpSize::i32Bit>},
-    {OPD(3, 0b01, 0x04), 1, &OpDispatchBuilder::AVX128_VPERMILImm<OpSize::i32Bit>},
-    {OPD(3, 0b01, 0x05), 1, &OpDispatchBuilder::AVX128_VPERMILImm<OpSize::i64Bit>},
-    {OPD(3, 0b01, 0x06), 1, &OpDispatchBuilder::AVX128_VPERM2},
-    {OPD(3, 0b01, 0x08), 1, &OpDispatchBuilder::AVX128_VectorRound<OpSize::i32Bit>},
-    {OPD(3, 0b01, 0x09), 1, &OpDispatchBuilder::AVX128_VectorRound<OpSize::i64Bit>},
-    {OPD(3, 0b01, 0x0A), 1, &OpDispatchBuilder::AVX128_InsertScalarRound<OpSize::i32Bit>},
-    {OPD(3, 0b01, 0x0B), 1, &OpDispatchBuilder::AVX128_InsertScalarRound<OpSize::i64Bit>},
-    {OPD(3, 0b01, 0x0C), 1, &OpDispatchBuilder::AVX128_VBLEND<OpSize::i32Bit>},
-    {OPD(3, 0b01, 0x0D), 1, &OpDispatchBuilder::AVX128_VBLEND<OpSize::i64Bit>},
-    {OPD(3, 0b01, 0x0E), 1, &OpDispatchBuilder::AVX128_VBLEND<OpSize::i16Bit>},
-    {OPD(3, 0b01, 0x0F), 1, &OpDispatchBuilder::AVX128_VPALIGNR},
-
-    {OPD(3, 0b01, 0x14), 1, &OpDispatchBuilder::AVX128_PExtr<OpSize::i8Bit>},
-    {OPD(3, 0b01, 0x15), 1, &OpDispatchBuilder::AVX128_PExtr<OpSize::i16Bit>},
-    {OPD(3, 0b01, 0x16), 1, &OpDispatchBuilder::AVX128_PExtr<OpSize::i32Bit>},
-    {OPD(3, 0b01, 0x17), 1, &OpDispatchBuilder::AVX128_PExtr<OpSize::i32Bit>},
-
-    {OPD(3, 0b01, 0x18), 1, &OpDispatchBuilder::AVX128_VINSERT},
-    {OPD(3, 0b01, 0x19), 1, &OpDispatchBuilder::AVX128_VEXTRACT128},
-    {OPD(3, 0b01, 0x1D), 1, &OpDispatchBuilder::AVX128_VCVTPS2PH},
-    {OPD(3, 0b01, 0x20), 1, &OpDispatchBuilder::AVX128_VPINSRB},
-    {OPD(3, 0b01, 0x21), 1, &OpDispatchBuilder::AVX128_VINSERTPS},
-    {OPD(3, 0b01, 0x22), 1, &OpDispatchBuilder::AVX128_VPINSRDQ},
-
-    {OPD(3, 0b01, 0x38), 1, &OpDispatchBuilder::AVX128_VINSERT},
-    {OPD(3, 0b01, 0x39), 1, &OpDispatchBuilder::AVX128_VEXTRACT128},
-
-    {OPD(3, 0b01, 0x40), 1, &OpDispatchBuilder::AVX128_VDPP<OpSize::i32Bit>},
-    {OPD(3, 0b01, 0x41), 1, &OpDispatchBuilder::AVX128_VDPP<OpSize::i64Bit>},
-    {OPD(3, 0b01, 0x42), 1, &OpDispatchBuilder::AVX128_VMPSADBW},
-
-    {OPD(3, 0b01, 0x46), 1, &OpDispatchBuilder::AVX128_VPERM2},
-
-    {OPD(3, 0b01, 0x4A), 1, &OpDispatchBuilder::AVX128_VectorVariableBlend<OpSize::i32Bit>},
-    {OPD(3, 0b01, 0x4B), 1, &OpDispatchBuilder::AVX128_VectorVariableBlend<OpSize::i64Bit>},
-    {OPD(3, 0b01, 0x4C), 1, &OpDispatchBuilder::AVX128_VectorVariableBlend<OpSize::i8Bit>},
-
-    {OPD(3, 0b01, 0x60), 1, &OpDispatchBuilder::AVX128_VPCMPESTRM},
-    {OPD(3, 0b01, 0x61), 1, &OpDispatchBuilder::AVX128_VPCMPESTRI},
-    {OPD(3, 0b01, 0x62), 1, &OpDispatchBuilder::AVX128_VPCMPISTRM},
-    {OPD(3, 0b01, 0x63), 1, &OpDispatchBuilder::AVX128_VPCMPISTRI},
-
-    {OPD(3, 0b01, 0xDF), 1, &OpDispatchBuilder::AVX128_VAESKeyGenAssist},
-  };
-#undef OPD
-
-#define OPD(group, pp, opcode) (((group - X86Tables::TYPE_VEX_GROUP_12) << 4) | (pp << 3) | (opcode))
-  static constexpr DispatchTableEntry VEX128TableGroupOps[] {
-    // VPSRLI
-    {OPD(X86Tables::TYPE_VEX_GROUP_12, 1, 0b010), 1,
-     &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftImmImpl, OpSize::i16Bit, IROps::OP_VUSHRI>},
-    // VPSLLI
-    {OPD(X86Tables::TYPE_VEX_GROUP_12, 1, 0b110), 1,
-     &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftImmImpl, OpSize::i16Bit, IROps::OP_VSHLI>},
-    // VPSRAI
-    {OPD(X86Tables::TYPE_VEX_GROUP_12, 1, 0b100), 1,
-     &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftImmImpl, OpSize::i16Bit, IROps::OP_VSSHRI>},
-
-    // VPSRLI
-    {OPD(X86Tables::TYPE_VEX_GROUP_13, 1, 0b010), 1,
-     &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftImmImpl, OpSize::i32Bit, IROps::OP_VUSHRI>},
-    // VPSLLI
-    {OPD(X86Tables::TYPE_VEX_GROUP_13, 1, 0b110), 1,
-     &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftImmImpl, OpSize::i32Bit, IROps::OP_VSHLI>},
-    // VPSRAI
-    {OPD(X86Tables::TYPE_VEX_GROUP_13, 1, 0b100), 1,
-     &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftImmImpl, OpSize::i32Bit, IROps::OP_VSSHRI>},
-
-    // VPSRLI
-    {OPD(X86Tables::TYPE_VEX_GROUP_14, 1, 0b010), 1,
-     &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftImmImpl, OpSize::i64Bit, IROps::OP_VUSHRI>},
-    // VPSRLDQ
-    {OPD(X86Tables::TYPE_VEX_GROUP_14, 1, 0b011), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ShiftDoubleImm, ShiftDirection::RIGHT>},
-    // VPSLLI
-    {OPD(X86Tables::TYPE_VEX_GROUP_14, 1, 0b110), 1,
-     &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VectorShiftImmImpl, OpSize::i64Bit, IROps::OP_VSHLI>},
-    // VPSLLDQ
-    {OPD(X86Tables::TYPE_VEX_GROUP_14, 1, 0b111), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_ShiftDoubleImm, ShiftDirection::LEFT>},
-
-    ///< Use the regular implementation. It just happens to be in the VEX table.
-    {OPD(X86Tables::TYPE_VEX_GROUP_15, 0, 0b010), 1, &OpDispatchBuilder::LDMXCSR},
-    {OPD(X86Tables::TYPE_VEX_GROUP_15, 0, 0b011), 1, &OpDispatchBuilder::STMXCSR},
-  };
-#undef OPD
-
-#define OPD(map_select, pp, opcode) (((map_select - 1) << 10) | (pp << 8) | (opcode))
-  constexpr DispatchTableEntry VEX128_PCLMUL[] = {
-    {OPD(3, 0b01, 0x44), 1, &OpDispatchBuilder::AVX128_VPCLMULQDQ},
-  };
-#undef OPD
-
-  InstallToTable(FEXCore::X86Tables::VEXTableOps, AVX128Table);
-  InstallToTable(FEXCore::X86Tables::VEXTableGroupOps, VEX128TableGroupOps);
-  if (CTX->HostFeatures.SupportsPMULL_128Bit) {
-    InstallToTable(FEXCore::X86Tables::VEXTableOps, VEX128_PCLMUL);
-  }
-
-  SaveAVXStateFunc = &OpDispatchBuilder::AVX128_SaveAVXState;
-  RestoreAVXStateFunc = &OpDispatchBuilder::AVX128_RestoreAVXState;
-  DefaultAVXStateFunc = &OpDispatchBuilder::AVX128_DefaultAVXState;
-}
-
 OpDispatchBuilder::RefPair OpDispatchBuilder::AVX128_LoadSource_WithOpSize(
   const X86Tables::DecodedOp& Op, const X86Tables::DecodedOperand& Operand, uint32_t Flags, bool NeedsHigh, MemoryAccessType AccessType) {
 
@@ -501,7 +41,7 @@ OpDispatchBuilder::RefPair OpDispatchBuilder::AVX128_LoadSource_WithOpSize(
     HighA.Offset += 16;
 
     if (Operand.IsSIB()) {
-      [[maybe_unused]] const bool IsVSIB = (Op->Flags & X86Tables::DecodeFlags::FLAG_VSIB_BYTE) != 0;
+      const bool IsVSIB = (Op->Flags & X86Tables::DecodeFlags::FLAG_VSIB_BYTE) != 0;
       LOGMAN_THROW_A_FMT(!IsVSIB, "VSIB uses LoadVSIB instead");
     }
 
@@ -515,7 +55,7 @@ OpDispatchBuilder::RefPair OpDispatchBuilder::AVX128_LoadSource_WithOpSize(
 
 OpDispatchBuilder::RefVSIB
 OpDispatchBuilder::AVX128_LoadVSIB(const X86Tables::DecodedOp& Op, const X86Tables::DecodedOperand& Operand, uint32_t Flags, bool NeedsHigh) {
-  [[maybe_unused]] const bool IsVSIB = (Op->Flags & X86Tables::DecodeFlags::FLAG_VSIB_BYTE) != 0;
+  const bool IsVSIB = (Op->Flags & X86Tables::DecodeFlags::FLAG_VSIB_BYTE) != 0;
   LOGMAN_THROW_A_FMT(Operand.IsSIB() && IsVSIB, "Trying to load VSIB for something that isn't the correct type!");
 
   // VSIB is a very special case which has a ton of encoded data.
@@ -954,8 +494,7 @@ void OpDispatchBuilder::AVX128_VMOVSHDUP(OpcodeArgs) {
                          [this](IR::OpSize ElementSize, Ref Src) { return _VTrn2(OpSize::i128Bit, ElementSize, Src, Src); });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VBROADCAST(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VBROADCAST(OpcodeArgs, IR::OpSize ElementSize) {
   const auto DstSize = GetDstSize(Op);
   const auto Is128Bit = DstSize == Core::CPUState::XMM_SSE_REG_SIZE;
   RefPair Src {};
@@ -981,14 +520,12 @@ void OpDispatchBuilder::AVX128_VBROADCAST(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Src);
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VPUNPCKL(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPUNPCKL(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize,
                           [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) { return _VZip(OpSize::i128Bit, _ElementSize, Src1, Src2); });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VPUNPCKH(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPUNPCKH(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize,
                           [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) { return _VZip2(OpSize::i128Bit, _ElementSize, Src1, Src2); });
 }
@@ -1011,8 +548,7 @@ void OpDispatchBuilder::AVX128_MOVVectorUnaligned(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Src);
 }
 
-template<IR::OpSize DstElementSize>
-void OpDispatchBuilder::AVX128_InsertCVTGPR_To_FPR(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_InsertCVTGPR_To_FPR(OpcodeArgs, IR::OpSize DstElementSize) {
   const auto SrcSize = OpSizeFromSrc(Op);
   const auto DstSize = OpSizeFromDst(Op);
 
@@ -1033,20 +569,19 @@ void OpDispatchBuilder::AVX128_InsertCVTGPR_To_FPR(OpcodeArgs) {
   } else {
     // In the case of cvtsi2s{s,d} where the source and destination are the same size,
     // then it is more optimal to load in to the FPR register directly and convert there.
-    auto Src2 = AVX128_LoadSource_WithOpSize(Op, Op->Src[1], Op->Flags, false);
+    auto Src2 = LoadSource(FPRClass, Op, Op->Src[1], Op->Flags);
     // Always signed
-    Result.Low = _VSToFVectorInsert(DstSize, DstElementSize, DstElementSize, Src1.Low, Src2.Low, false, false);
+    Result.Low = _VSToFVectorInsert(DstSize, DstElementSize, DstElementSize, Src1.Low, Src2, false, false);
   }
 
-  [[maybe_unused]] const auto Is128Bit = DstSize == OpSize::i128Bit;
+  const auto Is128Bit = DstSize == OpSize::i128Bit;
   LOGMAN_THROW_A_FMT(Is128Bit, "Programming Error: This should never occur!");
   Result.High = LoadZeroVector(OpSize::i128Bit);
 
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
-template<IR::OpSize SrcElementSize, bool HostRoundingMode>
-void OpDispatchBuilder::AVX128_CVTFPR_To_GPR(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_CVTFPR_To_GPR(OpcodeArgs, IR::OpSize SrcElementSize, bool HostRoundingMode) {
   // If loading a vector, use the full size, so we don't
   // unnecessarily zero extend the vector. Otherwise, if
   // memory, then we want to load the element size exactly.
@@ -1066,15 +601,13 @@ void OpDispatchBuilder::AVX128_VANDN(OpcodeArgs) {
                           [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) { return _VAndn(OpSize::i128Bit, _ElementSize, Src2, Src1); });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VPACKSS(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPACKSS(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize, [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) {
     return _VSQXTNPair(OpSize::i128Bit, _ElementSize, Src1, Src2);
   });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VPACKUS(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPACKUS(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize, [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) {
     return _VSQXTUNPair(OpSize::i128Bit, _ElementSize, Src1, Src2);
   });
@@ -1086,14 +619,12 @@ Ref OpDispatchBuilder::AVX128_PSIGNImpl(IR::OpSize ElementSize, Ref Src1, Ref Sr
   return _VMul(OpSize::i128Bit, ElementSize, Src1, Control);
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VPSIGN(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPSIGN(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize,
                           [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) { return AVX128_PSIGNImpl(_ElementSize, Src1, Src2); });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_UCOMISx(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_UCOMISx(OpcodeArgs, IR::OpSize ElementSize) {
   const auto SrcSize = Op->Src[0].IsGPR() ? GetGuestVectorLength() : ElementSize;
 
   auto Src1 = AVX128_LoadSource_WithOpSize(Op, Op->Dest, Op->Flags, false);
@@ -1131,8 +662,7 @@ void OpDispatchBuilder::AVX128_VectorScalarInsertALU(OpcodeArgs, FEXCore::IR::IR
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, RefPair {.Low = Result_Low, .High = High});
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VFCMP(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VFCMP(OpcodeArgs, IR::OpSize ElementSize) {
   const uint8_t CompType = Op->Src[2].Literal();
 
   struct {
@@ -1148,8 +678,7 @@ void OpDispatchBuilder::AVX128_VFCMP(OpcodeArgs) {
   });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_InsertScalarFCMP(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_InsertScalarFCMP(OpcodeArgs, IR::OpSize ElementSize) {
   // We load the full vector width when dealing with a source vector,
   // so that we don't do any unnecessary zero extension to the scalar
   // element that we're going to operate on.
@@ -1206,8 +735,7 @@ void OpDispatchBuilder::AVX128_MOVBetweenGPR_FPR(OpcodeArgs) {
   }
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_PExtr(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_PExtr(OpcodeArgs, IR::OpSize ElementSize) {
   const auto DstSize = OpSizeFromDst(Op);
 
   auto Src = AVX128_LoadSource_WithOpSize(Op, Op->Src[0], Op->Flags, false);
@@ -1218,7 +746,7 @@ void OpDispatchBuilder::AVX128_PExtr(OpcodeArgs) {
   // is the same except that REX.W or VEX.W is set to 1. Incredibly frustrating.
   // Use the destination size as the element size in this case.
   auto OverridenElementSize = ElementSize;
-  if constexpr (ElementSize == OpSize::i32Bit) {
+  if (ElementSize == OpSize::i32Bit) {
     OverridenElementSize = DstSize;
   }
 
@@ -1291,8 +819,7 @@ void OpDispatchBuilder::AVX128_ExtendVectorElements(OpcodeArgs, IR::OpSize Eleme
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_MOVMSK(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_MOVMSK(OpcodeArgs, IR::OpSize ElementSize) {
   const auto SrcSize = OpSizeFromSrc(Op);
   const auto Is128Bit = SrcSize == OpSize::i128Bit;
 
@@ -1474,8 +1001,7 @@ void OpDispatchBuilder::AVX128_VINSERTPS(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, AVX128_Zext(Result));
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VPHSUB(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPHSUB(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromDst(Op), ElementSize, [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) {
     return PHSUBOpImpl(OpSize::i128Bit, Src1, Src2, _ElementSize);
   });
@@ -1486,19 +1012,17 @@ void OpDispatchBuilder::AVX128_VPHSUBSW(OpcodeArgs) {
                           [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) { return PHSUBSOpImpl(OpSize::i128Bit, Src1, Src2); });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VADDSUBP(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VADDSUBP(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromDst(Op), ElementSize, [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) {
     return ADDSUBPOpImpl(OpSize::i128Bit, _ElementSize, Src1, Src2);
   });
 }
 
-template<IR::OpSize ElementSize, bool Signed>
-void OpDispatchBuilder::AVX128_VPMULL(OpcodeArgs) {
-  static_assert(ElementSize == OpSize::i32Bit, "Currently only handles 32-bit -> 64-bit");
+void OpDispatchBuilder::AVX128_VPMULL(OpcodeArgs, IR::OpSize ElementSize, bool Signed) {
+  LOGMAN_THROW_A_FMT(ElementSize == OpSize::i32Bit, "Currently only handles 32-bit -> 64-bit");
 
-  AVX128_VectorBinaryImpl(Op, OpSizeFromDst(Op), ElementSize, [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) -> Ref {
-    return PMULLOpImpl(OpSize::i128Bit, ElementSize, Signed, Src1, Src2);
+  AVX128_VectorBinaryImpl(Op, OpSizeFromDst(Op), ElementSize, [&](IR::OpSize _ElementSize, Ref Src1, Ref Src2) -> Ref {
+    return PMULLOpImpl(OpSize::i128Bit, _ElementSize, Signed, Src1, Src2);
   });
 }
 
@@ -1507,9 +1031,8 @@ void OpDispatchBuilder::AVX128_VPMULHRSW(OpcodeArgs) {
                           [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) -> Ref { return PMULHRSWOpImpl(OpSize::i128Bit, Src1, Src2); });
 }
 
-template<bool Signed>
-void OpDispatchBuilder::AVX128_VPMULHW(OpcodeArgs) {
-  AVX128_VectorBinaryImpl(Op, OpSizeFromDst(Op), OpSize::i16Bit, [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) -> Ref {
+void OpDispatchBuilder::AVX128_VPMULHW(OpcodeArgs, bool Signed) {
+  AVX128_VectorBinaryImpl(Op, OpSizeFromDst(Op), OpSize::i16Bit, [&](IR::OpSize _ElementSize, Ref Src1, Ref Src2) -> Ref {
     if (Signed) {
       return _VSMulH(OpSize::i128Bit, _ElementSize, Src1, Src2);
     } else {
@@ -1518,12 +1041,11 @@ void OpDispatchBuilder::AVX128_VPMULHW(OpcodeArgs) {
   });
 }
 
-template<IR::OpSize DstElementSize, IR::OpSize SrcElementSize>
-void OpDispatchBuilder::AVX128_InsertScalar_CVT_Float_To_Float(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_InsertScalar_CVT_Float_To_Float(OpcodeArgs, IR::OpSize DstElementSize, IR::OpSize SrcElementSize) {
   // Gotta be careful with this operation.
   // It inserts in to the lowest element, retaining the remainder of the lower 128-bits.
   // Then zero extends the top 128-bit.
-  const auto SrcSize = OpSizeFromSrc(Op);
+  const auto SrcSize = Op->Src[1].IsGPR() ? OpSize::i128Bit : SrcElementSize;
   auto Src1 = AVX128_LoadSource_WithOpSize(Op, Op->Src[0], Op->Flags, false);
   Ref Src2 = LoadSource_WithOpSize(FPRClass, Op, Op->Src[1], SrcSize, Op->Flags, {.AllowUpperGarbage = true});
 
@@ -1531,8 +1053,7 @@ void OpDispatchBuilder::AVX128_InsertScalar_CVT_Float_To_Float(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, AVX128_Zext(Result));
 }
 
-template<IR::OpSize DstElementSize, IR::OpSize SrcElementSize>
-void OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Float(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Float(OpcodeArgs, IR::OpSize DstElementSize, IR::OpSize SrcElementSize) {
   const auto SrcSize = OpSizeFromSrc(Op);
   const auto DstSize = OpSizeFromDst(Op);
 
@@ -1560,11 +1081,11 @@ void OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Float(OpcodeArgs) {
 
   RefPair Result {};
 
-  auto TransformLow = [this](Ref Src) -> Ref {
+  auto TransformLow = [&](Ref Src) -> Ref {
     return _Vector_FToF(OpSize::i128Bit, DstElementSize, Src, SrcElementSize);
   };
 
-  auto TransformHigh = [this](Ref Src) -> Ref {
+  auto TransformHigh = [&](Ref Src) -> Ref {
     return _VFCVTL2(OpSize::i128Bit, SrcElementSize, Src);
   };
 
@@ -1595,8 +1116,7 @@ void OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Float(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
-template<IR::OpSize SrcElementSize, bool HostRoundingMode>
-void OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Int(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Int(OpcodeArgs, IR::OpSize SrcElementSize, bool HostRoundingMode) {
   const auto SrcSize = GetSrcSize(Op);
 
   const auto Is128BitSrc = SrcSize == Core::CPUState::XMM_SSE_REG_SIZE;
@@ -1624,8 +1144,7 @@ void OpDispatchBuilder::AVX128_Vector_CVT_Float_To_Int(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
-template<IR::OpSize SrcElementSize, bool Widen>
-void OpDispatchBuilder::AVX128_Vector_CVT_Int_To_Float(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_Vector_CVT_Int_To_Float(OpcodeArgs, IR::OpSize SrcElementSize, bool Widen) {
   const auto Size = OpSizeFromDst(Op);
   const auto Is128Bit = Size == OpSize::i128Bit;
 
@@ -1641,7 +1160,7 @@ void OpDispatchBuilder::AVX128_Vector_CVT_Int_To_Float(OpcodeArgs) {
     }
   }();
 
-  auto Convert = [this](Ref Src, IROps Op) -> Ref {
+  auto Convert = [&](Ref Src, IROps Op) -> Ref {
     auto ElementSize = SrcElementSize;
     if (Widen) {
       DeriveOp(Extended, Op, _VSXTL(OpSize::i128Bit, ElementSize, Src));
@@ -1766,17 +1285,15 @@ void OpDispatchBuilder::AVX128_PHMINPOSUW(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, AVX128_Zext(Result));
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VectorRound(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VectorRound(OpcodeArgs, IR::OpSize ElementSize) {
   const auto Size = OpSizeFromSrc(Op);
   const auto Mode = Op->Src[1].Literal();
 
   AVX128_VectorUnaryImpl(Op, Size, ElementSize,
-                         [this, Mode](IR::OpSize, Ref Src) { return VectorRoundImpl(OpSize::i128Bit, ElementSize, Src, Mode); });
+                         [this, Mode](IR::OpSize ElementSize, Ref Src) { return VectorRoundImpl(OpSize::i128Bit, ElementSize, Src, Mode); });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_InsertScalarRound(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_InsertScalarRound(OpcodeArgs, IR::OpSize ElementSize) {
   // We load the full vector width when dealing with a source vector,
   // so that we don't do any unnecessary zero extension to the scalar
   // element that we're going to operate on.
@@ -1797,11 +1314,10 @@ void OpDispatchBuilder::AVX128_InsertScalarRound(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, AVX128_Zext(Result));
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VDPP(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VDPP(OpcodeArgs, IR::OpSize ElementSize) {
   const uint64_t Literal = Op->Src[2].Literal();
 
-  AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize, [this, Literal](IR::OpSize, Ref Src1, Ref Src2) {
+  AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize, [this, Literal](IR::OpSize ElementSize, Ref Src1, Ref Src2) {
     return DPPOpImpl(OpSize::i128Bit, Src1, Src2, Literal, ElementSize);
   });
 }
@@ -1869,8 +1385,7 @@ void OpDispatchBuilder::AVX128_VPSHUFW(OpcodeArgs, bool Low) {
   });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VSHUF(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VSHUF(OpcodeArgs, IR::OpSize ElementSize) {
   const auto SrcSize = GetSrcSize(Op);
   const auto Is128Bit = SrcSize == Core::CPUState::XMM_SSE_REG_SIZE;
   auto Shuffle = Op->Src[2].Literal();
@@ -1884,14 +1399,13 @@ void OpDispatchBuilder::AVX128_VSHUF(OpcodeArgs) {
   if (Is128Bit) {
     Result.High = LoadZeroVector(OpSize::i128Bit);
   } else {
-    constexpr uint8_t ShiftAmount = ElementSize == OpSize::i32Bit ? 0 : 2;
+    const uint8_t ShiftAmount = ElementSize == OpSize::i32Bit ? 0 : 2;
     Result.High = SHUFOpImpl(Op, OpSize::i128Bit, ElementSize, Src1.High, Src2.High, Shuffle >> ShiftAmount);
   }
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VPERMILImm(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPERMILImm(OpcodeArgs, IR::OpSize ElementSize) {
   const auto SrcSize = GetSrcSize(Op);
   const auto Is128Bit = SrcSize == Core::CPUState::XMM_SSE_REG_SIZE;
 
@@ -1900,7 +1414,7 @@ void OpDispatchBuilder::AVX128_VPERMILImm(OpcodeArgs) {
 
   RefPair Result = AVX128_Zext(LoadZeroVector(OpSize::i128Bit));
 
-  if constexpr (ElementSize == OpSize::i64Bit) {
+  if (ElementSize == OpSize::i64Bit) {
     auto DoSwizzle64 = [this](Ref Src, uint8_t Selector) -> Ref {
       switch (Selector) {
       case 0b00:
@@ -1928,9 +1442,8 @@ void OpDispatchBuilder::AVX128_VPERMILImm(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
-template<IROps IROp, IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VHADDP(OpcodeArgs) {
-  AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize, [this](IR::OpSize, Ref Src1, Ref Src2) {
+void OpDispatchBuilder::AVX128_VHADDP(OpcodeArgs, IROps IROp, IR::OpSize ElementSize) {
+  AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize, [&](IR::OpSize ElementSize, Ref Src1, Ref Src2) {
     DeriveOp(Res, IROp, _VFAddP(OpSize::i128Bit, ElementSize, Src1, Src2));
     return Res;
   });
@@ -1951,8 +1464,7 @@ void OpDispatchBuilder::AVX128_VPMADDWD(OpcodeArgs) {
                           [this](IR::OpSize _ElementSize, Ref Src1, Ref Src2) { return PMADDWDOpImpl(OpSize::i128Bit, Src1, Src2); });
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VBLEND(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VBLEND(OpcodeArgs, IR::OpSize ElementSize) {
   const auto SrcSize = OpSizeFromSrc(Op);
   const auto Is128Bit = SrcSize == OpSize::i128Bit;
   const uint64_t Selector = Op->Src[2].Literal();
@@ -1961,7 +1473,7 @@ void OpDispatchBuilder::AVX128_VBLEND(OpcodeArgs) {
   /// i16Bit: Reuses same bits, no shift
   /// i32Bit: Shift by 4
   /// i64Bit: Shift by 2
-  constexpr uint64_t SelectorShift = ElementSize == OpSize::i64Bit ? 2 : ElementSize == OpSize::i32Bit ? 4 : 0;
+  const uint64_t SelectorShift = ElementSize == OpSize::i64Bit ? 2 : ElementSize == OpSize::i32Bit ? 4 : 0;
 
   auto Src1 = AVX128_LoadSource_WithOpSize(Op, Op->Src[0], Op->Flags, !Is128Bit);
   auto Src2 = AVX128_LoadSource_WithOpSize(Op, Op->Src[1], Op->Flags, !Is128Bit);
@@ -1978,10 +1490,9 @@ void OpDispatchBuilder::AVX128_VBLEND(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VHSUBP(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VHSUBP(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromDst(Op), ElementSize,
-                          [this](IR::OpSize, Ref Src1, Ref Src2) { return HSUBPOpImpl(OpSize::i128Bit, ElementSize, Src1, Src2); });
+                          [&](IR::OpSize, Ref Src1, Ref Src2) { return HSUBPOpImpl(OpSize::i128Bit, ElementSize, Src1, Src2); });
 }
 
 void OpDispatchBuilder::AVX128_VPSHUFB(OpcodeArgs) {
@@ -2081,13 +1592,11 @@ void OpDispatchBuilder::AVX128_VMASKMOVImpl(OpcodeArgs, IR::OpSize ElementSize,
   }
 }
 
-template<bool IsStore>
-void OpDispatchBuilder::AVX128_VPMASKMOV(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPMASKMOV(OpcodeArgs, bool IsStore) {
   AVX128_VMASKMOVImpl(Op, OpSizeFromSrc(Op), OpSizeFromDst(Op), IsStore, Op->Src[0], Op->Src[1]);
 }
 
-template<IR::OpSize ElementSize, bool IsStore>
-void OpDispatchBuilder::AVX128_VMASKMOV(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VMASKMOV(OpcodeArgs, IR::OpSize ElementSize, bool IsStore) {
   AVX128_VMASKMOVImpl(Op, ElementSize, OpSizeFromDst(Op), IsStore, Op->Src[0], Op->Src[1]);
 }
 
@@ -2114,8 +1623,7 @@ void OpDispatchBuilder::AVX128_MASKMOV(OpcodeArgs) {
   _StoreMem(FPRClass, Size, MemDest, XMMReg, OpSize::i8Bit);
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VectorVariableBlend(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VectorVariableBlend(OpcodeArgs, IR::OpSize ElementSize) {
   const auto Size = OpSizeFromSrc(Op);
   const auto Is128Bit = Size == OpSize::i128Bit;
   const auto Src3Selector = Op->Src[2].Literal();
@@ -2130,7 +1638,7 @@ void OpDispatchBuilder::AVX128_VectorVariableBlend(OpcodeArgs) {
     Mask.High = AVX128_LoadXMMRegister(MaskRegister, true);
   }
 
-  auto Convert = [this](Ref Src1, Ref Src2, Ref Mask) {
+  auto Convert = [&](Ref Src1, Ref Src2, Ref Mask) {
     const auto ElementSizeBits = IR::OpSizeAsBits(ElementSize);
     Ref Shifted = _VSShrI(OpSize::i128Bit, ElementSize, Mask, ElementSizeBits - 1);
     return _VBSL(OpSize::i128Bit, Shifted, Src2, Src1);
@@ -2195,8 +1703,7 @@ void OpDispatchBuilder::AVX128_VPERM2(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VTESTP(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VTESTP(OpcodeArgs, IR::OpSize ElementSize) {
   const auto Size = GetSrcSize(Op);
   const auto Is128Bit = Size == Core::CPUState::XMM_SSE_REG_SIZE;
 
@@ -2212,8 +1719,6 @@ void OpDispatchBuilder::AVX128_VTESTP(OpcodeArgs) {
   // For 256-bit, we need to split up the operation. This is nontrivial.
   // Let's go the simple route here.
   Ref ZF, CFInv;
-  Ref ZeroConst = Constant(0);
-  Ref OneConst = Constant(1);
 
   const auto ElementSizeInBits = IR::OpSizeAsBits(ElementSize);
 
@@ -2255,7 +1760,7 @@ void OpDispatchBuilder::AVX128_VTESTP(OpcodeArgs) {
 
     // ExtGPR will either be [0, 8] or [0, 16] If 0 then set Flag.
     auto ExtGPR = _VExtractToGPR(OpSize::i128Bit, ElementSize, AddWide, 0);
-    CFInv = _Select(IR::COND_NEQ, ExtGPR, ZeroConst, OneConst, ZeroConst);
+    CFInv = To01(OpSize::i64Bit, ExtGPR);
   }
 
   // As in PTest, this sets Z appropriately while zeroing the rest of NZCV.
@@ -2294,10 +1799,7 @@ void OpDispatchBuilder::AVX128_PTest(OpcodeArgs) {
   Test1 = _VExtractToGPR(OpSize::i128Bit, OpSize::i16Bit, Test1, 0);
   Test2 = _VExtractToGPR(OpSize::i128Bit, OpSize::i16Bit, Test2, 0);
 
-  auto ZeroConst = Constant(0);
-  auto OneConst = Constant(1);
-
-  Test2 = _Select(FEXCore::IR::COND_NEQ, Test2, ZeroConst, OneConst, ZeroConst);
+  Test2 = To01(OpSize::i64Bit, Test2);
 
   // Careful, these flags are different between {V,}PTEST and VTESTP{S,D}
   // Set ZF according to Test1. SF will be zeroed since we do a 32-bit test on
@@ -2308,10 +1810,9 @@ void OpDispatchBuilder::AVX128_PTest(OpcodeArgs) {
   ZeroPF_AF();
 }
 
-template<IR::OpSize ElementSize>
-void OpDispatchBuilder::AVX128_VPERMILReg(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPERMILReg(OpcodeArgs, IR::OpSize ElementSize) {
   AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), ElementSize, [this](IR::OpSize _ElementSize, Ref Src, Ref Indices) {
-    return VPERMILRegOpImpl(OpSize::i128Bit, ElementSize, Src, Indices);
+    return VPERMILRegOpImpl(OpSize::i128Bit, _ElementSize, Src, Indices);
   });
 }
 
@@ -2338,6 +1839,11 @@ void OpDispatchBuilder::AVX128_VPERMD(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::AVX128_VPCLMULQDQ(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsPMULL_128Bit) {
+    UnimplementedOp(Op);
+    return;
+  }
+
   const auto Selector = static_cast<uint8_t>(Op->Src[2].Literal());
 
   AVX128_VectorBinaryImpl(Op, OpSizeFromSrc(Op), OpSize::iInvalid, [this, Selector](IR::OpSize, Ref Src1, Ref Src2) {
@@ -2518,7 +2024,7 @@ OpDispatchBuilder::RefPair OpDispatchBuilder::AVX128_VPGatherImpl(OpSize Size, O
   ///< BaseAddr doesn't need to exist, calculate that here.
   Ref BaseAddr = VSIB.BaseAddr;
   if (BaseAddr && VSIB.Displacement) {
-    BaseAddr = _Add(OpSize::i64Bit, BaseAddr, Constant(VSIB.Displacement));
+    BaseAddr = Add(OpSize::i64Bit, BaseAddr, VSIB.Displacement);
   } else if (VSIB.Displacement) {
     BaseAddr = Constant(VSIB.Displacement);
   } else if (!BaseAddr) {
@@ -2613,7 +2119,7 @@ OpDispatchBuilder::RefPair OpDispatchBuilder::AVX128_VPGatherQPSImpl(Ref Dest, R
   ///< BaseAddr doesn't need to exist, calculate that here.
   Ref BaseAddr = VSIB.BaseAddr;
   if (BaseAddr && VSIB.Displacement) {
-    BaseAddr = _Add(OpSize::i64Bit, BaseAddr, Constant(VSIB.Displacement));
+    BaseAddr = Add(OpSize::i64Bit, BaseAddr, VSIB.Displacement);
   } else if (VSIB.Displacement) {
     BaseAddr = Constant(VSIB.Displacement);
   } else if (!BaseAddr) {
@@ -2648,8 +2154,7 @@ OpDispatchBuilder::RefPair OpDispatchBuilder::AVX128_VPGatherQPSImpl(Ref Dest, R
   return Result;
 }
 
-template<OpSize AddrElementSize>
-void OpDispatchBuilder::AVX128_VPGATHER(OpcodeArgs) {
+void OpDispatchBuilder::AVX128_VPGATHER(OpcodeArgs, OpSize AddrElementSize) {
 
   const auto Size = OpSizeFromDst(Op);
   const auto Is128Bit = Size == OpSize::i128Bit;

FEXCore/Source/Interface/Core/OpcodeDispatcher/BaseTables.h

@@ -57,6 +57,7 @@ constexpr inline DispatchTableEntry OpDispatch_BaseOpTable[] = {
   {0xC2, 2, &OpDispatchBuilder::RETOp},
   {0xC8, 1, &OpDispatchBuilder::EnterOp},
   {0xC9, 1, &OpDispatchBuilder::LEAVEOp},
+  {0xCA, 2, &OpDispatchBuilder::RETFARIndirectOp},
   {0xCC, 2, &OpDispatchBuilder::INTOp},
   {0xCF, 1, &OpDispatchBuilder::IRETOp},
   {0xD7, 2, &OpDispatchBuilder::XLATOp},
@@ -75,33 +76,4 @@ constexpr inline DispatchTableEntry OpDispatch_BaseOpTable[] = {
   {0xFA, 2, &OpDispatchBuilder::PermissionRestrictedOp},
   {0xFC, 2, &OpDispatchBuilder::FLAGControlOp},
 };
-
-constexpr inline DispatchTableEntry OpDispatch_BaseOpTable_64[] = {
-  {0x63, 1, &OpDispatchBuilder::MOVSXDOp},
-  {0xA0, 4, &OpDispatchBuilder::MOVOffsetOp},
-};
-
-constexpr inline DispatchTableEntry OpDispatch_BaseOpTable_32[] = {
-  {0x06, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_ES_PREFIX>},
-  {0x07, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_ES_PREFIX>},
-  {0x0E, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_CS_PREFIX>},
-  {0x16, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_SS_PREFIX>},
-  {0x17, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_SS_PREFIX>},
-  {0x1E, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_DS_PREFIX>},
-  {0x1F, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_DS_PREFIX>},
-  {0x27, 1, &OpDispatchBuilder::DAAOp},
-  {0x2F, 1, &OpDispatchBuilder::DASOp},
-  {0x37, 1, &OpDispatchBuilder::AAAOp},
-  {0x3F, 1, &OpDispatchBuilder::AASOp},
-  {0x40, 8, &OpDispatchBuilder::INCOp},
-  {0x48, 8, &OpDispatchBuilder::DECOp},
-
-  {0x60, 1, &OpDispatchBuilder::PUSHAOp},
-  {0x61, 1, &OpDispatchBuilder::POPAOp},
-  {0xA0, 4, &OpDispatchBuilder::MOVOffsetOp},
-  {0xCE, 1, &OpDispatchBuilder::INTOp},
-  {0xD4, 1, &OpDispatchBuilder::AAMOp},
-  {0xD5, 1, &OpDispatchBuilder::AADOp},
-  {0xD6, 1, &OpDispatchBuilder::SALCOp},
-};
 } // namespace FEXCore::IR

FEXCore/Source/Interface/Core/OpcodeDispatcher/Crypto.cpp

@@ -19,6 +19,10 @@ class OrderedNode;
 #define OpcodeArgs [[maybe_unused]] FEXCore::X86Tables::DecodedOp Op
 
 void OpDispatchBuilder::SHA1NEXTEOp(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsSHA) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
 
@@ -36,6 +40,10 @@ void OpDispatchBuilder::SHA1NEXTEOp(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::SHA1MSG1Op(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsSHA) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
 
@@ -48,6 +56,10 @@ void OpDispatchBuilder::SHA1MSG1Op(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::SHA1MSG2Op(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsSHA) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
 
@@ -62,6 +74,10 @@ void OpDispatchBuilder::SHA1MSG2Op(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::SHA1RNDS4Op(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsSHA) {
+    UnimplementedOp(Op);
+    return;
+  }
   const uint64_t Imm8 = Op->Src[1].Literal() & 0b11;
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
@@ -100,6 +116,10 @@ void OpDispatchBuilder::SHA1RNDS4Op(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::SHA256MSG1Op(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsSHA) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
 
@@ -109,6 +129,10 @@ void OpDispatchBuilder::SHA256MSG1Op(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::SHA256MSG2Op(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsSHA) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
 
@@ -121,17 +145,11 @@ void OpDispatchBuilder::SHA256MSG2Op(OpcodeArgs) {
   StoreResult(FPRClass, Op, Result, OpSize::iInvalid);
 }
 
-Ref OpDispatchBuilder::BitwiseAtLeastTwo(Ref A, Ref B, Ref C) {
-  // Returns whether at least 2/3 of A/B/C is true.
-  // Expressed as (A & (B | C)) | (B & C)
-  //
-  // Equivalent to expression in SHA calculations: (A & B) ^ (A & C) ^ (B & C)
-  auto And = _And(OpSize::i32Bit, B, C);
-  auto Or = _Or(OpSize::i32Bit, B, C);
-  return _Or(OpSize::i32Bit, _And(OpSize::i32Bit, A, Or), And);
-}
-
 void OpDispatchBuilder::SHA256RNDS2Op(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsSHA) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
   // Hardcoded to XMM0
@@ -163,12 +181,20 @@ void OpDispatchBuilder::SHA256RNDS2Op(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::AESImcOp(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsAES) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
   Ref Result = _VAESImc(Src);
   StoreResult(FPRClass, Op, Result, OpSize::iInvalid);
 }
 
 void OpDispatchBuilder::AESEncOp(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsAES) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
   Ref Result = _VAESEnc(OpSize::i128Bit, Dest, Src, LoadZeroVector(OpSize::i128Bit));
@@ -177,7 +203,7 @@ void OpDispatchBuilder::AESEncOp(OpcodeArgs) {
 
 void OpDispatchBuilder::VAESEncOp(OpcodeArgs) {
   const auto DstSize = OpSizeFromDst(Op);
-  [[maybe_unused]] const auto Is128Bit = DstSize == OpSize::i128Bit;
+  const auto Is128Bit = DstSize == OpSize::i128Bit;
 
   // TODO: Handle 256-bit VAESENC.
   LOGMAN_THROW_A_FMT(Is128Bit, "256-bit VAESENC unimplemented");
@@ -190,6 +216,10 @@ void OpDispatchBuilder::VAESEncOp(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::AESEncLastOp(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsAES) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
   Ref Result = _VAESEncLast(OpSize::i128Bit, Dest, Src, LoadZeroVector(OpSize::i128Bit));
@@ -198,7 +228,7 @@ void OpDispatchBuilder::AESEncLastOp(OpcodeArgs) {
 
 void OpDispatchBuilder::VAESEncLastOp(OpcodeArgs) {
   const auto DstSize = OpSizeFromDst(Op);
-  [[maybe_unused]] const auto Is128Bit = DstSize == OpSize::i128Bit;
+  const auto Is128Bit = DstSize == OpSize::i128Bit;
 
   // TODO: Handle 256-bit VAESENCLAST.
   LOGMAN_THROW_A_FMT(Is128Bit, "256-bit VAESENCLAST unimplemented");
@@ -211,6 +241,10 @@ void OpDispatchBuilder::VAESEncLastOp(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::AESDecOp(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsAES) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
   Ref Result = _VAESDec(OpSize::i128Bit, Dest, Src, LoadZeroVector(OpSize::i128Bit));
@@ -219,7 +253,7 @@ void OpDispatchBuilder::AESDecOp(OpcodeArgs) {
 
 void OpDispatchBuilder::VAESDecOp(OpcodeArgs) {
   const auto DstSize = OpSizeFromDst(Op);
-  [[maybe_unused]] const auto Is128Bit = DstSize == OpSize::i128Bit;
+  const auto Is128Bit = DstSize == OpSize::i128Bit;
 
   // TODO: Handle 256-bit VAESDEC.
   LOGMAN_THROW_A_FMT(Is128Bit, "256-bit VAESDEC unimplemented");
@@ -232,6 +266,10 @@ void OpDispatchBuilder::VAESDecOp(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::AESDecLastOp(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsAES) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
   Ref Result = _VAESDecLast(OpSize::i128Bit, Dest, Src, LoadZeroVector(OpSize::i128Bit));
@@ -240,7 +278,7 @@ void OpDispatchBuilder::AESDecLastOp(OpcodeArgs) {
 
 void OpDispatchBuilder::VAESDecLastOp(OpcodeArgs) {
   const auto DstSize = OpSizeFromDst(Op);
-  [[maybe_unused]] const auto Is128Bit = DstSize == OpSize::i128Bit;
+  const auto Is128Bit = DstSize == OpSize::i128Bit;
 
   // TODO: Handle 256-bit VAESDECLAST.
   LOGMAN_THROW_A_FMT(Is128Bit, "256-bit VAESDECLAST unimplemented");
@@ -261,11 +299,20 @@ Ref OpDispatchBuilder::AESKeyGenAssistImpl(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::AESKeyGenAssist(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsAES) {
+    UnimplementedOp(Op);
+    return;
+  }
+
   Ref Result = AESKeyGenAssistImpl(Op);
   StoreResult(FPRClass, Op, Result, OpSize::iInvalid);
 }
 
 void OpDispatchBuilder::PCLMULQDQOp(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsPMULL_128Bit) {
+    UnimplementedOp(Op);
+    return;
+  }
   Ref Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags);
   Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);
   const auto Selector = static_cast<uint8_t>(Op->Src[1].Literal());
@@ -275,6 +322,10 @@ void OpDispatchBuilder::PCLMULQDQOp(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::VPCLMULQDQOp(OpcodeArgs) {
+  if (!CTX->HostFeatures.SupportsPMULL_128Bit) {
+    UnimplementedOp(Op);
+    return;
+  }
   const auto DstSize = OpSizeFromDst(Op);
 
   Ref Src1 = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags);

FEXCore/Source/Interface/Core/OpcodeDispatcher/Flags.cpp

@@ -201,7 +201,7 @@ void OpDispatchBuilder::FixupAF() {
   auto PFRaw = GetRFLAG(FEXCore::X86State::RFLAG_PF_RAW_LOC);
   auto AFRaw = GetRFLAG(FEXCore::X86State::RFLAG_AF_RAW_LOC);
 
-  // Again 64-bit as masking is more expensive given our ConstProp design.
+  // Again 64-bit as masking is more expensive.
   Ref XorRes = _Xor(OpSize::i64Bit, AFRaw, PFRaw);
   SetRFLAG<FEXCore::X86State::RFLAG_AF_RAW_LOC>(XorRes);
 }
@@ -267,8 +267,6 @@ Ref OpDispatchBuilder::IncrementByCarry(OpSize OpSize, Ref Src) {
 }
 
 Ref OpDispatchBuilder::CalculateFlags_ADC(IR::OpSize SrcSize, Ref Src1, Ref Src2) {
-  auto Zero = _InlineConstant(0);
-  auto One = _InlineConstant(1);
   auto OpSize = SrcSize == OpSize::i64Bit ? OpSize::i64Bit : OpSize::i32Bit;
   Ref Res;
 
@@ -288,11 +286,11 @@ Ref OpDispatchBuilder::CalculateFlags_ADC(IR::OpSize SrcSize, Ref Src1, Ref Src2
     Ref Src2PlusCF = IncrementByCarry(OpSize, Src2);
 
     // Need to zero-extend for the comparison.
-    Res = _Add(OpSize, Src1, Src2PlusCF);
+    Res = Add(OpSize, Src1, Src2PlusCF);
     Res = _Bfe(OpSize, IR::OpSizeAsBits(SrcSize), 0, Res);
 
     // TODO: We can fold that second Bfe in (cmp uxth).
-    auto SelectCFInv = _Select(FEXCore::IR::COND_UGE, Res, Src2PlusCF, One, Zero);
+    auto SelectCFInv = Select01(OpSize, CondClassType {COND_UGE}, Res, Src2PlusCF);
 
     SetNZ_ZeroCV(SrcSize, Res);
     SetCFInverted(SelectCFInv);
@@ -304,8 +302,6 @@ Ref OpDispatchBuilder::CalculateFlags_ADC(IR::OpSize SrcSize, Ref Src1, Ref Src2
 }
 
 Ref OpDispatchBuilder::CalculateFlags_SBB(IR::OpSize SrcSize, Ref Src1, Ref Src2) {
-  auto Zero = _InlineConstant(0);
-  auto One = _InlineConstant(1);
   auto OpSize = SrcSize == OpSize::i64Bit ? OpSize::i64Bit : OpSize::i32Bit;
 
   CalculateAF(Src1, Src2);
@@ -325,10 +321,10 @@ Ref OpDispatchBuilder::CalculateFlags_SBB(IR::OpSize SrcSize, Ref Src1, Ref Src2
 
     auto Src2PlusCF = IncrementByCarry(OpSize, Src2);
 
-    Res = _Sub(OpSize, Src1, Src2PlusCF);
+    Res = Sub(OpSize, Src1, Src2PlusCF);
     Res = _Bfe(OpSize, IR::OpSizeAsBits(SrcSize), 0, Res);
 
-    auto SelectCFInv = _Select(FEXCore::IR::COND_UGE, Src1, Src2PlusCF, One, Zero);
+    auto SelectCFInv = Select01(OpSize, CondClassType {COND_UGE}, Src1, Src2PlusCF);
 
     SetNZ_ZeroCV(SrcSize, Res);
     SetCFInverted(SelectCFInv);
@@ -349,10 +345,10 @@ Ref OpDispatchBuilder::CalculateFlags_SUB(IR::OpSize SrcSize, Ref Src1, Ref Src2
 
   Ref Res;
   if (SrcSize >= OpSize::i32Bit) {
-    Res = _SubWithFlags(SrcSize, Src1, Src2);
+    Res = SubWithFlags(SrcSize, Src1, Src2);
   } else {
     _SubNZCV(SrcSize, Src1, Src2);
-    Res = _Sub(OpSize::i32Bit, Src1, Src2);
+    Res = Sub(OpSize::i32Bit, Src1, Src2);
   }
 
   CalculatePF(Res);
@@ -379,10 +375,10 @@ Ref OpDispatchBuilder::CalculateFlags_ADD(IR::OpSize SrcSize, Ref Src1, Ref Src2
 
   Ref Res;
   if (SrcSize >= OpSize::i32Bit) {
-    Res = _AddWithFlags(SrcSize, Src1, Src2);
+    Res = AddWithFlags(SrcSize, Src1, Src2);
   } else {
     _AddNZCV(SrcSize, Src1, Src2);
-    Res = _Add(OpSize::i32Bit, Src1, Src2);
+    Res = Add(OpSize::i32Bit, Src1, Src2);
   }
 
   CalculatePF(Res);

FEXCore/Source/Interface/Core/OpcodeDispatcher/H0F38Tables.h

@@ -6,6 +6,7 @@ namespace FEXCore::IR {
 #define OPD(prefix, opcode) (((prefix) << 8) | opcode)
 constexpr uint16_t PF_38_NONE = 0;
 constexpr uint16_t PF_38_66 = (1U << 0);
+constexpr uint16_t PF_38_F2 = (1U << 1);
 constexpr uint16_t PF_38_F3 = (1U << 2);
 
 constexpr DispatchTableEntry OpDispatch_H0F38Table[] = {
@@ -71,9 +72,28 @@ constexpr DispatchTableEntry OpDispatch_H0F38Table[] = {
   {OPD(PF_38_66, 0x40), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VectorALUOp, IR::OP_VMUL, OpSize::i32Bit>},
   {OPD(PF_38_66, 0x41), 1, &OpDispatchBuilder::PHMINPOSUWOp},
 
+  {OPD(PF_38_NONE, 0xC8), 1, &OpDispatchBuilder::SHA1NEXTEOp},
+  {OPD(PF_38_NONE, 0xC9), 1, &OpDispatchBuilder::SHA1MSG1Op},
+  {OPD(PF_38_NONE, 0xCA), 1, &OpDispatchBuilder::SHA1MSG2Op},
+  {OPD(PF_38_NONE, 0xCB), 1, &OpDispatchBuilder::SHA256RNDS2Op},
+  {OPD(PF_38_NONE, 0xCC), 1, &OpDispatchBuilder::SHA256MSG1Op},
+  {OPD(PF_38_NONE, 0xCD), 1, &OpDispatchBuilder::SHA256MSG2Op},
+
+  {OPD(PF_38_66, 0xDB), 1, &OpDispatchBuilder::AESImcOp},
+  {OPD(PF_38_66, 0xDC), 1, &OpDispatchBuilder::AESEncOp},
+  {OPD(PF_38_66, 0xDD), 1, &OpDispatchBuilder::AESEncLastOp},
+  {OPD(PF_38_66, 0xDE), 1, &OpDispatchBuilder::AESDecOp},
+  {OPD(PF_38_66, 0xDF), 1, &OpDispatchBuilder::AESDecLastOp},
+
   {OPD(PF_38_NONE, 0xF0), 2, &OpDispatchBuilder::MOVBEOp},
   {OPD(PF_38_66, 0xF0), 2, &OpDispatchBuilder::MOVBEOp},
 
+  {OPD(PF_38_F2, 0xF0), 1, &OpDispatchBuilder::CRC32},
+  {OPD(PF_38_F2, 0xF1), 1, &OpDispatchBuilder::CRC32},
+
+  {OPD(PF_38_66 | PF_38_F2, 0xF0), 1, &OpDispatchBuilder::CRC32},
+  {OPD(PF_38_66 | PF_38_F2, 0xF1), 1, &OpDispatchBuilder::CRC32},
+
   {OPD(PF_38_66, 0xF6), 1, &OpDispatchBuilder::ADXOp},
   {OPD(PF_38_F3, 0xF6), 1, &OpDispatchBuilder::ADXOp},
 };

FEXCore/Source/Interface/Core/OpcodeDispatcher/H0F3ATables.h

@@ -29,6 +29,7 @@ constexpr auto OpDispatchTableGenH0F3A = []() consteval {
       {OPD(REX, PF_3A_66, 0x40), 1, &OpDispatchBuilder::DPPOp<OpSize::i32Bit>},
       {OPD(REX, PF_3A_66, 0x41), 1, &OpDispatchBuilder::DPPOp<OpSize::i64Bit>},
       {OPD(REX, PF_3A_66, 0x42), 1, &OpDispatchBuilder::MPSADBWOp},
+      {OPD(REX, PF_3A_66, 0x44), 1, &OpDispatchBuilder::PCLMULQDQOp},
 
       {OPD(REX, PF_3A_66, 0x60), 1, &OpDispatchBuilder::VPCMPESTRMOp},
       {OPD(REX, PF_3A_66, 0x61), 1, &OpDispatchBuilder::VPCMPESTRIOp},
@@ -36,6 +37,8 @@ constexpr auto OpDispatchTableGenH0F3A = []() consteval {
       {OPD(REX, PF_3A_66, 0x63), 1, &OpDispatchBuilder::VPCMPISTRIOp},
 
       {OPD(REX, PF_3A_NONE, 0xCC), 1, &OpDispatchBuilder::SHA1RNDS4Op},
+      {OPD(REX, PF_3A_66, 0xDF), 1, &OpDispatchBuilder::AESKeyGenAssist},
+
     };
     return std::to_array(Table);
   };
@@ -65,11 +68,6 @@ constexpr DispatchTableEntry OpDispatch_H0F3ATableNeedsREX0[] = {
   {OPD(0, PF_3A_66, 0x22), 1, &OpDispatchBuilder::PINSROp<OpSize::i32Bit>},
 };
 
-constexpr DispatchTableEntry OpDispatch_H0F3ATable_64[] = {
-  {OPD(1, PF_3A_66, 0x16), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PExtrOp, OpSize::i64Bit>},
-  {OPD(1, PF_3A_66, 0x22), 1, &OpDispatchBuilder::PINSROp<OpSize::i64Bit>},
-};
-
 #undef PF_3A_NONE
 #undef PF_3A_66
 

FEXCore/Source/Interface/Core/OpcodeDispatcher/PrimaryGroupTables.h

@@ -117,7 +117,9 @@ constexpr DispatchTableEntry OpDispatch_PrimaryGroupTables[] = {
   {OPD(FEXCore::X86Tables::TYPE_GROUP_5, OpToIndex(0xFF), 0), 1, &OpDispatchBuilder::INCOp}, // INC
   {OPD(FEXCore::X86Tables::TYPE_GROUP_5, OpToIndex(0xFF), 1), 1, &OpDispatchBuilder::DECOp}, // DEC
   {OPD(FEXCore::X86Tables::TYPE_GROUP_5, OpToIndex(0xFF), 2), 1, &OpDispatchBuilder::CALLAbsoluteOp},
+  {OPD(FEXCore::X86Tables::TYPE_GROUP_5, OpToIndex(0xFF), 3), 1, &OpDispatchBuilder::CALLFARIndirectOp},
   {OPD(FEXCore::X86Tables::TYPE_GROUP_5, OpToIndex(0xFF), 4), 1, &OpDispatchBuilder::JUMPAbsoluteOp},
+  {OPD(FEXCore::X86Tables::TYPE_GROUP_5, OpToIndex(0xFF), 5), 1, &OpDispatchBuilder::JUMPFARIndirectOp},
   {OPD(FEXCore::X86Tables::TYPE_GROUP_5, OpToIndex(0xFF), 6), 1, &OpDispatchBuilder::PUSHOp},
 
   // GROUP 11

FEXCore/Source/Interface/Core/OpcodeDispatcher/SecondaryGroupTables.h

@@ -69,10 +69,16 @@ constexpr DispatchTableEntry OpDispatch_SecondaryGroupTables[] = {
 
   // GROUP 9
   {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_NONE, 1), 1, &OpDispatchBuilder::CMPXCHGPairOp},
-  {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_F3, 1), 1, &OpDispatchBuilder::CMPXCHGPairOp},
+  {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_NONE, 6), 1, &OpDispatchBuilder::RDRANDOp<false>},
+  {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_NONE, 7), 1, &OpDispatchBuilder::RDRANDOp<true>},
+
   {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_66, 1), 1, &OpDispatchBuilder::CMPXCHGPairOp},
+  {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_66, 6), 1, &OpDispatchBuilder::RDRANDOp<false>},
+  {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_66, 7), 1, &OpDispatchBuilder::RDRANDOp<true>},
+
   {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_F2, 1), 1, &OpDispatchBuilder::CMPXCHGPairOp},
 
+  {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_F3, 1), 1, &OpDispatchBuilder::CMPXCHGPairOp},
   {OPD(FEXCore::X86Tables::TYPE_GROUP_9, PF_F3, 7), 1, &OpDispatchBuilder::RDPIDOp},
 
   // GROUP 12
@@ -156,18 +162,6 @@ constexpr DispatchTableEntry OpDispatch_SecondaryGroupTables[] = {
   {OPD(FEXCore::X86Tables::TYPE_GROUP_P, PF_F2, 0), 8, &OpDispatchBuilder::NOPOp},
 };
 
-constexpr DispatchTableEntry OpDispatch_SecondaryGroupTables_64[] = {
-  // GROUP 15
-  {OPD(FEXCore::X86Tables::TYPE_GROUP_15, PF_F3, 0), 1,
-   &OpDispatchBuilder::Bind<&OpDispatchBuilder::ReadSegmentReg, OpDispatchBuilder::Segment::FS>},
-  {OPD(FEXCore::X86Tables::TYPE_GROUP_15, PF_F3, 1), 1,
-   &OpDispatchBuilder::Bind<&OpDispatchBuilder::ReadSegmentReg, OpDispatchBuilder::Segment::GS>},
-  {OPD(FEXCore::X86Tables::TYPE_GROUP_15, PF_F3, 2), 1,
-   &OpDispatchBuilder::Bind<&OpDispatchBuilder::WriteSegmentReg, OpDispatchBuilder::Segment::FS>},
-  {OPD(FEXCore::X86Tables::TYPE_GROUP_15, PF_F3, 3), 1,
-   &OpDispatchBuilder::Bind<&OpDispatchBuilder::WriteSegmentReg, OpDispatchBuilder::Segment::GS>},
-};
-
 #undef OPD
 
 } // namespace FEXCore::IR

FEXCore/Source/Interface/Core/OpcodeDispatcher/SecondaryModRMTables.h

@@ -17,6 +17,7 @@ constexpr DispatchTableEntry OpDispatch_SecondaryModRMTables[] = {
   // REG /7
   {((3 << 3) | 0), 1, &OpDispatchBuilder::PermissionRestrictedOp},
   {((3 << 3) | 1), 1, &OpDispatchBuilder::RDTSCPOp},
+  {((3 << 3) | 4), 1, &OpDispatchBuilder::CLZeroOp},
 };
 
 } // namespace FEXCore::IR

FEXCore/Source/Interface/Core/OpcodeDispatcher/SecondaryTables.h

@@ -313,20 +313,4 @@ constexpr DispatchTableEntry OpDispatch_SecondaryOpSizeModTables[] = {
   {0xFD, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VectorALUOp, IR::OP_VADD, OpSize::i16Bit>},
   {0xFE, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VectorALUOp, IR::OP_VADD, OpSize::i32Bit>},
 };
-
-constexpr DispatchTableEntry OpDispatch_TwoByteOpTable_64[] = {
-  {0x05, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::SyscallOp, true>},
-  {0xA0, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_FS_PREFIX>},
-  {0xA1, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_FS_PREFIX>},
-  {0xA8, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_GS_PREFIX>},
-  {0xA9, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_GS_PREFIX>},
-};
-
-constexpr DispatchTableEntry OpDispatch_TwoByteOpTable_32[] = {
-  {0x05, 1, &OpDispatchBuilder::NOPOp},
-  {0xA0, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_FS_PREFIX>},
-  {0xA1, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_FS_PREFIX>},
-  {0xA8, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_GS_PREFIX>},
-  {0xA9, 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_GS_PREFIX>},
-};
 } // namespace FEXCore::IR

FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp

@@ -477,7 +477,7 @@ Ref OpDispatchBuilder::InsertScalar_CVT_Float_To_FloatImpl(OpcodeArgs, IR::OpSiz
   // We load the full vector width when dealing with a source vector,
   // so that we don't do any unnecessary zero extension to the scalar
   // element that we're going to operate on.
-  const auto SrcSize = OpSizeFromSrc(Op);
+  const auto SrcSize = Src2Op.IsGPR() ? OpSize::i128Bit : SrcElementSize;
 
   Ref Src1 = LoadSource_WithOpSize(FPRClass, Op, Src1Op, DstSize, Op->Flags);
   Ref Src2 = LoadSource_WithOpSize(FPRClass, Op, Src2Op, SrcSize, Op->Flags, {.AllowUpperGarbage = true});
@@ -3950,10 +3950,7 @@ void OpDispatchBuilder::PTestOpImpl(OpSize Size, Ref Dest, Ref Src) {
   Test1 = _VExtractToGPR(Size, OpSize::i16Bit, Test1, 0);
   Test2 = _VExtractToGPR(Size, OpSize::i16Bit, Test2, 0);
 
-  auto ZeroConst = Constant(0);
-  auto OneConst = Constant(1);
-
-  Test2 = _Select(FEXCore::IR::COND_NEQ, Test2, ZeroConst, OneConst, ZeroConst);
+  Test2 = To01(OpSize::i64Bit, Test2);
 
   // Careful, these flags are different between {V,}PTEST and VTESTP{S,D}
   // Set ZF according to Test1. SF will be zeroed since we do a 32-bit test on
@@ -3990,10 +3987,7 @@ void OpDispatchBuilder::VTESTOpImpl(OpSize SrcSize, IR::OpSize ElementSize, Ref
   Ref AndGPR = _VExtractToGPR(SrcSize, OpSize::i16Bit, MaxAnd, 0);
   Ref AndNotGPR = _VExtractToGPR(SrcSize, OpSize::i16Bit, MaxAndNot, 0);
 
-  Ref ZeroConst = Constant(0);
-  Ref OneConst = Constant(1);
-
-  Ref CFInv = _Select(IR::COND_NEQ, AndNotGPR, ZeroConst, OneConst, ZeroConst);
+  Ref CFInv = To01(OpSize::i64Bit, AndNotGPR);
 
   // As in PTest, this sets Z appropriately while zeroing the rest of NZCV.
   SetNZ_ZeroCV(OpSize::i32Bit, AndGPR);
@@ -5007,7 +5001,7 @@ void OpDispatchBuilder::VFMAddSubImpl(OpcodeArgs, bool AddSub, uint8_t Src1Idx,
 }
 
 OpDispatchBuilder::RefVSIB OpDispatchBuilder::LoadVSIB(const X86Tables::DecodedOp& Op, const X86Tables::DecodedOperand& Operand, uint32_t Flags) {
-  [[maybe_unused]] const bool IsVSIB = (Op->Flags & X86Tables::DecodeFlags::FLAG_VSIB_BYTE) != 0;
+  const bool IsVSIB = (Op->Flags & X86Tables::DecodeFlags::FLAG_VSIB_BYTE) != 0;
   LOGMAN_THROW_A_FMT(Operand.IsSIB() && IsVSIB, "Trying to load VSIB for something that isn't the correct type!");
 
   // VSIB is a very special case which has a ton of encoded data.
@@ -5084,7 +5078,7 @@ void OpDispatchBuilder::VPGATHER(OpcodeArgs) {
     ///< BaseAddr doesn't need to exist, calculate that here.
     Ref BaseAddr = VSIB.BaseAddr;
     if (BaseAddr && VSIB.Displacement) {
-      BaseAddr = _Add(OpSize::i64Bit, BaseAddr, Constant(VSIB.Displacement));
+      BaseAddr = Add(OpSize::i64Bit, BaseAddr, VSIB.Displacement);
     } else if (VSIB.Displacement) {
       BaseAddr = Constant(VSIB.Displacement);
     } else if (!BaseAddr) {

FEXCore/Source/Interface/Core/OpcodeDispatcher/X87.cpp

@@ -32,21 +32,25 @@ Ref OpDispatchBuilder::GetX87Top() {
 }
 
 void OpDispatchBuilder::SetX87FTW(Ref FTW) {
-  Ref X87Empty = Constant(static_cast<uint8_t>(FPState::X87Tag::Empty));
-  Ref NewAbridgedFTW {};
+  // For the output, we want a 1-bit for each pair not equal to 11 (Empty).
+  static_assert(static_cast<uint8_t>(FPState::X87Tag::Empty) == 0b11);
 
-  for (int i = 0; i < 8; i++) {
-    Ref RegTag = _Bfe(OpSize::i32Bit, 2, i * 2, FTW);
-    Ref RegValid = _Select(FEXCore::IR::COND_NEQ, RegTag, X87Empty, Constant(1), Constant(0));
+  // Make even bits 1 if the pair is equal to 11, and 0 otherwise.
+  FTW = _AndShift(OpSize::i32Bit, FTW, FTW, ShiftType::LSR, 1);
 
-    if (i) {
-      NewAbridgedFTW = _Orlshl(OpSize::i32Bit, NewAbridgedFTW, RegValid, i);
-    } else {
-      NewAbridgedFTW = RegValid;
-    }
-  }
+  // Invert FTW and clear the odd bits. Even bits are 1 if the pair
+  // is not equal to 11, and odd bits are 0.
+  FTW = _Andn(OpSize::i32Bit, Constant(0x55555555), FTW);
 
-  StoreContext(AbridgedFTWIndex, NewAbridgedFTW);
+  // All that's left is to compact away the odd bits. That is a Morton
+  // deinterleave operation, which has a standard solution. See
+  // https://stackoverflow.com/questions/3137266/how-to-de-interleave-bits-unmortonizing
+  FTW = _And(OpSize::i32Bit, _Orlshr(OpSize::i32Bit, FTW, FTW, 1), Constant(0x33333333));
+  FTW = _And(OpSize::i32Bit, _Orlshr(OpSize::i32Bit, FTW, FTW, 2), Constant(0x0f0f0f0f));
+  FTW = _Orlshr(OpSize::i32Bit, FTW, FTW, 4);
+
+  // ...and that's it. StoreContext implicitly does the final masking.
+  StoreContext(AbridgedFTWIndex, FTW);
 }
 
 void OpDispatchBuilder::SetX87Top(Ref Value) {
@@ -110,10 +114,10 @@ void OpDispatchBuilder::FILD(OpcodeArgs) {
   auto absolute = _Neg(OpSize::i64Bit, Data, CondClassType {COND_MI});
 
   // left justify the absolute integer
-  auto shift = _Sub(OpSize::i64Bit, Constant(63), _FindMSB(IR::OpSize::i64Bit, absolute));
+  auto shift = Sub(OpSize::i64Bit, Constant(63), _FindMSB(IR::OpSize::i64Bit, absolute));
   auto shifted = _Lshl(OpSize::i64Bit, absolute, shift);
 
-  auto adjusted_exponent = _Sub(OpSize::i64Bit, Constant(0x3fff + 63), shift);
+  auto adjusted_exponent = Sub(OpSize::i64Bit, Constant(0x3fff + 63), shift);
   auto zeroed_exponent = _Select(COND_EQ, absolute, zero, zero, adjusted_exponent);
   auto upper = _Or(OpSize::i64Bit, sign, zeroed_exponent);
 
@@ -160,12 +164,12 @@ void OpDispatchBuilder::FIST(OpcodeArgs, bool Truncate) {
     // Check for NaN/Infinity: exponent = 0x7fff
     SaveNZCV();
     _TestNZ(OpSize::i64Bit, Exponent, Constant(0x7fff));
-    Ref IsSpecial = _NZCVSelect(OpSize::i64Bit, {COND_EQ}, Constant(1), Constant(0));
+    Ref IsSpecial = _NZCVSelect01({COND_EQ});
 
     // For overflow detection, check if exponent indicates a value >= 2^15
     // Biased exponent for 2^15 is 0x3fff + 15 = 0x400e
-    _SubWithFlags(OpSize::i64Bit, Exponent, Constant(0x400e));
-    Ref IsOverflow = _NZCVSelect(OpSize::i64Bit, {COND_UGE}, Constant(1), Constant(0));
+    SubWithFlags(OpSize::i64Bit, Exponent, 0x400e);
+    Ref IsOverflow = _NZCVSelect01({COND_UGE});
 
     // Set Invalid Operation flag if overflow or special value
     Ref InvalidFlag = _Or(OpSize::i64Bit, IsSpecial, IsOverflow);
@@ -421,11 +425,13 @@ Ref OpDispatchBuilder::ReconstructX87StateFromFSW_Helper(Ref FSW) {
   auto C1 = _Bfe(OpSize::i32Bit, 1, 9, FSW);
   auto C2 = _Bfe(OpSize::i32Bit, 1, 10, FSW);
   auto C3 = _Bfe(OpSize::i32Bit, 1, 14, FSW);
+  auto IE = _Bfe(OpSize::i32Bit, 1, 0, FSW);
 
   SetRFLAG<FEXCore::X86State::X87FLAG_C0_LOC>(C0);
   SetRFLAG<FEXCore::X86State::X87FLAG_C1_LOC>(C1);
   SetRFLAG<FEXCore::X86State::X87FLAG_C2_LOC>(C2);
   SetRFLAG<FEXCore::X86State::X87FLAG_C3_LOC>(C3);
+  SetRFLAG<FEXCore::X86State::X87FLAG_IE_LOC>(IE);
   return Top;
 }
 
@@ -439,13 +445,13 @@ void OpDispatchBuilder::X87LDENV(OpcodeArgs) {
   auto NewFCW = _LoadMem(GPRClass, OpSize::i16Bit, Mem, OpSize::i16Bit);
   _StoreContext(OpSize::i16Bit, GPRClass, NewFCW, offsetof(FEXCore::Core::CPUState, FCW));
 
-  Ref MemLocation = _Add(OpSize::i64Bit, Mem, Constant(IR::OpSizeToSize(Size) * 1));
+  Ref MemLocation = Add(OpSize::i64Bit, Mem, IR::OpSizeToSize(Size) * 1);
   auto NewFSW = _LoadMem(GPRClass, Size, MemLocation, Size);
   ReconstructX87StateFromFSW_Helper(NewFSW);
 
   {
     // FTW
-    Ref MemLocation = _Add(OpSize::i64Bit, Mem, Constant(IR::OpSizeToSize(Size) * 2));
+    Ref MemLocation = Add(OpSize::i64Bit, Mem, IR::OpSizeToSize(Size) * 2);
     SetX87FTW(_LoadMem(GPRClass, Size, MemLocation, Size));
   }
 }
@@ -508,7 +514,6 @@ void OpDispatchBuilder::X87FNSAVE(OpcodeArgs) {
     _StoreMem(GPRClass, Size, ZeroConst, Mem, Constant(IR::OpSizeToSize(Size) * 6), Size, MEM_OFFSET_SXTX, 1);
   }
 
-  auto OneConst = Constant(1);
   auto SevenConst = Constant(7);
   const auto LoadSize = ReducedPrecisionMode ? OpSize::i64Bit : OpSize::i128Bit;
   for (int i = 0; i < 7; ++i) {
@@ -517,7 +522,7 @@ void OpDispatchBuilder::X87FNSAVE(OpcodeArgs) {
       data = _F80CVTTo(data, OpSize::i64Bit);
     }
     _StoreMem(FPRClass, OpSize::i128Bit, data, Mem, Constant((IR::OpSizeToSize(Size) * 7) + (10 * i)), OpSize::i8Bit, MEM_OFFSET_SXTX, 1);
-    Top = _And(OpSize::i32Bit, _Add(OpSize::i32Bit, Top, OneConst), SevenConst);
+    Top = _And(OpSize::i32Bit, Add(OpSize::i32Bit, Top, 1), SevenConst);
   }
 
   // The final st(7) needs a bit of special handling here
@@ -561,9 +566,7 @@ void OpDispatchBuilder::X87FRSTOR(OpcodeArgs) {
     SetX87FTW(_LoadMem(GPRClass, Size, Mem, Constant(IR::OpSizeToSize(Size) * 2), Size, MEM_OFFSET_SXTX, 1));
   }
 
-  auto OneConst = Constant(1);
   auto SevenConst = Constant(7);
-
   auto low = Constant(~0ULL);
   auto high = Constant(0xFFFF);
   Ref Mask = _VLoadTwoGPRs(low, high);
@@ -578,7 +581,7 @@ void OpDispatchBuilder::X87FRSTOR(OpcodeArgs) {
     }
     _StoreContextIndexed(Reg, Top, StoreSize, MMBaseOffset(), IR::OpSizeToSize(OpSize::i128Bit), FPRClass);
 
-    Top = _And(OpSize::i32Bit, _Add(OpSize::i32Bit, Top, OneConst), SevenConst);
+    Top = _And(OpSize::i32Bit, Add(OpSize::i32Bit, Top, 1), SevenConst);
   }
 
   // The final st(7) needs a bit of special handling here
@@ -677,6 +680,9 @@ void OpDispatchBuilder::FCOMI(OpcodeArgs, IR::OpSize Width, bool Integer, OpDisp
     SetRFLAG<FEXCore::X86State::RFLAG_PF_RAW_LOC>(PF);
   }
 
+  // Set Invalid Operation flag when unordered (NaN comparison)
+  SetRFLAG<FEXCore::X86State::X87FLAG_IE_LOC>(HostFlag_Unordered);
+
   if (PopTwice) {
     _PopStackDestroy();
     _PopStackDestroy();
@@ -698,6 +704,9 @@ void OpDispatchBuilder::FTST(OpcodeArgs) {
   SetRFLAG<FEXCore::X86State::X87FLAG_C1_LOC>(Constant(0));
   SetRFLAG<FEXCore::X86State::X87FLAG_C2_LOC>(HostFlag_Unordered);
   SetRFLAG<FEXCore::X86State::X87FLAG_C3_LOC>(HostFlag_ZF);
+
+  // Set Invalid Operation flag when unordered (NaN comparison)
+  SetRFLAG<FEXCore::X86State::X87FLAG_IE_LOC>(HostFlag_Unordered);
 }
 
 void OpDispatchBuilder::X87OpHelper(OpcodeArgs, FEXCore::IR::IROps IROp, bool ZeroC2) {
@@ -759,6 +768,11 @@ void OpDispatchBuilder::X87FNSTSW(OpcodeArgs) {
   StoreResult(GPRClass, Op, StatusWord, OpSize::iInvalid);
 }
 
+void OpDispatchBuilder::FNCLEX(OpcodeArgs) {
+  // Clear the exception flag bit
+  SetRFLAG<FEXCore::X86State::X87FLAG_IE_LOC>(_Constant(0));
+}
+
 void OpDispatchBuilder::FNINIT(OpcodeArgs) {
   auto Zero = Constant(0);
 
@@ -782,6 +796,7 @@ void OpDispatchBuilder::FNINIT(OpcodeArgs) {
   SetRFLAG<FEXCore::X86State::X87FLAG_C1_LOC>(Zero);
   SetRFLAG<FEXCore::X86State::X87FLAG_C2_LOC>(Zero);
   SetRFLAG<FEXCore::X86State::X87FLAG_C3_LOC>(Zero);
+  SetRFLAG<FEXCore::X86State::X87FLAG_IE_LOC>(Zero);
 }
 
 void OpDispatchBuilder::X87FFREE(OpcodeArgs) {
@@ -827,11 +842,7 @@ void OpDispatchBuilder::X87FCMOV(OpcodeArgs) {
   default: LOGMAN_MSG_A_FMT("Unhandled FCMOV op: 0x{:x}", Opcode); break;
   }
 
-  auto ZeroConst = Constant(0);
-  auto AllOneConst = Constant(0xffff'ffff'ffff'ffffull);
-
-  Ref SrcCond = SelectCC(CC, OpSize::i64Bit, AllOneConst, ZeroConst);
-  Ref VecCond = _VDupFromGPR(OpSize::i128Bit, OpSize::i64Bit, SrcCond);
+  Ref VecCond = _VDupFromGPR(OpSize::i128Bit, OpSize::i64Bit, SelectCC0All1(CC));
   _F80VBSLStack(OpSize::i128Bit, VecCond, Op->OP & 7, 0);
 }
 
@@ -847,11 +858,9 @@ void OpDispatchBuilder::X87FXAM(OpcodeArgs) {
   // Claim this is a normal number
   // We don't support anything else
   auto TopValid = _StackValidTag(0);
-  auto ZeroConst = Constant(0);
-  auto OneConst = Constant(1);
 
   // In the case of top being invalid then C3:C2:C0 is 0b101
-  auto C3 = _Select(FEXCore::IR::COND_NEQ, TopValid, OneConst, OneConst, ZeroConst);
+  auto C3 = Select01(OpSize::i32Bit, CondClassType {COND_NEQ}, TopValid, Constant(1));
 
   auto C2 = TopValid;
   auto C0 = C3; // Mirror C3 until something other than zero is supported

FEXCore/Source/Interface/Core/OpcodeDispatcher/X87F64.cpp

@@ -382,7 +382,7 @@ void OpDispatchBuilder::X87FXTRACTF64(OpcodeArgs) {
 
   // non zero case
   Ref ExpNZ = _Bfe(OpSize::i64Bit, 11, 52, Gpr);
-  ExpNZ = _Sub(OpSize::i64Bit, ExpNZ, Constant(1023));
+  ExpNZ = Sub(OpSize::i64Bit, ExpNZ, Constant(1023));
   Ref ExpNZV = _Float_FromGPR_S(OpSize::i64Bit, OpSize::i64Bit, ExpNZ);
 
   Ref SigNZ = _And(OpSize::i64Bit, Gpr, Constant(0x800f'ffff'ffff'ffffLL));

FEXCore/Source/Interface/Core/X86HelperGen.cpp

@@ -33,7 +33,7 @@ X86GeneratedCode::X86GeneratedCode() {
 
   CallbackReturn = reinterpret_cast<uint64_t>(CodePtr);
 
-  memcpy(reinterpret_cast<void*>(CallbackReturn), &SignalReturnCode.at(0), SignalReturnCode.size());
+  memcpy(reinterpret_cast<void*>(CallbackReturn), SignalReturnCode.data(), SignalReturnCode.size());
 
   mprotect(CodePtr, CODE_SIZE, PROT_READ);
 #endif

FEXCore/Source/Interface/Core/X86Tables.cpp

@@ -1,29 +0,0 @@
-// SPDX-License-Identifier: MIT
-/*
-$info$
-meta: frontend|x86-tables ~ Metadata that drives the frontend x86/64 decoding
-tags: frontend|x86-tables
-$end_info$
-*/
-
-#include "Interface/Core/X86Tables/X86Tables.h"
-
-#include <FEXCore/Core/Context.h>
-
-namespace FEXCore::X86Tables {
-
-void InitializeBaseTables(Context::OperatingMode Mode);
-void InitializeSecondaryTables(Context::OperatingMode Mode);
-void InitializeSecondaryGroupTables(Context::OperatingMode Mode);
-void InitializePrimaryGroupTables(Context::OperatingMode Mode);
-void InitializeH0F3ATables(Context::OperatingMode Mode);
-
-void InitializeInfoTables(Context::OperatingMode Mode) {
-  InitializeBaseTables(Mode);
-  InitializeSecondaryTables(Mode);
-  InitializeSecondaryGroupTables(Mode);
-  InitializePrimaryGroupTables(Mode);
-  InitializeH0F3ATables(Mode);
-}
-
-} // namespace FEXCore::X86Tables

FEXCore/Source/Interface/Core/X86Tables/BaseTables.cpp

@@ -15,300 +15,425 @@ $end_info$
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 
-std::array<X86InstInfo, MAX_PRIMARY_TABLE_SIZE> BaseOps = []() consteval {
+enum Primary_LUT {
+  ENTRY_06,
+  ENTRY_07,
+  ENTRY_0E,
+  ENTRY_16,
+  ENTRY_17,
+  ENTRY_1E,
+  ENTRY_1F,
+  ENTRY_27,
+  ENTRY_2F,
+  ENTRY_37,
+  ENTRY_3F,
+  ENTRY_40,
+  ENTRY_48,
+  ENTRY_60,
+  ENTRY_61,
+  ENTRY_63,
+  ENTRY_9A,
+  ENTRY_A0,
+  ENTRY_A1,
+  ENTRY_A2,
+  ENTRY_A3,
+  ENTRY_CE,
+  ENTRY_D4,
+  ENTRY_D5,
+  ENTRY_D6,
+  ENTRY_EA,
+  ENTRY_MAX,
+};
+
+constexpr std::array<X86InstInfo[2], ENTRY_MAX> Primary_ArchSelect_LUT = {{
+  // ENTRY_06
+  {
+    {"PUSH ES",  TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_ES_PREFIX> } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_07
+  {
+    {"POP ES",   TYPE_INST, GenFlagsSizes(SIZE_16BIT, SIZE_DEF) | FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_ES_PREFIX> } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_0E
+  {
+    {"PUSH CS",  TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_CS_PREFIX> } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_16
+  {
+    {"PUSH SS",  TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_SS_PREFIX> } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_17
+  {
+    {"POP SS",   TYPE_INST, GenFlagsSizes(SIZE_16BIT, SIZE_DEF) | FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_SS_PREFIX> } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_1E
+  {
+    {"PUSH DS",  TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::PUSHSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_DS_PREFIX> } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_1F
+  {
+    {"POP DS",   TYPE_INST, GenFlagsSizes(SIZE_16BIT, SIZE_DEF) | FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::POPSegmentOp, FEXCore::X86Tables::DecodeFlags::FLAG_DS_PREFIX> } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_27
+  {
+    {"DAA",      TYPE_INST, GenFlagsDstSize(SIZE_8BIT) | FLAGS_SF_DST_RAX, 0, { .OpDispatch = &IR::OpDispatchBuilder::DAAOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_2F
+  {
+    {"DAS",      TYPE_INST, GenFlagsDstSize(SIZE_8BIT) | FLAGS_SF_DST_RAX, 0, { .OpDispatch = &IR::OpDispatchBuilder::DASOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_37
+  {
+    {"AAA",      TYPE_INST, GenFlagsDstSize(SIZE_16BIT) | FLAGS_SF_DST_RAX, 0, { .OpDispatch = &IR::OpDispatchBuilder::AAAOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_3F
+  {
+    {"AAS",      TYPE_INST, GenFlagsDstSize(SIZE_16BIT) | FLAGS_SF_DST_RAX, 0, { .OpDispatch = &IR::OpDispatchBuilder::AASOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_40
+  {
+    {"INC",    TYPE_INST, FLAGS_SF_REX_IN_BYTE, 0, { .OpDispatch = &IR::OpDispatchBuilder::INCOp } },
+    // REX
+    {"", TYPE_REX_PREFIX, FLAGS_NONE, 0},
+  },
+  // ENTRY_48
+  {
+    {"DEC",    TYPE_INST, FLAGS_SF_REX_IN_BYTE, 0, { .OpDispatch = &IR::OpDispatchBuilder::DECOp } },
+    {"", TYPE_REX_PREFIX, FLAGS_NONE, 0},
+  },
+  // ENTRY_60
+  {
+    {"PUSHA",  TYPE_INST, FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::PUSHAOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_61
+  {
+    {"POPA",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS, 0, { .OpDispatch = &IR::OpDispatchBuilder::POPAOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_63
+  {
+    {"ARPL",   TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"MOVSXD", TYPE_INST, GenFlagsDstSize(SIZE_64BIT) | FLAGS_MODRM, 0, { .OpDispatch = &IR::OpDispatchBuilder::MOVSXDOp } },
+  },
+  // ENTRY_9A
+  {
+    {"CALLF",  TYPE_INST, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_A0
+  {
+    {"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX | FLAGS_MEM_OFFSET, 4, { .OpDispatch = &IR::OpDispatchBuilder::MOVOffsetOp } },
+    {"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX | FLAGS_MEM_OFFSET, 8, { .OpDispatch = &IR::OpDispatchBuilder::MOVOffsetOp } },
+  },
+  // ENTRY_A1
+  {
+    {"MOV",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_MEM_OFFSET, 4, { .OpDispatch = &IR::OpDispatchBuilder::MOVOffsetOp } },
+    {"MOV",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_MEM_OFFSET, 8, { .OpDispatch = &IR::OpDispatchBuilder::MOVOffsetOp } },
+  },
+  // ENTRY_A2
+  {
+    {"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_SRC_RAX | FLAGS_MEM_OFFSET, 4, { .OpDispatch = &IR::OpDispatchBuilder::MOVOffsetOp } },
+    {"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_SRC_RAX | FLAGS_MEM_OFFSET, 8, { .OpDispatch = &IR::OpDispatchBuilder::MOVOffsetOp } },
+  },
+  // ENTRY_A3
+  {
+    {"MOV",    TYPE_INST, FLAGS_SF_SRC_RAX | FLAGS_MEM_OFFSET, 4, { .OpDispatch = &IR::OpDispatchBuilder::MOVOffsetOp } },
+    {"MOV",    TYPE_INST, FLAGS_SF_SRC_RAX | FLAGS_MEM_OFFSET, 8, { .OpDispatch = &IR::OpDispatchBuilder::MOVOffsetOp } },
+  },
+  // ENTRY_CE
+  {
+    {"INTO",   TYPE_INST, FLAGS_NONE, 0, { .OpDispatch = &IR::OpDispatchBuilder::INTOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_D4
+  {
+    {"AAM",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX, 1, { .OpDispatch = &IR::OpDispatchBuilder::AAMOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_D5
+  {
+    {"AAD",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX, 1, { .OpDispatch = &IR::OpDispatchBuilder::AADOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_D6
+  {
+    {"SALC",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX | FLAGS_SF_SRC_RAX, 0, { .OpDispatch = &IR::OpDispatchBuilder::SALCOp } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  // ENTRY_EA
+  {
+    {"JMPF",   TYPE_INST, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+}};
+
+const std::array<X86InstInfo, MAX_PRIMARY_TABLE_SIZE> BaseOps = []() consteval {
   std::array<X86InstInfo, MAX_PRIMARY_TABLE_SIZE> Table{};
 
   constexpr U8U8InfoStruct BaseOpTable[] = {
     // Prefixes
     // Operand size overide
-    {0x66, 1, X86InstInfo{"",      TYPE_PREFIX, FLAGS_NONE,        0, nullptr}},
+    {0x66, 1, X86InstInfo{"",      TYPE_PREFIX, FLAGS_NONE,        0}},
     // Address size override
-    {0x67, 1, X86InstInfo{"",      TYPE_PREFIX, FLAGS_NONE,        0, nullptr}},
-    {0x26, 1, X86InstInfo{"ES",    TYPE_LEGACY_PREFIX, FLAGS_NONE, 0, nullptr}},
-    {0x2E, 1, X86InstInfo{"CS",    TYPE_LEGACY_PREFIX, FLAGS_NONE, 0, nullptr}},
-    {0x36, 1, X86InstInfo{"SS",    TYPE_LEGACY_PREFIX, FLAGS_NONE, 0, nullptr}},
-    {0x3E, 1, X86InstInfo{"DS",    TYPE_LEGACY_PREFIX, FLAGS_NONE, 0, nullptr}},
+    {0x67, 1, X86InstInfo{"",      TYPE_PREFIX, FLAGS_NONE,        0}},
+    {0x26, 1, X86InstInfo{"ES",    TYPE_LEGACY_PREFIX, FLAGS_NONE, 0}},
+    {0x2E, 1, X86InstInfo{"CS",    TYPE_LEGACY_PREFIX, FLAGS_NONE, 0}},
+    {0x36, 1, X86InstInfo{"SS",    TYPE_LEGACY_PREFIX, FLAGS_NONE, 0}},
+    {0x3E, 1, X86InstInfo{"DS",    TYPE_LEGACY_PREFIX, FLAGS_NONE, 0}},
     // These are still invalid on 64bit
-    {0x64, 1, X86InstInfo{"FS",    TYPE_PREFIX, FLAGS_NONE,        0, nullptr}},
-    {0x65, 1, X86InstInfo{"GS",    TYPE_PREFIX, FLAGS_NONE,        0, nullptr}},
-    {0xF0, 1, X86InstInfo{"LOCK",  TYPE_PREFIX, FLAGS_NONE,        0, nullptr}},
-    {0xF2, 1, X86InstInfo{"REPNE", TYPE_PREFIX, FLAGS_NONE,        0, nullptr}},
-    {0xF3, 1, X86InstInfo{"REP",   TYPE_PREFIX, FLAGS_NONE,        0, nullptr}},
+    {0x64, 1, X86InstInfo{"FS",    TYPE_PREFIX, FLAGS_NONE,        0}},
+    {0x65, 1, X86InstInfo{"GS",    TYPE_PREFIX, FLAGS_NONE,        0}},
+    {0xF0, 1, X86InstInfo{"LOCK",  TYPE_PREFIX, FLAGS_NONE,        0}},
+    {0xF2, 1, X86InstInfo{"REPNE", TYPE_PREFIX, FLAGS_NONE,        0}},
+    {0xF3, 1, X86InstInfo{"REP",   TYPE_PREFIX, FLAGS_NONE,        0}},
 
     // Instructions
-    {0x00, 1, X86InstInfo{"ADD",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0, nullptr}},
-    {0x01, 1, X86InstInfo{"ADD",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_DISPLACE_SIZE_DIV_2,                                       0, nullptr}},
-    {0x02, 1, X86InstInfo{"ADD",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0, nullptr}},
-    {0x03, 1, X86InstInfo{"ADD",    TYPE_INST, FLAGS_MODRM,                                                                   0, nullptr}},
-    {0x04, 1, X86InstInfo{"ADD",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1, nullptr}},
-    {0x05, 1, X86InstInfo{"ADD",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4, nullptr}},
+    {0x00, 1, X86InstInfo{"ADD",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0}},
+    {0x01, 1, X86InstInfo{"ADD",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_DISPLACE_SIZE_DIV_2,                                       0}},
+    {0x02, 1, X86InstInfo{"ADD",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0}},
+    {0x03, 1, X86InstInfo{"ADD",    TYPE_INST, FLAGS_MODRM,                                                                   0}},
+    {0x04, 1, X86InstInfo{"ADD",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1}},
+    {0x05, 1, X86InstInfo{"ADD",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4}},
 
-    {0x08, 1, X86InstInfo{"OR",     TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0, nullptr}},
-    {0x09, 1, X86InstInfo{"OR",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0, nullptr}},
-    {0x0A, 1, X86InstInfo{"OR",     TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0, nullptr}},
-    {0x0B, 1, X86InstInfo{"OR",     TYPE_INST, FLAGS_MODRM,                                                                   0, nullptr}},
-    {0x0C, 1, X86InstInfo{"OR",     TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX ,                              1, nullptr}},
-    {0x0D, 1, X86InstInfo{"OR",     TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4, nullptr}},
+    {0x06, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_06] }}},
+    {0x07, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_07] }}},
 
-    {0x10, 1, X86InstInfo{"ADC",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0, nullptr}},
-    {0x11, 1, X86InstInfo{"ADC",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_DISPLACE_SIZE_DIV_2,                                       0, nullptr}},
-    {0x12, 1, X86InstInfo{"ADC",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0, nullptr}},
-    {0x13, 1, X86InstInfo{"ADC",    TYPE_INST, FLAGS_MODRM,                                                                   0, nullptr}},
-    {0x14, 1, X86InstInfo{"ADC",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1, nullptr}},
-    {0x15, 1, X86InstInfo{"ADC",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4, nullptr}},
+    {0x08, 1, X86InstInfo{"OR",     TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0}},
+    {0x09, 1, X86InstInfo{"OR",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0}},
+    {0x0A, 1, X86InstInfo{"OR",     TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0}},
+    {0x0B, 1, X86InstInfo{"OR",     TYPE_INST, FLAGS_MODRM,                                                                   0}},
+    {0x0C, 1, X86InstInfo{"OR",     TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX ,                              1}},
+    {0x0D, 1, X86InstInfo{"OR",     TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4}},
+    {0x0E, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_0E] }}},
 
-    {0x18, 1, X86InstInfo{"SBB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0, nullptr}},
-    {0x19, 1, X86InstInfo{"SBB",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_DISPLACE_SIZE_DIV_2,                                       0, nullptr}},
-    {0x1A, 1, X86InstInfo{"SBB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0, nullptr}},
-    {0x1B, 1, X86InstInfo{"SBB",    TYPE_INST, FLAGS_MODRM,                                                                   0, nullptr}},
-    {0x1C, 1, X86InstInfo{"SBB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1, nullptr}},
-    {0x1D, 1, X86InstInfo{"SBB",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4, nullptr}},
+    {0x10, 1, X86InstInfo{"ADC",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0}},
+    {0x11, 1, X86InstInfo{"ADC",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_DISPLACE_SIZE_DIV_2,                                       0}},
+    {0x12, 1, X86InstInfo{"ADC",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0}},
+    {0x13, 1, X86InstInfo{"ADC",    TYPE_INST, FLAGS_MODRM,                                                                   0}},
+    {0x14, 1, X86InstInfo{"ADC",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1}},
+    {0x15, 1, X86InstInfo{"ADC",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4}},
+    {0x16, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_16] }}},
+    {0x17, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_17] }}},
 
-    {0x20, 1, X86InstInfo{"AND",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0, nullptr}},
-    {0x21, 1, X86InstInfo{"AND",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0, nullptr}},
-    {0x22, 1, X86InstInfo{"AND",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0, nullptr}},
-    {0x23, 1, X86InstInfo{"AND",    TYPE_INST, FLAGS_MODRM,                                                                   0, nullptr}},
-    {0x24, 1, X86InstInfo{"AND",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1, nullptr}},
-    {0x25, 1, X86InstInfo{"AND",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4, nullptr}},
+    {0x18, 1, X86InstInfo{"SBB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0}},
+    {0x19, 1, X86InstInfo{"SBB",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_DISPLACE_SIZE_DIV_2,                                       0}},
+    {0x1A, 1, X86InstInfo{"SBB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0}},
+    {0x1B, 1, X86InstInfo{"SBB",    TYPE_INST, FLAGS_MODRM,                                                                   0}},
+    {0x1C, 1, X86InstInfo{"SBB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1}},
+    {0x1D, 1, X86InstInfo{"SBB",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4}},
+    {0x1E, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_1E] }}},
+    {0x1F, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_1F] }}},
 
-    {0x28, 1, X86InstInfo{"SUB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0, nullptr}},
-    {0x29, 1, X86InstInfo{"SUB",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0, nullptr}},
-    {0x2A, 1, X86InstInfo{"SUB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0, nullptr}},
-    {0x2B, 1, X86InstInfo{"SUB",    TYPE_INST, FLAGS_MODRM,                                                                   0, nullptr}},
-    {0x2C, 1, X86InstInfo{"SUB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX ,                              1, nullptr}},
-    {0x2D, 1, X86InstInfo{"SUB",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4, nullptr}},
+    {0x20, 1, X86InstInfo{"AND",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0}},
+    {0x21, 1, X86InstInfo{"AND",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0}},
+    {0x22, 1, X86InstInfo{"AND",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0}},
+    {0x23, 1, X86InstInfo{"AND",    TYPE_INST, FLAGS_MODRM,                                                                   0}},
+    {0x24, 1, X86InstInfo{"AND",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1}},
+    {0x25, 1, X86InstInfo{"AND",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4}},
 
-    {0x30, 1, X86InstInfo{"XOR",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0, nullptr}},
-    {0x31, 1, X86InstInfo{"XOR",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0, nullptr}},
-    {0x32, 1, X86InstInfo{"XOR",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0, nullptr}},
-    {0x33, 1, X86InstInfo{"XOR",    TYPE_INST, FLAGS_MODRM,                                                                   0, nullptr}},
-    {0x34, 1, X86InstInfo{"XOR",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1, nullptr}},
-    {0x35, 1, X86InstInfo{"XOR",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4, nullptr}},
+    {0x27, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_27] }}},
+    {0x28, 1, X86InstInfo{"SUB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0}},
+    {0x29, 1, X86InstInfo{"SUB",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0}},
+    {0x2A, 1, X86InstInfo{"SUB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0}},
+    {0x2B, 1, X86InstInfo{"SUB",    TYPE_INST, FLAGS_MODRM,                                                                   0}},
+    {0x2C, 1, X86InstInfo{"SUB",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX ,                              1}},
+    {0x2D, 1, X86InstInfo{"SUB",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4}},
+    {0x2F, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_2F] }}},
 
-    {0x38, 1, X86InstInfo{"CMP",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0, nullptr}},
-    {0x39, 1, X86InstInfo{"CMP",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0, nullptr}},
-    {0x3A, 1, X86InstInfo{"CMP",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0, nullptr}},
-    {0x3B, 1, X86InstInfo{"CMP",    TYPE_INST, FLAGS_MODRM,                                                                   0, nullptr}},
-    {0x3C, 1, X86InstInfo{"CMP",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1, nullptr}},
-    {0x3D, 1, X86InstInfo{"CMP",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4, nullptr}},
+    {0x30, 1, X86InstInfo{"XOR",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0}},
+    {0x31, 1, X86InstInfo{"XOR",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0}},
+    {0x32, 1, X86InstInfo{"XOR",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0}},
+    {0x33, 1, X86InstInfo{"XOR",    TYPE_INST, FLAGS_MODRM,                                                                   0}},
+    {0x34, 1, X86InstInfo{"XOR",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1}},
+    {0x35, 1, X86InstInfo{"XOR",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4}},
 
-    {0x50, 8, X86InstInfo{"PUSH",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SF_REX_IN_BYTE | FLAGS_DEBUG_MEM_ACCESS ,                    0, nullptr}},
-    {0x58, 8, X86InstInfo{"POP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SF_REX_IN_BYTE | FLAGS_DEBUG_MEM_ACCESS ,                    0, nullptr}},
+    {0x37, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_37] }}},
+    {0x38, 1, X86InstInfo{"CMP",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                   0}},
+    {0x39, 1, X86InstInfo{"CMP",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                                   0}},
+    {0x3A, 1, X86InstInfo{"CMP",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,                                                   0}},
+    {0x3B, 1, X86InstInfo{"CMP",    TYPE_INST, FLAGS_MODRM,                                                                   0}},
+    {0x3C, 1, X86InstInfo{"CMP",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX  ,                              1}},
+    {0x3D, 1, X86InstInfo{"CMP",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2, 4}},
+    {0x3F, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_3F] }}},
 
-    {0x62, 1, X86InstInfo{"",       TYPE_GROUP_EVEX, FLAGS_NONE,                                                                           0, nullptr}},
+    {0x40, 8, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_40] }}},
+    {0x48, 8, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_48] }}},
 
-    {0x68, 1, X86InstInfo{"PUSH",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_DISPLACE_SIZE_DIV_2 | FLAGS_SRC_SEXT, 4, nullptr}},
-    {0x69, 1, X86InstInfo{"IMUL",   TYPE_INST, FLAGS_MODRM | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2,        4, nullptr}},
-    {0x6A, 1, X86InstInfo{"PUSH",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_SRC_SEXT ,            1, nullptr}},
-    {0x6B, 1, X86InstInfo{"IMUL",   TYPE_INST, FLAGS_MODRM | FLAGS_SRC_SEXT ,                                    1, nullptr}},
+    {0x50, 8, X86InstInfo{"PUSH",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SF_REX_IN_BYTE | FLAGS_DEBUG_MEM_ACCESS ,                    0}},
+    {0x58, 8, X86InstInfo{"POP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SF_REX_IN_BYTE | FLAGS_DEBUG_MEM_ACCESS ,                    0}},
+
+
+    {0x60, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_60] }}},
+    {0x61, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_61] }}},
+    {0x62, 1, X86InstInfo{"",       TYPE_GROUP_EVEX, FLAGS_NONE,                                                                           0}},
+    {0x63, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_63] }}},
+
+    {0x68, 1, X86InstInfo{"PUSH",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_DISPLACE_SIZE_DIV_2 | FLAGS_SRC_SEXT, 4}},
+    {0x69, 1, X86InstInfo{"IMUL",   TYPE_INST, FLAGS_MODRM | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2,        4}},
+    {0x6A, 1, X86InstInfo{"PUSH",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_SRC_SEXT ,            1}},
+    {0x6B, 1, X86InstInfo{"IMUL",   TYPE_INST, FLAGS_MODRM | FLAGS_SRC_SEXT ,                                    1}},
 
     // This should just throw a GP
-    {0x6C, 1, X86InstInfo{"INSB",   TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {0x6D, 1, X86InstInfo{"INSW",   TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {0x6E, 1, X86InstInfo{"OUTS",   TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {0x6F, 1, X86InstInfo{"OUTS",   TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
+    {0x6C, 1, X86InstInfo{"INSB",   TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {0x6D, 1, X86InstInfo{"INSW",   TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {0x6E, 1, X86InstInfo{"OUTS",   TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {0x6F, 1, X86InstInfo{"OUTS",   TYPE_INST, FLAGS_BLOCK_END, 0}},
 
-    {0x70, 1, X86InstInfo{"JO",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x71, 1, X86InstInfo{"JNO",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x72, 1, X86InstInfo{"JB",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x73, 1, X86InstInfo{"JNB",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x74, 1, X86InstInfo{"JZ",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x75, 1, X86InstInfo{"JNZ",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x76, 1, X86InstInfo{"JBE",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x77, 1, X86InstInfo{"JNBE",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x78, 1, X86InstInfo{"JS",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x79, 1, X86InstInfo{"JNS",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x7A, 1, X86InstInfo{"JP",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x7B, 1, X86InstInfo{"JNP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x7C, 1, X86InstInfo{"JL",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x7D, 1, X86InstInfo{"JNL",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x7E, 1, X86InstInfo{"JLE",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
-    {0x7F, 1, X86InstInfo{"JNLE",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1, nullptr}},
+    {0x70, 1, X86InstInfo{"JO",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x71, 1, X86InstInfo{"JNO",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x72, 1, X86InstInfo{"JB",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x73, 1, X86InstInfo{"JNB",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x74, 1, X86InstInfo{"JZ",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x75, 1, X86InstInfo{"JNZ",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x76, 1, X86InstInfo{"JBE",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x77, 1, X86InstInfo{"JNBE",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x78, 1, X86InstInfo{"JS",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x79, 1, X86InstInfo{"JNS",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x7A, 1, X86InstInfo{"JP",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x7B, 1, X86InstInfo{"JNP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x7C, 1, X86InstInfo{"JL",     TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x7D, 1, X86InstInfo{"JNL",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x7E, 1, X86InstInfo{"JLE",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
+    {0x7F, 1, X86InstInfo{"JNLE",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT , 1}},
 
-    {0x84, 1, X86InstInfo{"TEST",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,         0, nullptr}},
-    {0x85, 1, X86InstInfo{"TEST",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                         0, nullptr}},
-    {0x86, 1, X86InstInfo{"XCHG",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,         0, nullptr}},
-    {0x87, 1, X86InstInfo{"XCHG",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                         0, nullptr}},
+    {0x84, 1, X86InstInfo{"TEST",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,         0}},
+    {0x85, 1, X86InstInfo{"TEST",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                         0}},
+    {0x86, 1, X86InstInfo{"XCHG",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,         0}},
+    {0x87, 1, X86InstInfo{"XCHG",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                         0}},
 
-    {0x88, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,         0, nullptr}},
-    {0x89, 1, X86InstInfo{"MOV",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                         0, nullptr}},
-    {0x8A, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,         0, nullptr}},
-    {0x8B, 1, X86InstInfo{"MOV",    TYPE_INST, FLAGS_MODRM,                         0, nullptr}},
-    {0x8C, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                      0, nullptr}},
-    {0x8D, 1, X86InstInfo{"LEA",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_MODRM,                         0, nullptr}},
-    {0x8E, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_16BIT) | FLAGS_MODRM,                      0, nullptr}},
-    {0x8F, 1, X86InstInfo{"POP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_ZERO_REG | FLAGS_DEBUG_MEM_ACCESS, 0, nullptr}},
-    {0x90, 8, X86InstInfo{"XCHG",   TYPE_INST, FLAGS_SF_REX_IN_BYTE | FLAGS_SF_SRC_RAX, 0, nullptr}},
-    {0x98, 1, X86InstInfo{"CDQE",   TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SF_SRC_RAX,     0, nullptr}},
-    {0x99, 1, X86InstInfo{"CQO",    TYPE_INST, FLAGS_SF_DST_RDX | FLAGS_SF_SRC_RAX,     0, nullptr}},
+    {0x88, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,         0}},
+    {0x89, 1, X86InstInfo{"MOV",    TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                         0}},
+    {0x8A, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM,         0}},
+    {0x8B, 1, X86InstInfo{"MOV",    TYPE_INST, FLAGS_MODRM,                         0}},
+    {0x8C, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                      0}},
+    {0x8D, 1, X86InstInfo{"LEA",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_MODRM,                         0}},
+    {0x8E, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_16BIT) | FLAGS_MODRM,                      0}},
+    {0x8F, 1, X86InstInfo{"POP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_ZERO_REG | FLAGS_DEBUG_MEM_ACCESS, 0}},
+    {0x90, 8, X86InstInfo{"XCHG",   TYPE_INST, FLAGS_SF_REX_IN_BYTE | FLAGS_SF_SRC_RAX, 0}},
+    {0x98, 1, X86InstInfo{"CDQE",   TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SF_SRC_RAX,     0}},
+    {0x99, 1, X86InstInfo{"CQO",    TYPE_INST, FLAGS_SF_DST_RDX | FLAGS_SF_SRC_RAX,     0}},
 
     // These three are all X87 instructions
-    {0x9B, 1, X86InstInfo{"FWAIT",  TYPE_INST, FLAGS_NONE,                              0, nullptr}},
-    {0x9C, 1, X86InstInfo{"PUSHF",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF),         0, nullptr}},
-    {0x9D, 1, X86InstInfo{"POPF",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_BLOCK_END,         0, nullptr}},
+    {0x9B, 1, X86InstInfo{"FWAIT",  TYPE_INST, FLAGS_NONE,                              0}},
+    {0x9C, 1, X86InstInfo{"PUSHF",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF),         0}},
+    {0x9D, 1, X86InstInfo{"POPF",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_BLOCK_END,         0}},
 
-    {0x9E, 1, X86InstInfo{"SAHF",   TYPE_INST, FLAGS_NONE,                              0, nullptr}},
-    {0x9F, 1, X86InstInfo{"LAHF",   TYPE_INST, FLAGS_NONE,                              0, nullptr}},
+    {0x9E, 1, X86InstInfo{"SAHF",   TYPE_INST, FLAGS_NONE,                              0}},
+    {0x9F, 1, X86InstInfo{"LAHF",   TYPE_INST, FLAGS_NONE,                              0}},
 
-    {0xA4, 1, X86InstInfo{"MOVSB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_DEBUG_MEM_ACCESS,                                            0, nullptr}},
-    {0xA5, 1, X86InstInfo{"MOVS",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS,                                                            0, nullptr}},
-    {0xA6, 1, X86InstInfo{"CMPSB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_DEBUG_MEM_ACCESS,                                            0, nullptr}},
-    {0xA7, 1, X86InstInfo{"CMPS",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS,                                                            0, nullptr}},
+    {0xA0, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_A0] }}},
+    {0xA1, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_A1] }}},
+    {0xA2, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_A2] }}},
+    {0xA3, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_A3] }}},
 
-    {0xA8, 1, X86InstInfo{"TEST",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX ,                                             1, nullptr}},
-    {0xA9, 1, X86InstInfo{"TEST",   TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2,                4, nullptr}},
-    {0xAA, 1, X86InstInfo{"STOS",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_SF_SRC_RAX,                   0, nullptr}},
-    {0xAB, 1, X86InstInfo{"STOS",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS | FLAGS_SF_SRC_RAX,                                   0, nullptr}},
-    {0xAC, 1, X86InstInfo{"LODS",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX | FLAGS_DEBUG_MEM_ACCESS,                                                      0, nullptr}},
-    {0xAD, 1, X86InstInfo{"LODS",   TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_DEBUG_MEM_ACCESS,                                                      0, nullptr}},
-    {0xAE, 1, X86InstInfo{"SCAS",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_SF_SRC_RAX,                                   0, nullptr}},
-    {0xAF, 1, X86InstInfo{"SCAS",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS | FLAGS_SF_SRC_RAX,                                   0, nullptr}},
+    {0xA4, 1, X86InstInfo{"MOVSB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_DEBUG_MEM_ACCESS,                                            0}},
+    {0xA5, 1, X86InstInfo{"MOVS",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS,                                                            0}},
+    {0xA6, 1, X86InstInfo{"CMPSB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_DEBUG_MEM_ACCESS,                                            0}},
+    {0xA7, 1, X86InstInfo{"CMPS",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS,                                                            0}},
 
-    {0xB0, 8, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_REX_IN_BYTE ,                                         1, nullptr}},
-    {0xB8, 8, X86InstInfo{"MOV",    TYPE_INST, FLAGS_SF_REX_IN_BYTE | FLAGS_DISPLACE_SIZE_DIV_2 | FLAGS_DISPLACE_SIZE_MUL_2, 4, nullptr}},
+    {0xA8, 1, X86InstInfo{"TEST",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX ,                                             1}},
+    {0xA9, 1, X86InstInfo{"TEST",   TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2,                4}},
+    {0xAA, 1, X86InstInfo{"STOS",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_SF_SRC_RAX,                   0}},
+    {0xAB, 1, X86InstInfo{"STOS",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS | FLAGS_SF_SRC_RAX,                                   0}},
+    {0xAC, 1, X86InstInfo{"LODS",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX | FLAGS_DEBUG_MEM_ACCESS,                                                      0}},
+    {0xAD, 1, X86InstInfo{"LODS",   TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_DEBUG_MEM_ACCESS,                                                      0}},
+    {0xAE, 1, X86InstInfo{"SCAS",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_SF_SRC_RAX,                                   0}},
+    {0xAF, 1, X86InstInfo{"SCAS",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS | FLAGS_SF_SRC_RAX,                                   0}},
 
-    {0xC2, 1, X86InstInfo{"RET",    TYPE_INST, FLAGS_SETS_RIP | FLAGS_BLOCK_END,                                             2, nullptr}},
-    {0xC3, 1, X86InstInfo{"RET",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_BLOCK_END ,                                                0, nullptr}},
-    {0xC8, 1, X86InstInfo{"ENTER",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS ,                                      3, nullptr}},
-    {0xC9, 1, X86InstInfo{"LEAVE",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS ,                                                0, nullptr}},
-    {0xCA, 2, X86InstInfo{"RETF",   TYPE_PRIV, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_BLOCK_END,                                                              0, nullptr}},
-    {0xCC, 1, X86InstInfo{"INT3",   TYPE_INST, FLAGS_BLOCK_END,                                                                                      0, nullptr}},
-    {0xCD, 1, X86InstInfo{"INT",    TYPE_INST, DEFAULT_SYSCALL_FLAGS,                                                                  1, nullptr}},
-    {0xCF, 1, X86InstInfo{"IRET",   TYPE_INST, FLAGS_SETS_RIP | FLAGS_BLOCK_END,                                                                                    0, nullptr}},
+    {0xB0, 8, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_REX_IN_BYTE ,                                         1}},
+    {0xB8, 8, X86InstInfo{"MOV",    TYPE_INST, FLAGS_SF_REX_IN_BYTE | FLAGS_DISPLACE_SIZE_DIV_2 | FLAGS_DISPLACE_SIZE_MUL_2, 4}},
 
-    {0xD7, 1, X86InstInfo{"XLAT",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS,                                                                           0, nullptr}},
+    {0xC2, 1, X86InstInfo{"RET",    TYPE_INST, FLAGS_SETS_RIP | FLAGS_BLOCK_END,                                             2}},
+    {0xC3, 1, X86InstInfo{"RET",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_BLOCK_END ,                                                0}},
+    {0xC8, 1, X86InstInfo{"ENTER",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS ,                                      3}},
+    {0xC9, 1, X86InstInfo{"LEAVE",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS ,                                                0}},
+    {0xCA, 1, X86InstInfo{"RETF",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_BLOCK_END,                                                              2}},
+    {0xCB, 1, X86InstInfo{"RETF",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_BLOCK_END,                                                              0}},
+    {0xCC, 1, X86InstInfo{"INT3",   TYPE_INST, FLAGS_BLOCK_END,                                                                                      0}},
+    {0xCD, 1, X86InstInfo{"INT",    TYPE_INST, DEFAULT_SYSCALL_FLAGS,                                                                  1}},
+    {0xCE, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_CE] }}},
+    {0xCF, 1, X86InstInfo{"IRET",   TYPE_INST, FLAGS_SETS_RIP | FLAGS_BLOCK_END,                                                                                    0}},
 
-    {0xE0, 1, X86InstInfo{"LOOPNE", TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_SF_SRC_RCX,                             1, nullptr}},
-    {0xE1, 1, X86InstInfo{"LOOPE",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_SF_SRC_RCX,                             1, nullptr}},
-    {0xE2, 1, X86InstInfo{"LOOP",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_SF_SRC_RCX,                             1, nullptr}},
-    {0xE3, 1, X86InstInfo{"JrCXZ",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT ,                             1, nullptr}},
+    {0xD4, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_D4] }}},
+    {0xD5, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_D5] }}},
+    {0xD6, 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = Primary_ArchSelect_LUT[ENTRY_D6] }}},
+    {0xD7, 1, X86InstInfo{"XLAT",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS,                                                                           0}},
+
+    {0xE0, 1, X86InstInfo{"LOOPNE", TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_SF_SRC_RCX,                             1}},
+    {0xE1, 1, X86InstInfo{"LOOPE",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_SF_SRC_RCX,                             1}},
+    {0xE2, 1, X86InstInfo{"LOOP",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_SF_SRC_RCX,                             1}},
+    {0xE3, 1, X86InstInfo{"JrCXZ",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT ,                             1}},
 
     // Should just throw GP
-    {0xE4, 2, X86InstInfo{"IN",     TYPE_INST, FLAGS_BLOCK_END,                                                                                                      1, nullptr}},
-    {0xE6, 2, X86InstInfo{"OUT",    TYPE_INST, FLAGS_BLOCK_END,                                                                                                      1, nullptr}},
+    {0xE4, 2, X86InstInfo{"IN",     TYPE_INST, FLAGS_BLOCK_END,                                                                                                      1}},
+    {0xE6, 2, X86InstInfo{"OUT",    TYPE_INST, FLAGS_BLOCK_END,                                                                                                      1}},
 
-    {0xE8, 1, X86InstInfo{"CALL",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2 | FLAGS_BLOCK_END | FLAGS_CALL , 4, nullptr}},
-    {0xE9, 1, X86InstInfo{"JMP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2 | FLAGS_BLOCK_END , 4, nullptr}},
-    {0xEB, 1, X86InstInfo{"JMP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_BLOCK_END ,                             1, nullptr}},
+    {0xE8, 1, X86InstInfo{"CALL",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2 | FLAGS_BLOCK_END | FLAGS_CALL , 4}},
+    {0xE9, 1, X86InstInfo{"JMP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2 | FLAGS_BLOCK_END , 4}},
+    {0xEB, 1, X86InstInfo{"JMP",    TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_SRC_SEXT | FLAGS_BLOCK_END ,                             1}},
 
     // Should just throw GP
-    {0xEC, 2, X86InstInfo{"IN",     TYPE_INST, FLAGS_BLOCK_END,             0, nullptr}},
-    {0xEE, 2, X86InstInfo{"OUT",    TYPE_INST, FLAGS_BLOCK_END,             0, nullptr}},
+    {0xEC, 2, X86InstInfo{"IN",     TYPE_INST, FLAGS_BLOCK_END,             0}},
+    {0xEE, 2, X86InstInfo{"OUT",    TYPE_INST, FLAGS_BLOCK_END,             0}},
 
-    {0xF1, 1, X86InstInfo{"INT1",   TYPE_INST, FLAGS_BLOCK_END,               0, nullptr}},
-    {0xF4, 1, X86InstInfo{"HLT",    TYPE_INST, FLAGS_BLOCK_END,               0, nullptr}},
-    {0xF5, 1, X86InstInfo{"CMC",    TYPE_INST, FLAGS_NONE,                0, nullptr}},
-    {0xF8, 1, X86InstInfo{"CLC",    TYPE_INST, FLAGS_NONE,                0, nullptr}},
-    {0xF9, 1, X86InstInfo{"STC",    TYPE_INST, FLAGS_NONE,                0, nullptr}},
-    {0xFA, 1, X86InstInfo{"CLI",    TYPE_INST, FLAGS_NONE,                0, nullptr}},
-    {0xFB, 1, X86InstInfo{"STI",    TYPE_INST, FLAGS_NONE,                0, nullptr}},
-    {0xFC, 1, X86InstInfo{"CLD",    TYPE_INST, FLAGS_NONE,                0, nullptr}},
-    {0xFD, 1, X86InstInfo{"STD",    TYPE_INST, FLAGS_NONE,                0, nullptr}},
+    {0xF1, 1, X86InstInfo{"INT1",   TYPE_INST, FLAGS_BLOCK_END,               0}},
+    {0xF4, 1, X86InstInfo{"HLT",    TYPE_INST, FLAGS_BLOCK_END,               0}},
+    {0xF5, 1, X86InstInfo{"CMC",    TYPE_INST, FLAGS_NONE,                0}},
+    {0xF8, 1, X86InstInfo{"CLC",    TYPE_INST, FLAGS_NONE,                0}},
+    {0xF9, 1, X86InstInfo{"STC",    TYPE_INST, FLAGS_NONE,                0}},
+    {0xFA, 1, X86InstInfo{"CLI",    TYPE_INST, FLAGS_NONE,                0}},
+    {0xFB, 1, X86InstInfo{"STI",    TYPE_INST, FLAGS_NONE,                0}},
+    {0xFC, 1, X86InstInfo{"CLD",    TYPE_INST, FLAGS_NONE,                0}},
+    {0xFD, 1, X86InstInfo{"STD",    TYPE_INST, FLAGS_NONE,                0}},
 
     // Two Byte table
-    {0x0F, 1, X86InstInfo{"",   TYPE_SECONDARY_TABLE_PREFIX, FLAGS_NONE,  0, nullptr}},
+    {0x0F, 1, X86InstInfo{"",   TYPE_SECONDARY_TABLE_PREFIX, FLAGS_NONE,  0}},
 
     // x87 table
-    {0xD8, 8, X86InstInfo{"",   TYPE_X87_TABLE_PREFIX, FLAGS_MODRM,        0, nullptr}},
+    {0xD8, 8, X86InstInfo{"",   TYPE_X87_TABLE_PREFIX, FLAGS_MODRM,        0}},
 
     // ModRM table
     // MoreBytes field repurposed for valid bits mask
-    {0x80, 1, X86InstInfo{"",   TYPE_GROUP_1, FLAGS_MODRM, 0, nullptr}},
-    {0x81, 1, X86InstInfo{"",   TYPE_GROUP_1, FLAGS_MODRM, 1, nullptr}},
-    {0x82, 1, X86InstInfo{"",   TYPE_GROUP_1, FLAGS_MODRM, 2, nullptr}},
-    {0x83, 1, X86InstInfo{"",   TYPE_GROUP_1, FLAGS_MODRM, 3, nullptr}},
-    {0xC0, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 0, nullptr}},
-    {0xC1, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 1, nullptr}},
-    {0xD0, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 2, nullptr}},
-    {0xD1, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 3, nullptr}},
-    {0xD2, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 4, nullptr}},
-    {0xD3, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 5, nullptr}},
-    {0xF6, 1, X86InstInfo{"",   TYPE_GROUP_3, FLAGS_MODRM, 0, nullptr}},
-    {0xF7, 1, X86InstInfo{"",   TYPE_GROUP_3, FLAGS_MODRM, 1, nullptr}},
-    {0xFE, 1, X86InstInfo{"",   TYPE_GROUP_4, FLAGS_MODRM, 0, nullptr}},
-    {0xFF, 1, X86InstInfo{"",   TYPE_GROUP_5, FLAGS_MODRM, 0, nullptr}},
+    {0x80, 1, X86InstInfo{"",   TYPE_GROUP_1, FLAGS_MODRM, 0}},
+    {0x81, 1, X86InstInfo{"",   TYPE_GROUP_1, FLAGS_MODRM, 1}},
+    {0x82, 1, X86InstInfo{"",   TYPE_GROUP_1, FLAGS_MODRM, 2}},
+    {0x83, 1, X86InstInfo{"",   TYPE_GROUP_1, FLAGS_MODRM, 3}},
+    {0xC0, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 0}},
+    {0xC1, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 1}},
+    {0xD0, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 2}},
+    {0xD1, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 3}},
+    {0xD2, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 4}},
+    {0xD3, 1, X86InstInfo{"",   TYPE_GROUP_2, FLAGS_MODRM, 5}},
+    {0xF6, 1, X86InstInfo{"",   TYPE_GROUP_3, FLAGS_MODRM, 0}},
+    {0xF7, 1, X86InstInfo{"",   TYPE_GROUP_3, FLAGS_MODRM, 1}},
+    {0xFE, 1, X86InstInfo{"",   TYPE_GROUP_4, FLAGS_MODRM, 0}},
+    {0xFF, 1, X86InstInfo{"",   TYPE_GROUP_5, FLAGS_MODRM, 0}},
 
     // Group 11
-    {0xC6, 1, X86InstInfo{"",   TYPE_GROUP_11, FLAGS_MODRM, 0, nullptr}},
-    {0xC7, 1, X86InstInfo{"",   TYPE_GROUP_11, FLAGS_MODRM, 1, nullptr}},
+    {0xC6, 1, X86InstInfo{"",   TYPE_GROUP_11, FLAGS_MODRM, 0}},
+    {0xC7, 1, X86InstInfo{"",   TYPE_GROUP_11, FLAGS_MODRM, 1}},
 
     // VEX table
-    {0xC4, 2, X86InstInfo{"",   TYPE_VEX_TABLE_PREFIX, FLAGS_NONE, 0, nullptr}},
+    {0xC4, 2, X86InstInfo{"",   TYPE_VEX_TABLE_PREFIX, FLAGS_NONE, 0}},
   };
 
-  GenerateTable(&Table.at(0), BaseOpTable, std::size(BaseOpTable));
+  GenerateTable(Table.data(), BaseOpTable, std::size(BaseOpTable));
   IR::InstallToTable(Table, IR::OpDispatch_BaseOpTable);
 
   return Table;
 }();
 
-void InitializeBaseTables(Context::OperatingMode Mode) {
-  static constexpr U8U8InfoStruct BaseOpTable_64[] = {
-    {0x06, 2, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                     0, nullptr}},
-    {0x0E, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                     0, nullptr}},
-    {0x16, 2, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                     0, nullptr}},
-    {0x1E, 2, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                     0, nullptr}},
-    {0x27, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                     0, nullptr}},
-    {0x2F, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                     0, nullptr}},
-    {0x37, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                     0, nullptr}},
-    {0x3F, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                     0, nullptr}},
-
-    // REX
-    {0x40, 16, X86InstInfo{"", TYPE_REX_PREFIX, FLAGS_NONE,        0, nullptr}},
-    {0x60, 2, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                           0, nullptr}},
-    {0x63, 1, X86InstInfo{"MOVSXD", TYPE_INST, GenFlagsDstSize(SIZE_64BIT) | FLAGS_MODRM,                                                                         0, nullptr}},
-    {0x9A, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                           0, nullptr}},
-    {0xA0, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX | FLAGS_MEM_OFFSET, 8, nullptr}},
-    {0xA2, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_SRC_RAX | FLAGS_MEM_OFFSET, 8, nullptr}},
-    {0xA1, 1, X86InstInfo{"MOV",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_MEM_OFFSET, 8, nullptr}},
-    {0xA3, 1, X86InstInfo{"MOV",    TYPE_INST, FLAGS_SF_SRC_RAX | FLAGS_MEM_OFFSET, 8, nullptr}},
-    {0xCE, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                                    0, nullptr}},
-    {0xD4, 2, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                                    0, nullptr}},
-    // `L1OM` Larrabee instructions used this as an escape byte.
-    // FEX will never support this.
-    {0xD6, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                                    0, nullptr}},
-    {0xEA, 1, X86InstInfo{"[INV]",  TYPE_INVALID, FLAGS_NONE,                                                                                                      0, nullptr}},
-  };
-
-  static constexpr U8U8InfoStruct BaseOpTable_32[] = {
-    {0x06, 1, X86InstInfo{"PUSH ES",  TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_DEBUG_MEM_ACCESS,            0, nullptr}},
-    {0x07, 1, X86InstInfo{"POP ES",   TYPE_INST, GenFlagsSizes(SIZE_16BIT, SIZE_DEF) | FLAGS_DEBUG_MEM_ACCESS,    0, nullptr}},
-    {0x0E, 1, X86InstInfo{"PUSH CS",  TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_DEBUG_MEM_ACCESS,            0, nullptr}},
-    {0x16, 1, X86InstInfo{"PUSH SS",  TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_DEBUG_MEM_ACCESS,            0, nullptr}},
-    {0x17, 1, X86InstInfo{"POP SS",   TYPE_INST, GenFlagsSizes(SIZE_16BIT, SIZE_DEF) | FLAGS_DEBUG_MEM_ACCESS,    0, nullptr}},
-    {0x1E, 1, X86InstInfo{"PUSH DS",  TYPE_INST, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_DEBUG_MEM_ACCESS,            0, nullptr}},
-    {0x1F, 1, X86InstInfo{"POP DS",   TYPE_INST, GenFlagsSizes(SIZE_16BIT, SIZE_DEF) | FLAGS_DEBUG_MEM_ACCESS,    0, nullptr}},
-    {0x27, 1, X86InstInfo{"DAA",      TYPE_INST, GenFlagsDstSize(SIZE_8BIT) | FLAGS_SF_DST_RAX,                   0, nullptr}},
-    {0x2F, 1, X86InstInfo{"DAS",      TYPE_INST, GenFlagsDstSize(SIZE_8BIT) | FLAGS_SF_DST_RAX,                   0, nullptr}},
-    {0x37, 1, X86InstInfo{"AAA",      TYPE_INST, GenFlagsDstSize(SIZE_16BIT) | FLAGS_SF_DST_RAX,                  0, nullptr}},
-    {0x3F, 1, X86InstInfo{"AAS",      TYPE_INST, GenFlagsDstSize(SIZE_16BIT) | FLAGS_SF_DST_RAX,                  0, nullptr}},
-
-    {0x40, 8, X86InstInfo{"INC",    TYPE_INST, FLAGS_SF_REX_IN_BYTE,                                              0, nullptr}},
-    {0x48, 8, X86InstInfo{"DEC",    TYPE_INST, FLAGS_SF_REX_IN_BYTE,                                              0, nullptr}},
-    {0x60, 1, X86InstInfo{"PUSHA",  TYPE_INST, FLAGS_DEBUG_MEM_ACCESS,                                            0, nullptr}},
-    {0x61, 1, X86InstInfo{"POPA",   TYPE_INST, FLAGS_DEBUG_MEM_ACCESS,                                            0, nullptr}},
-    {0x63, 1, X86InstInfo{"ARPL",   TYPE_INVALID, FLAGS_NONE,                                                     0, nullptr}},
-
-    {0x9A, 1, X86InstInfo{"CALLF",  TYPE_INST, FLAGS_NONE,                                                        0, nullptr}},
-    {0xA0, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX | FLAGS_MEM_OFFSET, 4, nullptr}},
-    {0xA2, 1, X86InstInfo{"MOV",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_SRC_RAX | FLAGS_MEM_OFFSET, 4, nullptr}},
-    {0xA1, 1, X86InstInfo{"MOV",    TYPE_INST, FLAGS_SF_DST_RAX | FLAGS_MEM_OFFSET,                               4, nullptr}},
-    {0xA3, 1, X86InstInfo{"MOV",    TYPE_INST, FLAGS_SF_SRC_RAX | FLAGS_MEM_OFFSET,                               4, nullptr}},
-    {0xCE, 1, X86InstInfo{"INTO",   TYPE_INST, FLAGS_NONE,                                                        0, nullptr}},
-    {0xD4, 1, X86InstInfo{"AAM",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX,                    1, nullptr}},
-    {0xD5, 1, X86InstInfo{"AAD",    TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX,                    1, nullptr}},
-    {0xD6, 1, X86InstInfo{"SALC",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_SF_DST_RAX | FLAGS_SF_SRC_RAX, 0, nullptr}},
-    {0xEA, 1, X86InstInfo{"JMPF",   TYPE_INST, FLAGS_NONE,                                                        0, nullptr}},
-  };
-
-  if (Mode == Context::MODE_64BIT) {
-    GenerateTable(&BaseOps.at(0), BaseOpTable_64, std::size(BaseOpTable_64));
-    IR::InstallToTable(BaseOps, IR::OpDispatch_BaseOpTable_64);
-  }
-  else {
-    GenerateTable(&BaseOps.at(0), BaseOpTable_32, std::size(BaseOpTable_32));
-    IR::InstallToTable(BaseOps, IR::OpDispatch_BaseOpTable_32);
-  }
-}
 }
 

FEXCore/Source/Interface/Core/X86Tables/DDDTables.cpp

@@ -13,48 +13,48 @@ $end_info$
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
 
-std::array<X86InstInfo, MAX_3DNOW_TABLE_SIZE> DDDNowOps = []() consteval {
+constexpr std::array<X86InstInfo, MAX_3DNOW_TABLE_SIZE> DDDNowOps = []() consteval {
   std::array<X86InstInfo, MAX_3DNOW_TABLE_SIZE> Table{};
   constexpr U8U8InfoStruct DDDNowOpTable[] = {
-    {0x0C, 1, X86InstInfo{"PI2FW",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x0D, 1, X86InstInfo{"PI2FD",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x1C, 1, X86InstInfo{"PF2IW",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x1D, 1, X86InstInfo{"PF2ID",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0x0C, 1, X86InstInfo{"PI2FW",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x0D, 1, X86InstInfo{"PI2FD",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x1C, 1, X86InstInfo{"PF2IW",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x1D, 1, X86InstInfo{"PF2ID",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
 
     // Inverse 3DNow! These two instructions are Geode product line specific
     // No CPUID for these, you're expected to read ID_CONFIG_MSR (1250h) bit 1
-    {0x86, 1, X86InstInfo{"PFRCPV",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x87, 1, X86InstInfo{"PFRSQRTV", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0x86, 1, X86InstInfo{"PFRCPV",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x87, 1, X86InstInfo{"PFRSQRTV", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
 
-    {0x8A, 1, X86InstInfo{"PFNACC",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x8E, 1, X86InstInfo{"PFPNACC",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0x8A, 1, X86InstInfo{"PFNACC",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x8E, 1, X86InstInfo{"PFPNACC",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
 
-    {0x90, 1, X86InstInfo{"PFCMPGE",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x94, 1, X86InstInfo{"PFMIN",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x96, 1, X86InstInfo{"PFRCP",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x97, 1, X86InstInfo{"PFRSQRT",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0x90, 1, X86InstInfo{"PFCMPGE",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x94, 1, X86InstInfo{"PFMIN",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x96, 1, X86InstInfo{"PFRCP",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x97, 1, X86InstInfo{"PFRSQRT",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
 
-    {0x9A, 1, X86InstInfo{"PFSUB",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0x9E, 1, X86InstInfo{"PFADD",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0x9A, 1, X86InstInfo{"PFSUB",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0x9E, 1, X86InstInfo{"PFADD",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
 
-    {0xA0, 1, X86InstInfo{"PFCMPGT",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0xA4, 1, X86InstInfo{"PFMAX",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0xA6, 1, X86InstInfo{"PFRCPIT1", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0xA7, 1, X86InstInfo{"PFRSQIT1", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0xA0, 1, X86InstInfo{"PFCMPGT",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0xA4, 1, X86InstInfo{"PFMAX",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0xA6, 1, X86InstInfo{"PFRCPIT1", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0xA7, 1, X86InstInfo{"PFRSQIT1", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
 
-    {0xAA, 1, X86InstInfo{"PFSUBR",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0xAE, 1, X86InstInfo{"PFACC",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0xAA, 1, X86InstInfo{"PFSUBR",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0xAE, 1, X86InstInfo{"PFACC",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
 
-    {0xB0, 1, X86InstInfo{"PFCMPEQ",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0xB4, 1, X86InstInfo{"PFMUL",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0xB6, 1, X86InstInfo{"PFRCPIT2", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0xB7, 1, X86InstInfo{"PMULHRW",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0xB0, 1, X86InstInfo{"PFCMPEQ",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0xB4, 1, X86InstInfo{"PFMUL",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0xB6, 1, X86InstInfo{"PFRCPIT2", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0xB7, 1, X86InstInfo{"PMULHRW",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
 
-    {0xBB, 1, X86InstInfo{"PSWAPD",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {0xBF, 1, X86InstInfo{"PAVGUSB",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
+    {0xBB, 1, X86InstInfo{"PSWAPD",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {0xBF, 1, X86InstInfo{"PAVGUSB",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
   };
 
-  GenerateTable(&Table.at(0), DDDNowOpTable, std::size(DDDNowOpTable));
+  GenerateTable(Table.data(), DDDNowOpTable, std::size(DDDNowOpTable));
 
   IR::InstallToTable(Table, IR::OpDispatch_DDDTable);
   return Table;

FEXCore/Source/Interface/Core/X86Tables/H0F38Tables.cpp

@@ -13,7 +13,7 @@ $end_info$
 
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
-std::array<X86InstInfo, MAX_0F_38_TABLE_SIZE> H0F38TableOps = []() consteval {
+constexpr std::array<X86InstInfo, MAX_0F_38_TABLE_SIZE> H0F38TableOps = []() consteval {
   std::array<X86InstInfo, MAX_0F_38_TABLE_SIZE> Table{};
 
 #define OPD(prefix, opcode) (((prefix) << 8) | opcode)
@@ -23,103 +23,103 @@ std::array<X86InstInfo, MAX_0F_38_TABLE_SIZE> H0F38TableOps = []() consteval {
   constexpr uint16_t PF_38_F3   = (1U << 2);
 
   constexpr U16U8InfoStruct H0F38Table[] = {
-    {OPD(PF_38_NONE, 0x00), 1, X86InstInfo{"PSHUFB",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x00), 1, X86InstInfo{"PSHUFB",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x01), 1, X86InstInfo{"PHADDW",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x01), 1, X86InstInfo{"PHADDW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x02), 1, X86InstInfo{"PHADDD",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x02), 1, X86InstInfo{"PHADDD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x03), 1, X86InstInfo{"PHADDSW",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x03), 1, X86InstInfo{"PHADDSW",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x04), 1, X86InstInfo{"PMADDUBSW",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x04), 1, X86InstInfo{"PMADDUBSW",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x05), 1, X86InstInfo{"PHSUBW",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x05), 1, X86InstInfo{"PHSUBW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x06), 1, X86InstInfo{"PHSUBD",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x06), 1, X86InstInfo{"PHSUBD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x07), 1, X86InstInfo{"PHSUBSW",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x07), 1, X86InstInfo{"PHSUBSW",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x08), 1, X86InstInfo{"PSIGNB",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x08), 1, X86InstInfo{"PSIGNB",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x09), 1, X86InstInfo{"PSIGNW",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x09), 1, X86InstInfo{"PSIGNW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x0A), 1, X86InstInfo{"PSIGND",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x0A), 1, X86InstInfo{"PSIGND",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x0B), 1, X86InstInfo{"PMULHRSW",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x0B), 1, X86InstInfo{"PMULHRSW",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
+    {OPD(PF_38_NONE, 0x00), 1, X86InstInfo{"PSHUFB",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x00), 1, X86InstInfo{"PSHUFB",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x01), 1, X86InstInfo{"PHADDW",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x01), 1, X86InstInfo{"PHADDW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x02), 1, X86InstInfo{"PHADDD",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x02), 1, X86InstInfo{"PHADDD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x03), 1, X86InstInfo{"PHADDSW",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x03), 1, X86InstInfo{"PHADDSW",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x04), 1, X86InstInfo{"PMADDUBSW",  TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x04), 1, X86InstInfo{"PMADDUBSW",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x05), 1, X86InstInfo{"PHSUBW",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x05), 1, X86InstInfo{"PHSUBW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x06), 1, X86InstInfo{"PHSUBD",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x06), 1, X86InstInfo{"PHSUBD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x07), 1, X86InstInfo{"PHSUBSW",    TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x07), 1, X86InstInfo{"PHSUBSW",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x08), 1, X86InstInfo{"PSIGNB",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x08), 1, X86InstInfo{"PSIGNB",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x09), 1, X86InstInfo{"PSIGNW",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x09), 1, X86InstInfo{"PSIGNW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x0A), 1, X86InstInfo{"PSIGND",     TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x0A), 1, X86InstInfo{"PSIGND",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x0B), 1, X86InstInfo{"PMULHRSW",   TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x0B), 1, X86InstInfo{"PMULHRSW",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
 
-    {OPD(PF_38_66,   0x10), 1, X86InstInfo{"PBLENDVB",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x14), 1, X86InstInfo{"BLENDVPS",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x15), 1, X86InstInfo{"BLENDVPD",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x17), 1, X86InstInfo{"PTEST",      TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x1C), 1, X86InstInfo{"PABSB",      TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x1C), 1, X86InstInfo{"PABSB",      TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x1D), 1, X86InstInfo{"PABSW",      TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x1D), 1, X86InstInfo{"PABSW",      TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0x1E), 1, X86InstInfo{"PABSD",      TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0, nullptr}},
-    {OPD(PF_38_66,   0x1E), 1, X86InstInfo{"PABSD",      TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
+    {OPD(PF_38_66,   0x10), 1, X86InstInfo{"PBLENDVB",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x14), 1, X86InstInfo{"BLENDVPS",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x15), 1, X86InstInfo{"BLENDVPD",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x17), 1, X86InstInfo{"PTEST",      TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x1C), 1, X86InstInfo{"PABSB",      TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x1C), 1, X86InstInfo{"PABSB",      TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x1D), 1, X86InstInfo{"PABSW",      TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x1D), 1, X86InstInfo{"PABSW",      TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0x1E), 1, X86InstInfo{"PABSD",      TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 0}},
+    {OPD(PF_38_66,   0x1E), 1, X86InstInfo{"PABSD",      TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
 
-    {OPD(PF_38_66,   0x20), 1, X86InstInfo{"PMOVSXBW",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x21), 1, X86InstInfo{"PMOVSXBD",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x22), 1, X86InstInfo{"PMOVSXBQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_16BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x23), 1, X86InstInfo{"PMOVSXWD",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x24), 1, X86InstInfo{"PMOVSXWQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x25), 1, X86InstInfo{"PMOVSXDQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x28), 1, X86InstInfo{"PMULDQ",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x29), 1, X86InstInfo{"PCMPEQQ",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x2A), 1, X86InstInfo{"MOVNTDQA",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x2B), 1, X86InstInfo{"PACKUSDW",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
+    {OPD(PF_38_66,   0x20), 1, X86InstInfo{"PMOVSXBW",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x21), 1, X86InstInfo{"PMOVSXBD",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x22), 1, X86InstInfo{"PMOVSXBQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_16BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x23), 1, X86InstInfo{"PMOVSXWD",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x24), 1, X86InstInfo{"PMOVSXWQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x25), 1, X86InstInfo{"PMOVSXDQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x28), 1, X86InstInfo{"PMULDQ",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x29), 1, X86InstInfo{"PCMPEQQ",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x2A), 1, X86InstInfo{"MOVNTDQA",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x2B), 1, X86InstInfo{"PACKUSDW",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
 
-    {OPD(PF_38_66,   0x30), 1, X86InstInfo{"PMOVZXBW",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x31), 1, X86InstInfo{"PMOVZXBD",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x32), 1, X86InstInfo{"PMOVZXBQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_16BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x33), 1, X86InstInfo{"PMOVZXWD",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x34), 1, X86InstInfo{"PMOVZXWQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x35), 1, X86InstInfo{"PMOVZXDQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x37), 1, X86InstInfo{"PCMPGTQ",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x38), 1, X86InstInfo{"PMINSB",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x39), 1, X86InstInfo{"PMINSD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x3A), 1, X86InstInfo{"PMINUW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x3B), 1, X86InstInfo{"PMINUD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x3C), 1, X86InstInfo{"PMAXSB",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x3D), 1, X86InstInfo{"PMAXSD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x3E), 1, X86InstInfo{"PMAXUW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x3F), 1, X86InstInfo{"PMAXUD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
+    {OPD(PF_38_66,   0x30), 1, X86InstInfo{"PMOVZXBW",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x31), 1, X86InstInfo{"PMOVZXBD",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x32), 1, X86InstInfo{"PMOVZXBQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_16BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x33), 1, X86InstInfo{"PMOVZXWD",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x34), 1, X86InstInfo{"PMOVZXWQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x35), 1, X86InstInfo{"PMOVZXDQ",   TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x37), 1, X86InstInfo{"PCMPGTQ",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x38), 1, X86InstInfo{"PMINSB",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x39), 1, X86InstInfo{"PMINSD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x3A), 1, X86InstInfo{"PMINUW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x3B), 1, X86InstInfo{"PMINUD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x3C), 1, X86InstInfo{"PMAXSB",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x3D), 1, X86InstInfo{"PMAXSD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x3E), 1, X86InstInfo{"PMAXUW",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x3F), 1, X86InstInfo{"PMAXUD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
 
-    {OPD(PF_38_66,   0x40), 1, X86InstInfo{"PMULLD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0x41), 1, X86InstInfo{"PHMINPOSUW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
+    {OPD(PF_38_66,   0x40), 1, X86InstInfo{"PMULLD",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0x41), 1, X86InstInfo{"PHMINPOSUW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
 
-    {OPD(PF_38_NONE, 0xC8), 1, X86InstInfo{"SHA1NEXTE",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0xC9), 1, X86InstInfo{"SHA1MSG1",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0xCA), 1, X86InstInfo{"SHA1MSG2",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
+    {OPD(PF_38_NONE, 0xC8), 1, X86InstInfo{"SHA1NEXTE",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0xC9), 1, X86InstInfo{"SHA1MSG1",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0xCA), 1, X86InstInfo{"SHA1MSG2",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
 
-    {OPD(PF_38_NONE, 0xCB), 1, X86InstInfo{"SHA256RNDS2", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0xCC), 1, X86InstInfo{"SHA256MSG1",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_NONE, 0xCD), 1, X86InstInfo{"SHA256MSG2",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
+    {OPD(PF_38_NONE, 0xCB), 1, X86InstInfo{"SHA256RNDS2", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0xCC), 1, X86InstInfo{"SHA256MSG1",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_NONE, 0xCD), 1, X86InstInfo{"SHA256MSG2",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
 
-    {OPD(PF_38_66,   0xDB), 1, X86InstInfo{"AESIMC",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0xDC), 1, X86InstInfo{"AESENC",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0xDD), 1, X86InstInfo{"AESENCLAST", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0xDE), 1, X86InstInfo{"AESDEC",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
-    {OPD(PF_38_66,   0xDF), 1, X86InstInfo{"AESDECLAST", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
+    {OPD(PF_38_66,   0xDB), 1, X86InstInfo{"AESIMC",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0xDC), 1, X86InstInfo{"AESENC",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0xDD), 1, X86InstInfo{"AESENCLAST", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0xDE), 1, X86InstInfo{"AESDEC",     TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
+    {OPD(PF_38_66,   0xDF), 1, X86InstInfo{"AESDECLAST", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0}},
 
-    {OPD(PF_38_NONE, 0xF0), 1, X86InstInfo{"MOVBE",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(PF_38_NONE, 0xF1), 1, X86InstInfo{"MOVBE",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
+    {OPD(PF_38_NONE, 0xF0), 1, X86InstInfo{"MOVBE",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(PF_38_NONE, 0xF1), 1, X86InstInfo{"MOVBE",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0}},
 
-    {OPD(PF_38_66, 0xF0), 1, X86InstInfo{"MOVBE",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(PF_38_66, 0xF1), 1, X86InstInfo{"MOVBE",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
+    {OPD(PF_38_66, 0xF0), 1, X86InstInfo{"MOVBE",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(PF_38_66, 0xF1), 1, X86InstInfo{"MOVBE",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0}},
 
-    {OPD(PF_38_F2,   0xF0), 1, X86InstInfo{"CRC32",      TYPE_INST, GenFlagsSizes(SIZE_DEF, SIZE_8BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPD(PF_38_F2,   0xF1), 1, X86InstInfo{"CRC32",      TYPE_INST, FLAGS_MODRM, 0, nullptr}},
-    {OPD(PF_38_66 | PF_38_F2,   0xF0), 1, X86InstInfo{"CRC32",      TYPE_INST, GenFlagsSizes(SIZE_DEF, SIZE_8BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPD(PF_38_66 | PF_38_F2,   0xF1), 1, X86InstInfo{"CRC32",      TYPE_INST, FLAGS_MODRM, 0, nullptr}},
+    {OPD(PF_38_F2,   0xF0), 1, X86InstInfo{"CRC32",      TYPE_INST, GenFlagsSizes(SIZE_DEF, SIZE_8BIT) | FLAGS_MODRM, 0}},
+    {OPD(PF_38_F2,   0xF1), 1, X86InstInfo{"CRC32",      TYPE_INST, FLAGS_MODRM, 0}},
+    {OPD(PF_38_66 | PF_38_F2,   0xF0), 1, X86InstInfo{"CRC32",      TYPE_INST, GenFlagsSizes(SIZE_DEF, SIZE_8BIT) | FLAGS_MODRM, 0}},
+    {OPD(PF_38_66 | PF_38_F2,   0xF1), 1, X86InstInfo{"CRC32",      TYPE_INST, FLAGS_MODRM, 0}},
 
-    {OPD(PF_38_66,   0xF6), 1, X86InstInfo{"ADCX",       TYPE_INST, FLAGS_MODRM | FLAGS_NO_OVERLAY66, 0, nullptr}},
-    {OPD(PF_38_F3,   0xF6), 1, X86InstInfo{"ADOX",       TYPE_INST, FLAGS_MODRM, 0, nullptr}},
+    {OPD(PF_38_66,   0xF6), 1, X86InstInfo{"ADCX",       TYPE_INST, FLAGS_MODRM | FLAGS_NO_OVERLAY66, 0}},
+    {OPD(PF_38_F3,   0xF6), 1, X86InstInfo{"ADOX",       TYPE_INST, FLAGS_MODRM, 0}},
   };
 #undef OPD
 
-  GenerateTable(&Table.at(0), H0F38Table, std::size(H0F38Table));
+  GenerateTable(Table.data(), H0F38Table, std::size(H0F38Table));
 
   IR::InstallToTable(Table, IR::OpDispatch_H0F38Table);
   return Table;

FEXCore/Source/Interface/Core/X86Tables/H0F3ATables.cpp

@@ -19,71 +19,82 @@ using namespace InstFlags;
 constexpr uint16_t PF_3A_NONE = 0;
 constexpr uint16_t PF_3A_66   = 1;
 
-std::array<X86InstInfo, MAX_0F_3A_TABLE_SIZE> H0F3ATableOps = []() consteval {
+enum H0F3A_LUT {
+  ENTRY_1_3A_66_16,
+  ENTRY_1_3A_66_22,
+  ENTRY_MAX,
+};
+
+constexpr std::array<X86InstInfo[2], ENTRY_MAX> H0F3A_ArchSelect_LUT = {{
+  // ENTRY_1_3A_66_16
+  {
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"PEXTRQ",          TYPE_INST, GenFlagsSizes(SIZE_64BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::PExtrOp, IR::OpSize::i64Bit> }},
+  },
+  // ENTRY_1_3A_66_22
+  {
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"PINSRQ",          TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_SRC_GPR, 1, { .OpDispatch = &IR::OpDispatchBuilder::PINSROp<IR::OpSize::i64Bit> }},
+  },
+}};
+
+constexpr std::array<X86InstInfo, MAX_0F_3A_TABLE_SIZE> H0F3ATableOps = []() consteval {
   std::array<X86InstInfo, MAX_0F_3A_TABLE_SIZE> Table{};
   auto TableGen = []<uint16_t REX>() consteval {
     constexpr U16U8InfoStruct Table[] = {
-      {OPD(REX, PF_3A_NONE, 0x0F), 1, X86InstInfo{"PALIGNR",         TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x08), 1, X86InstInfo{"ROUNDPS",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x09), 1, X86InstInfo{"ROUNDPD",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x0A), 1, X86InstInfo{"ROUNDSS",         TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x0B), 1, X86InstInfo{"ROUNDSD",         TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x0C), 1, X86InstInfo{"BLENDPS",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,           1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x0D), 1, X86InstInfo{"BLENDPD",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,           1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x0E), 1, X86InstInfo{"PBLENDW",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,           1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x0F), 1, X86InstInfo{"PALIGNR",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
+      {OPD(REX, PF_3A_NONE, 0x0F), 1, X86InstInfo{"PALIGNR",         TYPE_INST, GenFlagsSameSize(SIZE_64BIT)  | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+      {OPD(REX, PF_3A_66,   0x08), 1, X86InstInfo{"ROUNDPS",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x09), 1, X86InstInfo{"ROUNDPD",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x0A), 1, X86InstInfo{"ROUNDSS",         TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x0B), 1, X86InstInfo{"ROUNDSD",         TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x0C), 1, X86InstInfo{"BLENDPS",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,           1}},
+      {OPD(REX, PF_3A_66,   0x0D), 1, X86InstInfo{"BLENDPD",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,           1}},
+      {OPD(REX, PF_3A_66,   0x0E), 1, X86InstInfo{"PBLENDW",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,           1}},
+      {OPD(REX, PF_3A_66,   0x0F), 1, X86InstInfo{"PALIGNR",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
 
-      {OPD(REX, PF_3A_66,   0x14), 1, X86InstInfo{"PEXTRB",          TYPE_INST, GenFlagsSizes(SIZE_32BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x15), 1, X86InstInfo{"PEXTRW",          TYPE_INST, GenFlagsSizes(SIZE_16BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x17), 1, X86InstInfo{"EXTRACTPS",       TYPE_INST, GenFlagsSizes(SIZE_32BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1, nullptr}},
+      {OPD(REX, PF_3A_66,   0x14), 1, X86InstInfo{"PEXTRB",          TYPE_INST, GenFlagsSizes(SIZE_32BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x15), 1, X86InstInfo{"PEXTRW",          TYPE_INST, GenFlagsSizes(SIZE_16BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x17), 1, X86InstInfo{"EXTRACTPS",       TYPE_INST, GenFlagsSizes(SIZE_32BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1}},
 
-      {OPD(REX, PF_3A_66,   0x20), 1, X86InstInfo{"PINSRB",          TYPE_INST, GenFlagsDstSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_SRC_GPR,           1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x21), 1, X86InstInfo{"INSERTPS",        TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,           1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x40), 1, X86InstInfo{"DPPS",            TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x41), 1, X86InstInfo{"DPPD",            TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x42), 1, X86InstInfo{"MPSADBW",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x44), 1, X86InstInfo{"PCLMULQDQ",       TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
+      {OPD(REX, PF_3A_66,   0x20), 1, X86InstInfo{"PINSRB",          TYPE_INST, GenFlagsDstSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_SRC_GPR,           1}},
+      {OPD(REX, PF_3A_66,   0x21), 1, X86InstInfo{"INSERTPS",        TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS,           1}},
+      {OPD(REX, PF_3A_66,   0x40), 1, X86InstInfo{"DPPS",            TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x41), 1, X86InstInfo{"DPPD",            TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x42), 1, X86InstInfo{"MPSADBW",         TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x44), 1, X86InstInfo{"PCLMULQDQ",       TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
 
-      {OPD(REX, PF_3A_66,   0x60), 1, X86InstInfo{"PCMPESTRM",       TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x61), 1, X86InstInfo{"PCMPESTRI",       TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x62), 1, X86InstInfo{"PCMPISTRM",       TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
-      {OPD(REX, PF_3A_66,   0x63), 1, X86InstInfo{"PCMPISTRI",       TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
+      {OPD(REX, PF_3A_66,   0x60), 1, X86InstInfo{"PCMPESTRM",       TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x61), 1, X86InstInfo{"PCMPESTRI",       TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x62), 1, X86InstInfo{"PCMPISTRM",       TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
+      {OPD(REX, PF_3A_66,   0x63), 1, X86InstInfo{"PCMPISTRI",       TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
 
-      {OPD(REX, PF_3A_NONE, 0xCC), 1, X86InstInfo{"SHA1RNDS4",       TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
+      {OPD(REX, PF_3A_NONE, 0xCC), 1, X86InstInfo{"SHA1RNDS4",       TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
 
-      {OPD(REX, PF_3A_66,   0xDF), 1, X86InstInfo{"AESKEYGENASSIST", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1, nullptr}},
+      {OPD(REX, PF_3A_66,   0xDF), 1, X86InstInfo{"AESKEYGENASSIST", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 1}},
     };
     return std::to_array(Table);
   };
   constexpr auto H0F3ATable_IgnoresREX0 = TableGen.template operator()<0>();
   constexpr auto H0F3ATable_IgnoresREX1 = TableGen.template operator()<1>();
 
-  GenerateTable(&Table.at(0), &H0F3ATable_IgnoresREX0.at(0), H0F3ATable_IgnoresREX0.size());
-  GenerateTable(&Table.at(0), &H0F3ATable_IgnoresREX1.at(0), H0F3ATable_IgnoresREX1.size());
+  GenerateTable(Table.data(), H0F3ATable_IgnoresREX0.data(), H0F3ATable_IgnoresREX0.size());
+  GenerateTable(Table.data(), H0F3ATable_IgnoresREX1.data(), H0F3ATable_IgnoresREX1.size());
 
-  constexpr U16U8InfoStruct TableNeedsREX[] = {
-    {OPD(0, PF_3A_66,   0x16), 1, X86InstInfo{"PEXTRD",          TYPE_INST, GenFlagsSizes(SIZE_32BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1, nullptr}},
-    {OPD(0, PF_3A_66,   0x22), 1, X86InstInfo{"PINSRD",          TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_SRC_GPR,           1, nullptr}},
+  constexpr U16U8InfoStruct TableNeedsREX0[] = {
+    {OPD(0, PF_3A_66,   0x16), 1, X86InstInfo{"PEXTRD",          TYPE_INST, GenFlagsSizes(SIZE_32BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1}},
+    {OPD(0, PF_3A_66,   0x22), 1, X86InstInfo{"PINSRD",          TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_32BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_SRC_GPR,           1}},
   };
-  GenerateTable(&Table.at(0), TableNeedsREX, std::size(TableNeedsREX));
+  GenerateTable(Table.data(), TableNeedsREX0, std::size(TableNeedsREX0));
+
+  constexpr U16U8InfoStruct TableNeedsREX1[] = {
+    {OPD(1, PF_3A_66,   0x16), 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = H0F3A_ArchSelect_LUT[ENTRY_1_3A_66_16] }}},
+    {OPD(1, PF_3A_66,   0x22), 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = H0F3A_ArchSelect_LUT[ENTRY_1_3A_66_22] }}},
+  };
+  GenerateTable(Table.data(), TableNeedsREX1, std::size(TableNeedsREX1));
 
   IR::InstallToTable(Table, IR::OpDispatch_H0F3ATableIgnoreREX);
   IR::InstallToTable(Table, IR::OpDispatch_H0F3ATableNeedsREX0);
 
   return Table;
 }();
-
-void InitializeH0F3ATables(Context::OperatingMode Mode) {
-  static constexpr U16U8InfoStruct H0F3ATable_64[] = {
-    {OPD(1, PF_3A_66,   0x16), 1, X86InstInfo{"PEXTRQ",          TYPE_INST, GenFlagsSizes(SIZE_64BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_DST_GPR | FLAGS_XMM_FLAGS, 1, nullptr}},
-    {OPD(1, PF_3A_66,   0x22), 1, X86InstInfo{"PINSRQ",          TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_SRC_GPR,           1, nullptr}},
-  };
-
-#undef OPD
-
-  if (Mode == Context::MODE_64BIT) {
-    GenerateTable(&H0F3ATableOps.at(0), H0F3ATable_64, std::size(H0F3ATable_64));
-    IR::InstallToTable(H0F3ATableOps, IR::OpDispatch_H0F3ATable_64);
-  }
-}
 }

FEXCore/Source/Interface/Core/X86Tables/PrimaryGroupTables.cpp

@@ -14,168 +14,197 @@ $end_info$
 
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
-std::array<X86InstInfo, MAX_INST_GROUP_TABLE_SIZE> PrimaryInstGroupOps = []() consteval {
+enum PrimaryGroup_LUT {
+  ENTRY_1_82_0,
+  ENTRY_1_82_1,
+  ENTRY_1_82_2,
+  ENTRY_1_82_3,
+  ENTRY_1_82_4,
+  ENTRY_1_82_5,
+  ENTRY_1_82_6,
+  ENTRY_1_82_7,
+  ENTRY_MAX,
+};
+
+constexpr std::array<X86InstInfo[2], ENTRY_MAX> PrimaryGroup_ArchSelect_LUT = {{
+  {
+    {"ADD",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, { .OpDispatch = &IR::OpDispatchBuilder::SecondaryALUOp }},
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  {
+    {"OR",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, { .OpDispatch = &IR::OpDispatchBuilder::SecondaryALUOp }},
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  {
+    {"ADC",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::ADCOp, 1> }},
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  {
+    {"SBB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::SBBOp, 1> }},
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  {
+    {"AND",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, { .OpDispatch = &IR::OpDispatchBuilder::SecondaryALUOp }},
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  {
+    {"SUB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, { .OpDispatch = &IR::OpDispatchBuilder::SecondaryALUOp }},
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  {
+    {"XOR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, { .OpDispatch = &IR::OpDispatchBuilder::SecondaryALUOp }},
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+  {
+    {"CMP",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::CMPOp, 1> }},
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+  },
+}};
+
+constexpr std::array<X86InstInfo, MAX_INST_GROUP_TABLE_SIZE> PrimaryInstGroupOps = []() consteval {
   std::array<X86InstInfo, MAX_INST_GROUP_TABLE_SIZE> Table{};
 #define OPD(group, prefix, Reg) (((group - FEXCore::X86Tables::TYPE_GROUP_1) << 6) | (prefix) << 3 | (Reg))
   constexpr U16U8InfoStruct PrimaryGroupOpTable[] = {
     // GROUP_1 | 0x80 | reg
-    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 0), 1, X86InstInfo{"ADD",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 1), 1, X86InstInfo{"OR",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 2), 1, X86InstInfo{"ADC",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 3), 1, X86InstInfo{"SBB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 4), 1, X86InstInfo{"AND",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 5), 1, X86InstInfo{"SUB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 6), 1, X86InstInfo{"XOR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 7), 1, X86InstInfo{"CMP",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 0), 1, X86InstInfo{"ADD",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 1), 1, X86InstInfo{"OR",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 2), 1, X86InstInfo{"ADC",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 3), 1, X86InstInfo{"SBB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 4), 1, X86InstInfo{"AND",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 5), 1, X86InstInfo{"SUB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 6), 1, X86InstInfo{"XOR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x80), 7), 1, X86InstInfo{"CMP",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
 
-    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 0), 1, X86InstInfo{"ADD",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 1), 1, X86InstInfo{"OR",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 2), 1, X86InstInfo{"ADC",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 3), 1, X86InstInfo{"SBB",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 4), 1, X86InstInfo{"AND",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 5), 1, X86InstInfo{"SUB",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 6), 1, X86InstInfo{"XOR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 7), 1, X86InstInfo{"CMP",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 0), 1, X86InstInfo{"ADD",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 1), 1, X86InstInfo{"OR",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 2), 1, X86InstInfo{"ADC",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 3), 1, X86InstInfo{"SBB",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 4), 1, X86InstInfo{"AND",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 5), 1, X86InstInfo{"SUB",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 6), 1, X86InstInfo{"XOR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x81), 7), 1, X86InstInfo{"CMP",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
 
-    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 0), 1, X86InstInfo{"ADD",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 1), 1, X86InstInfo{"OR",   TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 2), 1, X86InstInfo{"ADC",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 3), 1, X86InstInfo{"SBB",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 4), 1, X86InstInfo{"AND",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 5), 1, X86InstInfo{"SUB",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 6), 1, X86InstInfo{"XOR",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 7), 1, X86InstInfo{"CMP",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1, nullptr}},
+    // Duplicates the 0x80 opcode group
+    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 0), 1, X86InstInfo{"",  TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = PrimaryGroup_ArchSelect_LUT[ENTRY_1_82_0] }}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 1), 1, X86InstInfo{"",  TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = PrimaryGroup_ArchSelect_LUT[ENTRY_1_82_1] }}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 2), 1, X86InstInfo{"",  TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = PrimaryGroup_ArchSelect_LUT[ENTRY_1_82_2] }}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 3), 1, X86InstInfo{"",  TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = PrimaryGroup_ArchSelect_LUT[ENTRY_1_82_3] }}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 4), 1, X86InstInfo{"",  TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = PrimaryGroup_ArchSelect_LUT[ENTRY_1_82_4] }}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 5), 1, X86InstInfo{"",  TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = PrimaryGroup_ArchSelect_LUT[ENTRY_1_82_5] }}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 6), 1, X86InstInfo{"",  TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = PrimaryGroup_ArchSelect_LUT[ENTRY_1_82_6] }}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 7), 1, X86InstInfo{"",  TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = PrimaryGroup_ArchSelect_LUT[ENTRY_1_82_7] }}},
+
+    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 0), 1, X86InstInfo{"ADD",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 1), 1, X86InstInfo{"OR",   TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 2), 1, X86InstInfo{"ADC",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 3), 1, X86InstInfo{"SBB",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 4), 1, X86InstInfo{"AND",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 5), 1, X86InstInfo{"SUB",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 6), 1, X86InstInfo{"XOR",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1}},
+    {OPD(TYPE_GROUP_1, OpToIndex(0x83), 7), 1, X86InstInfo{"CMP",  TYPE_INST, FLAGS_SRC_SEXT | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     1}},
 
     // GROUP 2
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 0), 1, X86InstInfo{"ROL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 1), 1, X86InstInfo{"ROR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 2), 1, X86InstInfo{"RCL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 3), 1, X86InstInfo{"RCR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 4), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 5), 1, X86InstInfo{"SHR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 6), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 7), 1, X86InstInfo{"SAR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 0), 1, X86InstInfo{"ROL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 1), 1, X86InstInfo{"ROR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 2), 1, X86InstInfo{"RCL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 3), 1, X86InstInfo{"RCR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 4), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 5), 1, X86InstInfo{"SHR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 6), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC0), 7), 1, X86InstInfo{"SAR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
 
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 0), 1, X86InstInfo{"ROL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 1), 1, X86InstInfo{"ROR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 2), 1, X86InstInfo{"RCL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 3), 1, X86InstInfo{"RCR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 4), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 5), 1, X86InstInfo{"SHR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 6), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 7), 1, X86InstInfo{"SAR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1, nullptr}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 0), 1, X86InstInfo{"ROL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 1), 1, X86InstInfo{"ROR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 2), 1, X86InstInfo{"RCL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 3), 1, X86InstInfo{"RCR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 4), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 5), 1, X86InstInfo{"SHR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 6), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xC1), 7), 1, X86InstInfo{"SAR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      1}},
 
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 0), 1, X86InstInfo{"ROL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 1), 1, X86InstInfo{"ROR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 2), 1, X86InstInfo{"RCL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 3), 1, X86InstInfo{"RCR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 4), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 5), 1, X86InstInfo{"SHR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 6), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 7), 1, X86InstInfo{"SAR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 0), 1, X86InstInfo{"ROL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 1), 1, X86InstInfo{"ROR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 2), 1, X86InstInfo{"RCL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 3), 1, X86InstInfo{"RCR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 4), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 5), 1, X86InstInfo{"SHR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 6), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD0), 7), 1, X86InstInfo{"SAR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
 
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 0), 1, X86InstInfo{"ROL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 1), 1, X86InstInfo{"ROR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 2), 1, X86InstInfo{"RCL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 3), 1, X86InstInfo{"RCR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 4), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 5), 1, X86InstInfo{"SHR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 6), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 7), 1, X86InstInfo{"SAR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 0), 1, X86InstInfo{"ROL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 1), 1, X86InstInfo{"ROR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 2), 1, X86InstInfo{"RCL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 3), 1, X86InstInfo{"RCR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 4), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 5), 1, X86InstInfo{"SHR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 6), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD1), 7), 1, X86InstInfo{"SAR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
 
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 0), 1, X86InstInfo{"ROL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 1), 1, X86InstInfo{"ROR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 2), 1, X86InstInfo{"RCL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 3), 1, X86InstInfo{"RCR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 4), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 5), 1, X86InstInfo{"SHR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 6), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 7), 1, X86InstInfo{"SAR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0, nullptr}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 0), 1, X86InstInfo{"ROL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 1), 1, X86InstInfo{"ROR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 2), 1, X86InstInfo{"RCL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 3), 1, X86InstInfo{"RCR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 4), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 5), 1, X86InstInfo{"SHR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 6), 1, X86InstInfo{"SHL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD2), 7), 1, X86InstInfo{"SAR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                   0}},
 
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 0), 1, X86InstInfo{"ROL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 1), 1, X86InstInfo{"ROR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 2), 1, X86InstInfo{"RCL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 3), 1, X86InstInfo{"RCR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 4), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 5), 1, X86InstInfo{"SHR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 6), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0, nullptr}},
-    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 7), 1, X86InstInfo{"SAR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0, nullptr}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 0), 1, X86InstInfo{"ROL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 1), 1, X86InstInfo{"ROR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 2), 1, X86InstInfo{"RCL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 3), 1, X86InstInfo{"RCR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 4), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 5), 1, X86InstInfo{"SHR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 6), 1, X86InstInfo{"SHL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0}},
+    {OPD(TYPE_GROUP_2, OpToIndex(0xD3), 7), 1, X86InstInfo{"SAR",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_SRC_RCX,                                   0}},
 
     // GROUP 3
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 0), 1, X86InstInfo{"TEST", TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 1), 1, X86InstInfo{"TEST", TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 2), 1, X86InstInfo{"NOT",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 3), 1, X86InstInfo{"NEG",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 4), 1, X86InstInfo{"MUL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 5), 1, X86InstInfo{"IMUL", TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 6), 1, X86InstInfo{"DIV",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 7), 1, X86InstInfo{"IDIV", TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0, nullptr}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 0), 1, X86InstInfo{"TEST", TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 1), 1, X86InstInfo{"TEST", TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 2), 1, X86InstInfo{"NOT",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 3), 1, X86InstInfo{"NEG",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 4), 1, X86InstInfo{"MUL",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 5), 1, X86InstInfo{"IMUL", TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 6), 1, X86InstInfo{"DIV",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF6), 7), 1, X86InstInfo{"IDIV", TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      0}},
 
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 0), 1, X86InstInfo{"TEST", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 1), 1, X86InstInfo{"TEST", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 2), 1, X86InstInfo{"NOT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 3), 1, X86InstInfo{"NEG",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 4), 1, X86InstInfo{"MUL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 5), 1, X86InstInfo{"IMUL", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 6), 1, X86InstInfo{"DIV",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
-    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 7), 1, X86InstInfo{"IDIV", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0, nullptr}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 0), 1, X86InstInfo{"TEST", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 1), 1, X86InstInfo{"TEST", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT64BIT | FLAGS_DISPLACE_SIZE_DIV_2,                          4}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 2), 1, X86InstInfo{"NOT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 3), 1, X86InstInfo{"NEG",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 4), 1, X86InstInfo{"MUL",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 5), 1, X86InstInfo{"IMUL", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 6), 1, X86InstInfo{"DIV",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
+    {OPD(TYPE_GROUP_3, OpToIndex(0xF7), 7), 1, X86InstInfo{"IDIV", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                      0}},
 
     // GROUP 4
-    {OPD(TYPE_GROUP_4, OpToIndex(0xFE), 0), 1, X86InstInfo{"INC",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     0, nullptr}},
-    {OPD(TYPE_GROUP_4, OpToIndex(0xFE), 1), 1, X86InstInfo{"DEC",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     0, nullptr}},
-    {OPD(TYPE_GROUP_4, OpToIndex(0xFE), 2), 6, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                                       0, nullptr}},
+    {OPD(TYPE_GROUP_4, OpToIndex(0xFE), 0), 1, X86InstInfo{"INC",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     0}},
+    {OPD(TYPE_GROUP_4, OpToIndex(0xFE), 1), 1, X86InstInfo{"DEC",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                     0}},
+    {OPD(TYPE_GROUP_4, OpToIndex(0xFE), 2), 6, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                                       0}},
 
     // GROUP 5
-    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 0), 1, X86InstInfo{"INC",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                     0, nullptr}},
-    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 1), 1, X86InstInfo{"DEC",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                     0, nullptr}},
-    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 2), 1, X86InstInfo{"CALL",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_MODRM | FLAGS_BLOCK_END | FLAGS_CALL , 0, nullptr}},
-    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 3), 1, X86InstInfo{"CALLF", TYPE_INST, FLAGS_SETS_RIP | FLAGS_MODRM | FLAGS_BLOCK_END,                  0, nullptr}},
-    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 4), 1, X86InstInfo{"JMP",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_MODRM | FLAGS_BLOCK_END , 0, nullptr}},
-    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 5), 1, X86InstInfo{"JMPF",  TYPE_INST, FLAGS_SETS_RIP | FLAGS_MODRM | FLAGS_BLOCK_END,                  0, nullptr}},
-    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 6), 1, X86InstInfo{"PUSH",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_MODRM,                                                     0, nullptr}},
-    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                                       0, nullptr}},
+    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 0), 1, X86InstInfo{"INC",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                     0}},
+    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 1), 1, X86InstInfo{"DEC",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,                                                     0}},
+    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 2), 1, X86InstInfo{"CALL",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_MODRM | FLAGS_BLOCK_END | FLAGS_CALL , 0}},
+    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 3), 1, X86InstInfo{"CALLF", TYPE_INST, FLAGS_SETS_RIP | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY | FLAGS_BLOCK_END,                  0}},
+    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 4), 1, X86InstInfo{"JMP",   TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SETS_RIP | FLAGS_MODRM | FLAGS_BLOCK_END , 0}},
+    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 5), 1, X86InstInfo{"JMPF",  TYPE_INST, FLAGS_SETS_RIP | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY | FLAGS_BLOCK_END,                  0}},
+    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 6), 1, X86InstInfo{"PUSH",  TYPE_INST, GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_DEBUG_MEM_ACCESS | FLAGS_MODRM,                                                     0}},
+    {OPD(TYPE_GROUP_5, OpToIndex(0xFF), 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                                       0}},
 
     // GROUP 11
-    {OPD(TYPE_GROUP_11, OpToIndex(0xC6), 0), 1, X86InstInfo{"MOV",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST  | FLAGS_SRC_SEXT,                   1, nullptr}},
-    {OPD(TYPE_GROUP_11, OpToIndex(0xC6), 1), 5, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,                                                       0, nullptr}},
-    {OPD(TYPE_GROUP_11, OpToIndex(0xC6), 7), 1, X86InstInfo{"XABORT", TYPE_INST, FLAGS_MODRM,                                                       1, nullptr}},
-    {OPD(TYPE_GROUP_11, OpToIndex(0xC7), 0), 1, X86InstInfo{"MOV",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2,                                   4, nullptr}},
-    {OPD(TYPE_GROUP_11, OpToIndex(0xC7), 1), 5, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,                                                       0, nullptr}},
-    {OPD(TYPE_GROUP_11, OpToIndex(0xC7), 7), 1, X86InstInfo{"XBEGIN", TYPE_INST, FLAGS_MODRM | FLAGS_SRC_SEXT | FLAGS_SETS_RIP | FLAGS_DISPLACE_SIZE_DIV_2,                                                       4, nullptr}},
+    {OPD(TYPE_GROUP_11, OpToIndex(0xC6), 0), 1, X86InstInfo{"MOV",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST  | FLAGS_SRC_SEXT,                   1}},
+    {OPD(TYPE_GROUP_11, OpToIndex(0xC6), 1), 5, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,                                                       0}},
+    {OPD(TYPE_GROUP_11, OpToIndex(0xC6), 7), 1, X86InstInfo{"XABORT", TYPE_INST, FLAGS_MODRM,                                                       1}},
+    {OPD(TYPE_GROUP_11, OpToIndex(0xC7), 0), 1, X86InstInfo{"MOV",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SRC_SEXT | FLAGS_DISPLACE_SIZE_DIV_2,                                   4}},
+    {OPD(TYPE_GROUP_11, OpToIndex(0xC7), 1), 5, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,                                                       0}},
+    {OPD(TYPE_GROUP_11, OpToIndex(0xC7), 7), 1, X86InstInfo{"XBEGIN", TYPE_INST, FLAGS_MODRM | FLAGS_SRC_SEXT | FLAGS_SETS_RIP | FLAGS_DISPLACE_SIZE_DIV_2,                                                       4}},
   };
 
-  GenerateTable(&Table.at(0), PrimaryGroupOpTable, std::size(PrimaryGroupOpTable));
+  GenerateTable(Table.data(), PrimaryGroupOpTable, std::size(PrimaryGroupOpTable));
 
   IR::InstallToTable(Table, IR::OpDispatch_PrimaryGroupTables);
   return Table;
 }();
 
-void InitializePrimaryGroupTables(Context::OperatingMode Mode) {
-  const U16U8InfoStruct PrimaryGroupOpTable_64[] = {
-    // Invalid in 64bit mode
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 0), 8, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,                                                        0, nullptr}},
-  };
-
-  const U16U8InfoStruct PrimaryGroupOpTable_32[] = {
-    // Duplicates the 0x80 opcode group
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 0), 1, X86InstInfo{"ADD",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 1), 1, X86InstInfo{"OR",   TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 2), 1, X86InstInfo{"ADC",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 3), 1, X86InstInfo{"SBB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 4), 1, X86InstInfo{"AND",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 5), 1, X86InstInfo{"SUB",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 6), 1, X86InstInfo{"XOR",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-    {OPD(TYPE_GROUP_1, OpToIndex(0x82), 7), 1, X86InstInfo{"CMP",  TYPE_INST, GenFlagsSameSize(SIZE_8BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST,                                      1, nullptr}},
-  };
-
-#undef OPD
-
-  if (Mode == Context::MODE_64BIT) {
-    GenerateTable(&PrimaryInstGroupOps.at(0), PrimaryGroupOpTable_64, std::size(PrimaryGroupOpTable_64));
-  }
-  else {
-    GenerateTable(&PrimaryInstGroupOps.at(0), PrimaryGroupOpTable_32, std::size(PrimaryGroupOpTable_32));
-  }
-}
-
 }

FEXCore/Source/Interface/Core/X86Tables/SecondaryGroupTables.cpp

@@ -13,14 +13,45 @@ $end_info$
 
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
-std::array<X86InstInfo, MAX_INST_SECOND_GROUP_TABLE_SIZE> SecondInstGroupOps = []() consteval {
-  std::array<X86InstInfo, MAX_INST_SECOND_GROUP_TABLE_SIZE> Table{};
+constexpr uint16_t PF_NONE = 0;
+constexpr uint16_t PF_F3   = 1;
+constexpr uint16_t PF_66   = 2;
+constexpr uint16_t PF_F2   = 3;
 #define OPD(group, prefix, Reg) (((group - FEXCore::X86Tables::TYPE_GROUP_6) << 5) | (prefix) << 3 | (Reg))
-  constexpr uint16_t PF_NONE = 0;
-  constexpr uint16_t PF_F3   = 1;
-  constexpr uint16_t PF_66   = 2;
-  constexpr uint16_t PF_F2   = 3;
 
+enum SecondGroup_LUT {
+  ENTRY_15_F3_0,
+  ENTRY_15_F3_1,
+  ENTRY_15_F3_2,
+  ENTRY_15_F3_3,
+  ENTRY_MAX,
+};
+
+constexpr std::array<X86InstInfo[2], ENTRY_MAX> SecondGroup_ArchSelect_LUT = {{
+  // ENTRY_15_F3_0
+  {
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"RDFSBASE", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::ReadSegmentReg, IR::OpDispatchBuilder::Segment::FS> } },
+  },
+  // ENTRY_15_F3_1
+  {
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"RDGSBASE", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::ReadSegmentReg, IR::OpDispatchBuilder::Segment::GS> } },
+  },
+  // ENTRY_15_F3_2
+  {
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"WRFSBASE", TYPE_INST, GenFlagsDstSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_REG_ONLY, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::WriteSegmentReg, IR::OpDispatchBuilder::Segment::FS> } },
+  },
+  // ENTRY_15_F3_3
+  {
+    {"", TYPE_INVALID, FLAGS_NONE, 0, { .OpDispatch = nullptr } },
+    {"WRGSBASE", TYPE_INST, GenFlagsDstSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_REG_ONLY, 0, { .OpDispatch = &IR::OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::WriteSegmentReg, IR::OpDispatchBuilder::Segment::GS> } },
+  },
+}};
+
+constexpr auto SecondInstGroupOps = []() consteval {
+  std::array<X86InstInfo, MAX_INST_SECOND_GROUP_TABLE_SIZE> Table{};
   constexpr U16U8InfoStruct SecondaryExtensionOpTable[] = {
     // GROUP 1
     // GROUP 2
@@ -30,115 +61,115 @@ std::array<X86InstInfo, MAX_INST_SECOND_GROUP_TABLE_SIZE> SecondInstGroupOps = [
     // Pulls from other MODRM table
 
     // GROUP 6
-    {OPD(TYPE_GROUP_6, PF_NONE, 0), 1, X86InstInfo{"SLDT",  TYPE_UNDEC, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_NONE, 1), 1, X86InstInfo{"STR",   TYPE_PRIV, FLAGS_MODRM | FLAGS_SF_MOD_DST,  0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_NONE, 2), 1, X86InstInfo{"LLDT",  TYPE_PRIV, FLAGS_NONE,       0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_NONE, 3), 1, X86InstInfo{"LTR",   TYPE_INST, FLAGS_NONE,       0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_NONE, 4), 1, X86InstInfo{"VERR",  TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_NONE, 5), 1, X86InstInfo{"VERW",  TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_NONE, 6), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,    0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_NONE, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,    0, nullptr}},
+    {OPD(TYPE_GROUP_6, PF_NONE, 0), 1, X86InstInfo{"SLDT",  TYPE_UNDEC, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPD(TYPE_GROUP_6, PF_NONE, 1), 1, X86InstInfo{"STR",   TYPE_PRIV, FLAGS_MODRM | FLAGS_SF_MOD_DST,  0}},
+    {OPD(TYPE_GROUP_6, PF_NONE, 2), 1, X86InstInfo{"LLDT",  TYPE_PRIV, FLAGS_NONE,       0}},
+    {OPD(TYPE_GROUP_6, PF_NONE, 3), 1, X86InstInfo{"LTR",   TYPE_INST, FLAGS_NONE,       0}},
+    {OPD(TYPE_GROUP_6, PF_NONE, 4), 1, X86InstInfo{"VERR",  TYPE_UNDEC, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_6, PF_NONE, 5), 1, X86InstInfo{"VERW",  TYPE_UNDEC, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_6, PF_NONE, 6), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,    0}},
+    {OPD(TYPE_GROUP_6, PF_NONE, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,    0}},
 
-    {OPD(TYPE_GROUP_6, PF_F3, 0), 1, X86InstInfo{"SLDT",    TYPE_UNDEC, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F3, 1), 1, X86InstInfo{"STR",     TYPE_PRIV, FLAGS_MODRM | FLAGS_SF_MOD_DST,  0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F3, 2), 1, X86InstInfo{"LLDT",    TYPE_PRIV, FLAGS_NONE,       0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F3, 3), 1, X86InstInfo{"LTR",     TYPE_INST, FLAGS_NONE,       0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F3, 4), 1, X86InstInfo{"VERR",    TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F3, 5), 1, X86InstInfo{"VERW",    TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F3, 6), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F3, 7), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0, nullptr}},
+    {OPD(TYPE_GROUP_6, PF_F3, 0), 1, X86InstInfo{"SLDT",    TYPE_UNDEC, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPD(TYPE_GROUP_6, PF_F3, 1), 1, X86InstInfo{"STR",     TYPE_PRIV, FLAGS_MODRM | FLAGS_SF_MOD_DST,  0}},
+    {OPD(TYPE_GROUP_6, PF_F3, 2), 1, X86InstInfo{"LLDT",    TYPE_PRIV, FLAGS_NONE,       0}},
+    {OPD(TYPE_GROUP_6, PF_F3, 3), 1, X86InstInfo{"LTR",     TYPE_INST, FLAGS_NONE,       0}},
+    {OPD(TYPE_GROUP_6, PF_F3, 4), 1, X86InstInfo{"VERR",    TYPE_UNDEC, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_6, PF_F3, 5), 1, X86InstInfo{"VERW",    TYPE_UNDEC, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_6, PF_F3, 6), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0}},
+    {OPD(TYPE_GROUP_6, PF_F3, 7), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0}},
 
-    {OPD(TYPE_GROUP_6, PF_66, 0), 1, X86InstInfo{"SLDT",    TYPE_UNDEC, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_66, 1), 1, X86InstInfo{"STR",     TYPE_PRIV, FLAGS_MODRM | FLAGS_SF_MOD_DST,  0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_66, 2), 1, X86InstInfo{"LLDT",    TYPE_PRIV, FLAGS_NONE,       0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_66, 3), 1, X86InstInfo{"LTR",     TYPE_INST, FLAGS_NONE,       0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_66, 4), 1, X86InstInfo{"VERR",    TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_66, 5), 1, X86InstInfo{"VERW",    TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_66, 6), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_66, 7), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0, nullptr}},
+    {OPD(TYPE_GROUP_6, PF_66, 0), 1, X86InstInfo{"SLDT",    TYPE_UNDEC, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPD(TYPE_GROUP_6, PF_66, 1), 1, X86InstInfo{"STR",     TYPE_PRIV, FLAGS_MODRM | FLAGS_SF_MOD_DST,  0}},
+    {OPD(TYPE_GROUP_6, PF_66, 2), 1, X86InstInfo{"LLDT",    TYPE_PRIV, FLAGS_NONE,       0}},
+    {OPD(TYPE_GROUP_6, PF_66, 3), 1, X86InstInfo{"LTR",     TYPE_INST, FLAGS_NONE,       0}},
+    {OPD(TYPE_GROUP_6, PF_66, 4), 1, X86InstInfo{"VERR",    TYPE_UNDEC, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_6, PF_66, 5), 1, X86InstInfo{"VERW",    TYPE_UNDEC, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_6, PF_66, 6), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0}},
+    {OPD(TYPE_GROUP_6, PF_66, 7), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0}},
 
-    {OPD(TYPE_GROUP_6, PF_F2, 0), 1, X86InstInfo{"SLDT",    TYPE_UNDEC, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F2, 1), 1, X86InstInfo{"STR",     TYPE_PRIV, FLAGS_MODRM | FLAGS_SF_MOD_DST,  0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F2, 2), 1, X86InstInfo{"LLDT",    TYPE_PRIV, FLAGS_NONE,       0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F2, 3), 1, X86InstInfo{"LTR",     TYPE_INST, FLAGS_NONE,       0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F2, 4), 1, X86InstInfo{"VERR",    TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F2, 5), 1, X86InstInfo{"VERW",    TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F2, 6), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0, nullptr}},
-    {OPD(TYPE_GROUP_6, PF_F2, 7), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0, nullptr}},
+    {OPD(TYPE_GROUP_6, PF_F2, 0), 1, X86InstInfo{"SLDT",    TYPE_UNDEC, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPD(TYPE_GROUP_6, PF_F2, 1), 1, X86InstInfo{"STR",     TYPE_PRIV, FLAGS_MODRM | FLAGS_SF_MOD_DST,  0}},
+    {OPD(TYPE_GROUP_6, PF_F2, 2), 1, X86InstInfo{"LLDT",    TYPE_PRIV, FLAGS_NONE,       0}},
+    {OPD(TYPE_GROUP_6, PF_F2, 3), 1, X86InstInfo{"LTR",     TYPE_INST, FLAGS_NONE,       0}},
+    {OPD(TYPE_GROUP_6, PF_F2, 4), 1, X86InstInfo{"VERR",    TYPE_UNDEC, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_6, PF_F2, 5), 1, X86InstInfo{"VERW",    TYPE_UNDEC, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_6, PF_F2, 6), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0}},
+    {OPD(TYPE_GROUP_6, PF_F2, 7), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE,    0}},
 
     // GROUP 7
-    {OPD(TYPE_GROUP_7, PF_NONE, 0), 1, X86InstInfo{"SGDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,         0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_NONE, 1), 1, X86InstInfo{"SIDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,         0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_NONE, 2), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE, 0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_NONE, 3), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE, 0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_NONE, 4), 1, X86InstInfo{"SMSW", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,          0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_NONE, 5), 1, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,            0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_NONE, 6), 1, X86InstInfo{"LMSW", TYPE_INST, FLAGS_MODRM,          0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_NONE, 7), 1, X86InstInfo{"INVLPG", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
+    {OPD(TYPE_GROUP_7, PF_NONE, 0), 1, X86InstInfo{"SGDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,         0}},
+    {OPD(TYPE_GROUP_7, PF_NONE, 1), 1, X86InstInfo{"SIDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,         0}},
+    {OPD(TYPE_GROUP_7, PF_NONE, 2), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE, 0}},
+    {OPD(TYPE_GROUP_7, PF_NONE, 3), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE, 0}},
+    {OPD(TYPE_GROUP_7, PF_NONE, 4), 1, X86InstInfo{"SMSW", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,          0}},
+    {OPD(TYPE_GROUP_7, PF_NONE, 5), 1, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,            0}},
+    {OPD(TYPE_GROUP_7, PF_NONE, 6), 1, X86InstInfo{"LMSW", TYPE_INST, FLAGS_MODRM,          0}},
+    {OPD(TYPE_GROUP_7, PF_NONE, 7), 1, X86InstInfo{"INVLPG", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
 
-    {OPD(TYPE_GROUP_7, PF_F3, 0), 1, X86InstInfo{"SGDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F3, 1), 1, X86InstInfo{"SIDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F3, 2), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F3, 3), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F3, 4), 1, X86InstInfo{"SMSW", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,            0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F3, 5), 1, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,              0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F3, 6), 1, X86InstInfo{"LMSW", TYPE_INST, FLAGS_MODRM,            0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F3, 7), 1, X86InstInfo{"INVLPG", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,   0, nullptr}},
+    {OPD(TYPE_GROUP_7, PF_F3, 0), 1, X86InstInfo{"SGDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0}},
+    {OPD(TYPE_GROUP_7, PF_F3, 1), 1, X86InstInfo{"SIDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0}},
+    {OPD(TYPE_GROUP_7, PF_F3, 2), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_7, PF_F3, 3), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_7, PF_F3, 4), 1, X86InstInfo{"SMSW", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,            0}},
+    {OPD(TYPE_GROUP_7, PF_F3, 5), 1, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,              0}},
+    {OPD(TYPE_GROUP_7, PF_F3, 6), 1, X86InstInfo{"LMSW", TYPE_INST, FLAGS_MODRM,            0}},
+    {OPD(TYPE_GROUP_7, PF_F3, 7), 1, X86InstInfo{"INVLPG", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,   0}},
 
-    {OPD(TYPE_GROUP_7, PF_66, 0), 1, X86InstInfo{"SGDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_66, 1), 1, X86InstInfo{"SIDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_66, 2), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_66, 3), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_66, 4), 1, X86InstInfo{"SMSW", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,            0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_66, 5), 1, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,              0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_66, 6), 1, X86InstInfo{"LMSW", TYPE_INST, FLAGS_MODRM,            0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_66, 7), 1, X86InstInfo{"INVLPG", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,   0, nullptr}},
+    {OPD(TYPE_GROUP_7, PF_66, 0), 1, X86InstInfo{"SGDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0}},
+    {OPD(TYPE_GROUP_7, PF_66, 1), 1, X86InstInfo{"SIDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0}},
+    {OPD(TYPE_GROUP_7, PF_66, 2), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_7, PF_66, 3), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_7, PF_66, 4), 1, X86InstInfo{"SMSW", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,            0}},
+    {OPD(TYPE_GROUP_7, PF_66, 5), 1, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,              0}},
+    {OPD(TYPE_GROUP_7, PF_66, 6), 1, X86InstInfo{"LMSW", TYPE_INST, FLAGS_MODRM,            0}},
+    {OPD(TYPE_GROUP_7, PF_66, 7), 1, X86InstInfo{"INVLPG", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,   0}},
 
-    {OPD(TYPE_GROUP_7, PF_F2, 0), 1, X86InstInfo{"SGDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F2, 1), 1, X86InstInfo{"SIDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F2, 2), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F2, 3), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F2, 4), 1, X86InstInfo{"SMSW", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,            0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F2, 5), 1, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,              0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F2, 6), 1, X86InstInfo{"LMSW", TYPE_INST, FLAGS_MODRM,            0, nullptr}},
-    {OPD(TYPE_GROUP_7, PF_F2, 7), 1, X86InstInfo{"INVLPG", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,   0, nullptr}},
+    {OPD(TYPE_GROUP_7, PF_F2, 0), 1, X86InstInfo{"SGDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0}},
+    {OPD(TYPE_GROUP_7, PF_F2, 1), 1, X86InstInfo{"SIDT", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,           0}},
+    {OPD(TYPE_GROUP_7, PF_F2, 2), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_7, PF_F2, 3), 1, X86InstInfo{"",     TYPE_SECOND_GROUP_MODRM, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_7, PF_F2, 4), 1, X86InstInfo{"SMSW", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,            0}},
+    {OPD(TYPE_GROUP_7, PF_F2, 5), 1, X86InstInfo{"",     TYPE_INVALID, FLAGS_NONE,              0}},
+    {OPD(TYPE_GROUP_7, PF_F2, 6), 1, X86InstInfo{"LMSW", TYPE_INST, FLAGS_MODRM,            0}},
+    {OPD(TYPE_GROUP_7, PF_F2, 7), 1, X86InstInfo{"INVLPG", TYPE_SECOND_GROUP_MODRM, FLAGS_MODRM | FLAGS_SF_MOD_DST,   0}},
 
     // GROUP 8
-    {OPD(TYPE_GROUP_8, PF_NONE, 0), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_NONE, 1), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_NONE, 2), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_NONE, 3), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_NONE, 4), 1, X86InstInfo{"BT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_NONE, 5), 1, X86InstInfo{"BTS", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_NONE, 6), 1, X86InstInfo{"BTR", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_NONE, 7), 1, X86InstInfo{"BTC", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST, 1, nullptr}},
+    {OPD(TYPE_GROUP_8, PF_NONE, 0), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_8, PF_NONE, 1), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_8, PF_NONE, 2), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_8, PF_NONE, 3), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_8, PF_NONE, 4), 1, X86InstInfo{"BT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST, 1}},
+    {OPD(TYPE_GROUP_8, PF_NONE, 5), 1, X86InstInfo{"BTS", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST, 1}},
+    {OPD(TYPE_GROUP_8, PF_NONE, 6), 1, X86InstInfo{"BTR", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST, 1}},
+    {OPD(TYPE_GROUP_8, PF_NONE, 7), 1, X86InstInfo{"BTC", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST, 1}},
 
-    {OPD(TYPE_GROUP_8, PF_F3, 0), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F3, 1), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F3, 2), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F3, 3), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F3, 4), 1, X86InstInfo{"BT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F3, 5), 1, X86InstInfo{"BTS", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F3, 6), 1, X86InstInfo{"BTR", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F3, 7), 1, X86InstInfo{"BTC", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
+    {OPD(TYPE_GROUP_8, PF_F3, 0), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_F3, 1), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_F3, 2), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_F3, 3), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_F3, 4), 1, X86InstInfo{"BT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_F3, 5), 1, X86InstInfo{"BTS", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_F3, 6), 1, X86InstInfo{"BTR", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_F3, 7), 1, X86InstInfo{"BTC", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
 
-    {OPD(TYPE_GROUP_8, PF_66, 0), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_66, 1), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_66, 2), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_66, 3), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_66, 4), 1, X86InstInfo{"BT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_66, 5), 1, X86InstInfo{"BTS", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_66, 6), 1, X86InstInfo{"BTR", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_66, 7), 1, X86InstInfo{"BTC", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
+    {OPD(TYPE_GROUP_8, PF_66, 0), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_66, 1), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_66, 2), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_66, 3), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_66, 4), 1, X86InstInfo{"BT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_66, 5), 1, X86InstInfo{"BTS", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_66, 6), 1, X86InstInfo{"BTR", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_66, 7), 1, X86InstInfo{"BTC", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
 
-    {OPD(TYPE_GROUP_8, PF_F2, 0), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F2, 1), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F2, 2), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F2, 3), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F2, 4), 1, X86InstInfo{"BT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F2, 5), 1, X86InstInfo{"BTS", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F2, 6), 1, X86InstInfo{"BTR", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
-    {OPD(TYPE_GROUP_8, PF_F2, 7), 1, X86InstInfo{"BTC", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1, nullptr}},
+    {OPD(TYPE_GROUP_8, PF_F2, 0), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_F2, 1), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_F2, 2), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_F2, 3), 1, X86InstInfo{"",    TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_8, PF_F2, 4), 1, X86InstInfo{"BT",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_F2, 5), 1, X86InstInfo{"BTS", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_F2, 6), 1, X86InstInfo{"BTR", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
+    {OPD(TYPE_GROUP_8, PF_F2, 7), 1, X86InstInfo{"BTC", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,   1}},
 
     // GROUP 9
 
@@ -147,302 +178,302 @@ std::array<X86InstInfo, MAX_INST_SECOND_GROUP_TABLE_SIZE> SecondInstGroupOps = [
     // RDRAND/RDSEED only work with no prefix (Other than 66h)
     // CMPXCHG8B/16B works with all prefixes
     // Tooling fails to decode CMPXCHG with prefix
-    {OPD(TYPE_GROUP_9, PF_NONE, 0), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_NONE, 1), 1, X86InstInfo{"CMPXCHG8B/16B", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_NONE, 2), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_NONE, 3), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_NONE, 4), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_NONE, 5), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_NONE, 6), 1, X86InstInfo{"RDRAND",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_NONE, 7), 1, X86InstInfo{"RDSEED",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0, nullptr}},
+    {OPD(TYPE_GROUP_9, PF_NONE, 0), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_9, PF_NONE, 1), 1, X86InstInfo{"CMPXCHG8B/16B", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_9, PF_NONE, 2), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_9, PF_NONE, 3), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_9, PF_NONE, 4), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_9, PF_NONE, 5), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,   0}},
+    {OPD(TYPE_GROUP_9, PF_NONE, 6), 1, X86InstInfo{"RDRAND",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0}},
+    {OPD(TYPE_GROUP_9, PF_NONE, 7), 1, X86InstInfo{"RDSEED",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0}},
 
-    {OPD(TYPE_GROUP_9, PF_F3, 0), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F3, 1), 1, X86InstInfo{"CMPXCHG8B/16B", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F3, 2), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F3, 3), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F3, 4), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F3, 5), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F3, 6), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F3, 7), 1, X86InstInfo{"RDPID",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0, nullptr}},
+    {OPD(TYPE_GROUP_9, PF_F3, 0), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F3, 1), 1, X86InstInfo{"CMPXCHG8B/16B", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_9, PF_F3, 2), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F3, 3), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F3, 4), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F3, 5), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F3, 6), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F3, 7), 1, X86InstInfo{"RDPID",      TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0}},
 
-    {OPD(TYPE_GROUP_9, PF_66, 0), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_66, 1), 1, X86InstInfo{"CMPXCHG8B/16B", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_66, 2), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_66, 3), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_66, 4), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_66, 5), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_66, 6), 1, X86InstInfo{"RDRAND",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_66, 7), 1, X86InstInfo{"RDSEED",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0, nullptr}},
+    {OPD(TYPE_GROUP_9, PF_66, 0), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_66, 1), 1, X86InstInfo{"CMPXCHG8B/16B", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_9, PF_66, 2), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_66, 3), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_66, 4), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_66, 5), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_66, 6), 1, X86InstInfo{"RDRAND",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0}},
+    {OPD(TYPE_GROUP_9, PF_66, 7), 1, X86InstInfo{"RDSEED",     TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY, 0}},
 
-    {OPD(TYPE_GROUP_9, PF_F2, 0), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F2, 1), 1, X86InstInfo{"CMPXCHG8B/16B", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F2, 2), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F2, 3), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F2, 4), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F2, 5), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F2, 6), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
-    {OPD(TYPE_GROUP_9, PF_F2, 7), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0, nullptr}},
+    {OPD(TYPE_GROUP_9, PF_F2, 0), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F2, 1), 1, X86InstInfo{"CMPXCHG8B/16B", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_9, PF_F2, 2), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F2, 3), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F2, 4), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F2, 5), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F2, 6), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
+    {OPD(TYPE_GROUP_9, PF_F2, 7), 1, X86InstInfo{"",           TYPE_INVALID, FLAGS_NONE,     0}},
 
     // GROUP 10
-    {OPD(TYPE_GROUP_10, PF_NONE, 0), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_NONE, 1), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_NONE, 2), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_NONE, 3), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_NONE, 4), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_NONE, 5), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_NONE, 6), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_NONE, 7), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0, nullptr}},
+    {OPD(TYPE_GROUP_10, PF_NONE, 0), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {OPD(TYPE_GROUP_10, PF_NONE, 1), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {OPD(TYPE_GROUP_10, PF_NONE, 2), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {OPD(TYPE_GROUP_10, PF_NONE, 3), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {OPD(TYPE_GROUP_10, PF_NONE, 4), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {OPD(TYPE_GROUP_10, PF_NONE, 5), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {OPD(TYPE_GROUP_10, PF_NONE, 6), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0}},
+    {OPD(TYPE_GROUP_10, PF_NONE, 7), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END, 0}},
 
-    {OPD(TYPE_GROUP_10, PF_F3, 0), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F3, 1), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F3, 2), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F3, 3), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F3, 4), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F3, 5), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F3, 6), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F3, 7), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
+    {OPD(TYPE_GROUP_10, PF_F3, 0), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F3, 1), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F3, 2), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F3, 3), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F3, 4), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F3, 5), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F3, 6), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F3, 7), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
 
-    {OPD(TYPE_GROUP_10, PF_66, 0), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_66, 1), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_66, 2), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_66, 3), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_66, 4), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_66, 5), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_66, 6), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_66, 7), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
+    {OPD(TYPE_GROUP_10, PF_66, 0), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_66, 1), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_66, 2), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_66, 3), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_66, 4), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_66, 5), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_66, 6), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_66, 7), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
 
-    {OPD(TYPE_GROUP_10, PF_F2, 0), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F2, 1), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F2, 2), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F2, 3), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F2, 4), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F2, 5), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F2, 6), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
-    {OPD(TYPE_GROUP_10, PF_F2, 7), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0, nullptr}},
+    {OPD(TYPE_GROUP_10, PF_F2, 0), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F2, 1), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F2, 2), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F2, 3), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F2, 4), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F2, 5), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F2, 6), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
+    {OPD(TYPE_GROUP_10, PF_F2, 7), 1, X86InstInfo{"UD1", TYPE_INST, FLAGS_BLOCK_END,   0}},
 
     // GROUP 12
-    {OPD(TYPE_GROUP_12, PF_NONE, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_NONE, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_NONE, 2), 1, X86InstInfo{"PSRLW", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_NONE, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_NONE, 4), 1, X86InstInfo{"PSRAW", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_NONE, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_NONE, 6), 1, X86InstInfo{"PSLLW", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_NONE, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
+    {OPD(TYPE_GROUP_12, PF_NONE, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_12, PF_NONE, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_12, PF_NONE, 2), 1, X86InstInfo{"PSRLW", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+    {OPD(TYPE_GROUP_12, PF_NONE, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_12, PF_NONE, 4), 1, X86InstInfo{"PSRAW", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+    {OPD(TYPE_GROUP_12, PF_NONE, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_12, PF_NONE, 6), 1, X86InstInfo{"PSLLW", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+    {OPD(TYPE_GROUP_12, PF_NONE, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
 
-    {OPD(TYPE_GROUP_12, PF_66, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_66, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_66, 2), 1, X86InstInfo{"PSRLW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_66, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_66, 4), 1, X86InstInfo{"PSRAW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_66, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_66, 6), 1, X86InstInfo{"PSLLW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_66, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
+    {OPD(TYPE_GROUP_12, PF_66, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_12, PF_66, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_12, PF_66, 2), 1, X86InstInfo{"PSRLW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_12, PF_66, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_12, PF_66, 4), 1, X86InstInfo{"PSRAW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_12, PF_66, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_12, PF_66, 6), 1, X86InstInfo{"PSLLW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_12, PF_66, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
 
-    {OPD(TYPE_GROUP_12, PF_F3, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F3, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F3, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F3, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F3, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F3, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
+    {OPD(TYPE_GROUP_12, PF_F3, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F3, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F3, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F3, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F3, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F3, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
 
-    {OPD(TYPE_GROUP_12, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F2, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_12, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
+    {OPD(TYPE_GROUP_12, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F2, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_12, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
 
     // GROUP 13
-    {OPD(TYPE_GROUP_13, PF_NONE, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_NONE, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_NONE, 2), 1, X86InstInfo{"PSRLD", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_NONE, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_NONE, 4), 1, X86InstInfo{"PSRAD", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_NONE, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_NONE, 6), 1, X86InstInfo{"PSLLD", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_NONE, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
+    {OPD(TYPE_GROUP_13, PF_NONE, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_13, PF_NONE, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_13, PF_NONE, 2), 1, X86InstInfo{"PSRLD", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+    {OPD(TYPE_GROUP_13, PF_NONE, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_13, PF_NONE, 4), 1, X86InstInfo{"PSRAD", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+    {OPD(TYPE_GROUP_13, PF_NONE, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_13, PF_NONE, 6), 1, X86InstInfo{"PSLLD", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+    {OPD(TYPE_GROUP_13, PF_NONE, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
 
-    {OPD(TYPE_GROUP_13, PF_66, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_66, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_66, 2), 1, X86InstInfo{"PSRLD", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_66, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_66, 4), 1, X86InstInfo{"PSRAD", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_66, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_66, 6), 1, X86InstInfo{"PSLLD", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_66, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
+    {OPD(TYPE_GROUP_13, PF_66, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_13, PF_66, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_13, PF_66, 2), 1, X86InstInfo{"PSRLD", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_13, PF_66, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_13, PF_66, 4), 1, X86InstInfo{"PSRAD", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_13, PF_66, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_13, PF_66, 6), 1, X86InstInfo{"PSLLD", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_13, PF_66, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
 
-    {OPD(TYPE_GROUP_13, PF_F3, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F3, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F3, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F3, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F3, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F3, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
+    {OPD(TYPE_GROUP_13, PF_F3, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F3, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F3, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F3, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F3, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F3, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
 
-    {OPD(TYPE_GROUP_13, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F2, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_13, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
+    {OPD(TYPE_GROUP_13, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F2, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_13, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
 
     // GROUP 14
-    {OPD(TYPE_GROUP_14, PF_NONE, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_NONE, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_NONE, 2), 1, X86InstInfo{"PSRLQ", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_NONE, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_NONE, 4), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_NONE, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_NONE, 6), 1, X86InstInfo{"PSLLQ", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_NONE, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0, nullptr}},
+    {OPD(TYPE_GROUP_14, PF_NONE, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_14, PF_NONE, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_14, PF_NONE, 2), 1, X86InstInfo{"PSRLQ", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+    {OPD(TYPE_GROUP_14, PF_NONE, 3), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_14, PF_NONE, 4), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_14, PF_NONE, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
+    {OPD(TYPE_GROUP_14, PF_NONE, 6), 1, X86InstInfo{"PSLLQ", TYPE_INST, GenFlagsSameSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS | FLAGS_SF_MMX, 1}},
+    {OPD(TYPE_GROUP_14, PF_NONE, 7), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                      0}},
 
-    {OPD(TYPE_GROUP_14, PF_66, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_66, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_66, 2), 1, X86InstInfo{"PSRLQ", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_66, 3), 1, X86InstInfo{"PSRLDQ",TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_66, 4), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_66, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_66, 6), 1, X86InstInfo{"PSLLQ", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_66, 7), 1, X86InstInfo{"PSLLDQ",TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1, nullptr}},
+    {OPD(TYPE_GROUP_14, PF_66, 0), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_14, PF_66, 1), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_14, PF_66, 2), 1, X86InstInfo{"PSRLQ", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_14, PF_66, 3), 1, X86InstInfo{"PSRLDQ",TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_14, PF_66, 4), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_14, PF_66, 5), 1, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE,                                        0}},
+    {OPD(TYPE_GROUP_14, PF_66, 6), 1, X86InstInfo{"PSLLQ", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
+    {OPD(TYPE_GROUP_14, PF_66, 7), 1, X86InstInfo{"PSLLDQ",TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS,  1}},
 
-    {OPD(TYPE_GROUP_14, PF_F3, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F3, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F3, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F3, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F3, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F3, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
+    {OPD(TYPE_GROUP_14, PF_F3, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F3, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F3, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F3, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F3, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F3, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
 
-    {OPD(TYPE_GROUP_14, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F2, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
-    {OPD(TYPE_GROUP_14, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0, nullptr}},
+    {OPD(TYPE_GROUP_14, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F2, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
+    {OPD(TYPE_GROUP_14, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                             0}},
 
     // GROUP 15
-    {OPD(TYPE_GROUP_15, PF_NONE, 0), 1, X86InstInfo{"FXSAVE",          TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,       0, nullptr}}, // MMX/x87
-    {OPD(TYPE_GROUP_15, PF_NONE, 1), 1, X86InstInfo{"FXRSTOR",         TYPE_INST, FLAGS_MODRM,       0, nullptr}}, // MMX/x87
-    {OPD(TYPE_GROUP_15, PF_NONE, 2), 1, X86InstInfo{"LDMXCSR",         TYPE_INST, GenFlagsSameSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_NONE, 3), 1, X86InstInfo{"STMXCSR",         TYPE_INST, GenFlagsSameSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_NONE, 4), 1, X86InstInfo{"XSAVE",           TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_NONE, 5), 1, X86InstInfo{"LFENCE/XRSTOR",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_NONE, 6), 1, X86InstInfo{"MFENCE/XSAVEOPT", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_NONE, 7), 1, X86InstInfo{"SFENCE/CLFLUSH",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0, nullptr}},
+    {OPD(TYPE_GROUP_15, PF_NONE, 0), 1, X86InstInfo{"FXSAVE",          TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,       0}}, // MMX/x87
+    {OPD(TYPE_GROUP_15, PF_NONE, 1), 1, X86InstInfo{"FXRSTOR",         TYPE_INST, FLAGS_MODRM,       0}}, // MMX/x87
+    {OPD(TYPE_GROUP_15, PF_NONE, 2), 1, X86InstInfo{"LDMXCSR",         TYPE_INST, GenFlagsSameSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_15, PF_NONE, 3), 1, X86InstInfo{"STMXCSR",         TYPE_INST, GenFlagsSameSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_15, PF_NONE, 4), 1, X86InstInfo{"XSAVE",           TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0}},
+    {OPD(TYPE_GROUP_15, PF_NONE, 5), 1, X86InstInfo{"LFENCE/XRSTOR",   TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0}},
+    {OPD(TYPE_GROUP_15, PF_NONE, 6), 1, X86InstInfo{"MFENCE/XSAVEOPT", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0}},
+    {OPD(TYPE_GROUP_15, PF_NONE, 7), 1, X86InstInfo{"SFENCE/CLFLUSH",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0}},
 
-    {OPD(TYPE_GROUP_15, PF_F3, 0), 1, X86InstInfo{"RDFSBASE", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY,          0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F3, 1), 1, X86InstInfo{"RDGSBASE", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY,          0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F3, 2), 1, X86InstInfo{"WRFSBASE", TYPE_INST, GenFlagsDstSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_REG_ONLY,          0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F3, 3), 1, X86InstInfo{"WRGSBASE", TYPE_INST, GenFlagsDstSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_REG_ONLY,          0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F3, 5), 1, X86InstInfo{"INCSSPQ", TYPE_INST, FLAGS_MODRM,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F3, 6), 1, X86InstInfo{"UMONITOR/CLRSSBSY", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,       0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
+    {OPD(TYPE_GROUP_15, PF_F3, 0), 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = SecondGroup_ArchSelect_LUT[ENTRY_15_F3_0] }}},
+    {OPD(TYPE_GROUP_15, PF_F3, 1), 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = SecondGroup_ArchSelect_LUT[ENTRY_15_F3_1] }}},
+    {OPD(TYPE_GROUP_15, PF_F3, 2), 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = SecondGroup_ArchSelect_LUT[ENTRY_15_F3_2] }}},
+    {OPD(TYPE_GROUP_15, PF_F3, 3), 1, X86InstInfo{"", TYPE_ARCH_DISPATCHER, FLAGS_NONE, 0, { .Indirect = SecondGroup_ArchSelect_LUT[ENTRY_15_F3_3] }}},
+    {OPD(TYPE_GROUP_15, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_F3, 5), 1, X86InstInfo{"INCSSPQ", TYPE_INST, FLAGS_MODRM,                    0}},
+    {OPD(TYPE_GROUP_15, PF_F3, 6), 1, X86InstInfo{"UMONITOR/CLRSSBSY", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,       0}},
+    {OPD(TYPE_GROUP_15, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
 
-    {OPD(TYPE_GROUP_15, PF_66, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_66, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_66, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_66, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_66, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_66, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_66, 6), 1, X86InstInfo{"CLWB", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_66, 7), 1, X86InstInfo{"CLFLUSHOPT", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0, nullptr}},
+    {OPD(TYPE_GROUP_15, PF_66, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_66, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_66, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_66, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_66, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_66, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_66, 6), 1, X86InstInfo{"CLWB", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0}},
+    {OPD(TYPE_GROUP_15, PF_66, 7), 1, X86InstInfo{"CLFLUSHOPT", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,      0}},
 
-    {OPD(TYPE_GROUP_15, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F2, 6), 1, X86InstInfo{"UMWAIT", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,       0, nullptr}},
-    {OPD(TYPE_GROUP_15, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0, nullptr}},
+    {OPD(TYPE_GROUP_15, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
+    {OPD(TYPE_GROUP_15, PF_F2, 6), 1, X86InstInfo{"UMWAIT", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_DST,       0}},
+    {OPD(TYPE_GROUP_15, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                    0}},
 
     // GROUP 16
     // AMD documentation claims again that this entire group is n/a to prefix
     // Tooling once again fails to disassemble oens with the prefix. Disable until proven otherwise
-    {OPD(TYPE_GROUP_16, PF_NONE, 0), 1, X86InstInfo{"PREFETCH NTA", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_NONE, 1), 1, X86InstInfo{"PREFETCH T0",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_NONE, 2), 1, X86InstInfo{"PREFETCH T1",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_NONE, 3), 1, X86InstInfo{"PREFETCH T2",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_NONE, 4), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_NONE, 5), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_NONE, 6), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_NONE, 7), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM, 0, nullptr}},
+    {OPD(TYPE_GROUP_16, PF_NONE, 0), 1, X86InstInfo{"PREFETCH NTA", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_16, PF_NONE, 1), 1, X86InstInfo{"PREFETCH T0",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_16, PF_NONE, 2), 1, X86InstInfo{"PREFETCH T1",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_16, PF_NONE, 3), 1, X86InstInfo{"PREFETCH T2",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_16, PF_NONE, 4), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_16, PF_NONE, 5), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_16, PF_NONE, 6), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM, 0}},
+    {OPD(TYPE_GROUP_16, PF_NONE, 7), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM, 0}},
 
-    {OPD(TYPE_GROUP_16, PF_F3, 0), 1, X86InstInfo{"PREFETCH NTA", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F3, 1), 1, X86InstInfo{"PREFETCH T0",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F3, 2), 1, X86InstInfo{"PREFETCH T1",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F3, 3), 1, X86InstInfo{"PREFETCH T2",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F3, 4), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F3, 5), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F3, 6), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F3, 7), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
+    {OPD(TYPE_GROUP_16, PF_F3, 0), 1, X86InstInfo{"PREFETCH NTA", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_F3, 1), 1, X86InstInfo{"PREFETCH T0",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_F3, 2), 1, X86InstInfo{"PREFETCH T1",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_F3, 3), 1, X86InstInfo{"PREFETCH T2",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_F3, 4), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_F3, 5), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_F3, 6), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_F3, 7), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
 
-    {OPD(TYPE_GROUP_16, PF_66, 0), 1, X86InstInfo{"PREFETCH NTA", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_66, 1), 1, X86InstInfo{"PREFETCH T0",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_66, 2), 1, X86InstInfo{"PREFETCH T1",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_66, 3), 1, X86InstInfo{"PREFETCH T2",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_66, 4), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_66, 5), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_66, 6), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_66, 7), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
+    {OPD(TYPE_GROUP_16, PF_66, 0), 1, X86InstInfo{"PREFETCH NTA", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_66, 1), 1, X86InstInfo{"PREFETCH T0",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_66, 2), 1, X86InstInfo{"PREFETCH T1",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_66, 3), 1, X86InstInfo{"PREFETCH T2",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_66, 4), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_66, 5), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_66, 6), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_66, 7), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
 
-    {OPD(TYPE_GROUP_16, PF_F2, 0), 1, X86InstInfo{"PREFETCH NTA", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F2, 1), 1, X86InstInfo{"PREFETCH T0",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F2, 2), 1, X86InstInfo{"PREFETCH T1",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F2, 3), 1, X86InstInfo{"PREFETCH T2",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F2, 4), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F2, 5), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F2, 6), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
-    {OPD(TYPE_GROUP_16, PF_F2, 7), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0, nullptr}},
+    {OPD(TYPE_GROUP_16, PF_F2, 0), 1, X86InstInfo{"PREFETCH NTA", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_F2, 1), 1, X86InstInfo{"PREFETCH T0",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_F2, 2), 1, X86InstInfo{"PREFETCH T1",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_F2, 3), 1, X86InstInfo{"PREFETCH T2",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_16, PF_F2, 4), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_F2, 5), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_F2, 6), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
+    {OPD(TYPE_GROUP_16, PF_F2, 7), 1, X86InstInfo{"NOP",          TYPE_INST, FLAGS_MODRM,   0}},
 
     // GROUP 17
-    {OPD(TYPE_GROUP_17, PF_NONE, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_NONE, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_NONE, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_NONE, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_NONE, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_NONE, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_NONE, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_NONE, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0, nullptr}},
+    {OPD(TYPE_GROUP_17, PF_NONE, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0}},
+    {OPD(TYPE_GROUP_17, PF_NONE, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0}},
+    {OPD(TYPE_GROUP_17, PF_NONE, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0}},
+    {OPD(TYPE_GROUP_17, PF_NONE, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0}},
+    {OPD(TYPE_GROUP_17, PF_NONE, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0}},
+    {OPD(TYPE_GROUP_17, PF_NONE, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0}},
+    {OPD(TYPE_GROUP_17, PF_NONE, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0}},
+    {OPD(TYPE_GROUP_17, PF_NONE, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                          0}},
 
-    {OPD(TYPE_GROUP_17, PF_F3, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F3, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F3, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F3, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F3, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F3, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
+    {OPD(TYPE_GROUP_17, PF_F3, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F3, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F3, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F3, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F3, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F3, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F3, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F3, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
 
-    {OPD(TYPE_GROUP_17, PF_66, 0), 1, X86InstInfo{"EXTRQ", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS, 2, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_66, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_66, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_66, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_66, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_66, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_66, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_66, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
+    {OPD(TYPE_GROUP_17, PF_66, 0), 1, X86InstInfo{"EXTRQ", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_SF_MOD_REG_ONLY | FLAGS_XMM_FLAGS, 2}},
+    {OPD(TYPE_GROUP_17, PF_66, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_66, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_66, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_66, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_66, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_66, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_66, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
 
-    {OPD(TYPE_GROUP_17, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F2, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
-    {OPD(TYPE_GROUP_17, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0, nullptr}},
+    {OPD(TYPE_GROUP_17, PF_F2, 0), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F2, 1), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F2, 2), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F2, 3), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F2, 4), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F2, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F2, 6), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
+    {OPD(TYPE_GROUP_17, PF_F2, 7), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE,                                            0}},
 
     // GROUP P
     // AMD documentation claims n/a for all instructions in Group P
@@ -450,54 +481,48 @@ std::array<X86InstInfo, MAX_INST_SECOND_GROUP_TABLE_SIZE> SecondInstGroupOps = [
     // It claims that /3 is still Prefetch Mod
     // Tooling fails to decode past the /2 encoding but runs fine in hardware
     // Hardware also runs all the prefixes correctly
-    {OPD(TYPE_GROUP_P, PF_NONE, 0), 1, X86InstInfo{"PREFETCH Ex",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_NONE, 1), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_NONE, 2), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_NONE, 3), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_NONE, 4), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_NONE, 5), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_NONE, 6), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_NONE, 7), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0, nullptr}},
+    {OPD(TYPE_GROUP_P, PF_NONE, 0), 1, X86InstInfo{"PREFETCH Ex",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_P, PF_NONE, 1), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_P, PF_NONE, 2), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_P, PF_NONE, 3), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_P, PF_NONE, 4), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_P, PF_NONE, 5), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_P, PF_NONE, 6), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
+    {OPD(TYPE_GROUP_P, PF_NONE, 7), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY, 0}},
 
-    {OPD(TYPE_GROUP_P, PF_F3, 0), 1, X86InstInfo{"PREFETCH Ex",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F3, 1), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F3, 2), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F3, 3), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F3, 4), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F3, 5), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F3, 6), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F3, 7), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
+    {OPD(TYPE_GROUP_P, PF_F3, 0), 1, X86InstInfo{"PREFETCH Ex",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F3, 1), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F3, 2), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F3, 3), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F3, 4), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F3, 5), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F3, 6), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F3, 7), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
 
-    {OPD(TYPE_GROUP_P, PF_66, 0), 1, X86InstInfo{"PREFETCH Ex",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_66, 1), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_66, 2), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_66, 3), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_66, 4), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_66, 5), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_66, 6), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_66, 7), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
+    {OPD(TYPE_GROUP_P, PF_66, 0), 1, X86InstInfo{"PREFETCH Ex",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_66, 1), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_66, 2), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_66, 3), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_66, 4), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_66, 5), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_66, 6), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_66, 7), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
 
-    {OPD(TYPE_GROUP_P, PF_F2, 0), 1, X86InstInfo{"PREFETCH Ex",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F2, 1), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F2, 2), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F2, 3), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F2, 4), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F2, 5), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F2, 6), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
-    {OPD(TYPE_GROUP_P, PF_F2, 7), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0, nullptr}},
+    {OPD(TYPE_GROUP_P, PF_F2, 0), 1, X86InstInfo{"PREFETCH Ex",  TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F2, 1), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F2, 2), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F2, 3), 1, X86InstInfo{"PREFETCH Mod", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F2, 4), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F2, 5), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F2, 6), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
+    {OPD(TYPE_GROUP_P, PF_F2, 7), 1, X86InstInfo{"PREFETCH Res", TYPE_INST, FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY,   0}},
   };
 #undef OPD
 
-  GenerateTable(&Table.at(0), SecondaryExtensionOpTable, std::size(SecondaryExtensionOpTable));
+  GenerateTable(Table.data(), SecondaryExtensionOpTable, std::size(SecondaryExtensionOpTable));
 
   IR::InstallToTable(Table, IR::OpDispatch_SecondaryGroupTables);
   return Table;
 }();
 
-void InitializeSecondaryGroupTables(Context::OperatingMode Mode) {
-  if (Mode == Context::MODE_64BIT) {
-    IR::InstallToTable(SecondInstGroupOps, IR::OpDispatch_SecondaryGroupTables_64);
-  }
-}
-
 }

FEXCore/Source/Interface/Core/X86Tables/SecondaryModRMTables.cpp

@@ -12,51 +12,51 @@ $end_info$
 
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
-std::array<X86InstInfo, MAX_SECOND_MODRM_TABLE_SIZE> SecondModRMTableOps = []() consteval {
+constexpr std::array<X86InstInfo, MAX_SECOND_MODRM_TABLE_SIZE> SecondModRMTableOps = []() consteval {
   std::array<X86InstInfo, MAX_SECOND_MODRM_TABLE_SIZE> Table{};
   constexpr U8U8InfoStruct SecondaryModRMExtensionOpTable[] = {
     // REG /1
-    {((0 << 3) | 0), 1, X86InstInfo{"MONITOR",  TYPE_INST,    FLAGS_NONE, 0, nullptr}},
-    {((0 << 3) | 1), 1, X86InstInfo{"MWAIT",    TYPE_INST,    FLAGS_NONE, 0, nullptr}},
-    {((0 << 3) | 2), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((0 << 3) | 3), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((0 << 3) | 4), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((0 << 3) | 5), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((0 << 3) | 6), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((0 << 3) | 7), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+    {((0 << 3) | 0), 1, X86InstInfo{"MONITOR",  TYPE_INST,    FLAGS_NONE, 0}},
+    {((0 << 3) | 1), 1, X86InstInfo{"MWAIT",    TYPE_INST,    FLAGS_NONE, 0}},
+    {((0 << 3) | 2), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((0 << 3) | 3), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((0 << 3) | 4), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((0 << 3) | 5), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((0 << 3) | 6), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((0 << 3) | 7), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
 
     // REG /2
-    {((1 << 3) | 0), 1, X86InstInfo{"XGETBV",   TYPE_INST,    FLAGS_NONE, 0, nullptr}},
-    {((1 << 3) | 1), 1, X86InstInfo{"XSETBV",   TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((1 << 3) | 2), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((1 << 3) | 3), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((1 << 3) | 4), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((1 << 3) | 5), 1, X86InstInfo{"XEND",     TYPE_INST, FLAGS_NONE, 0, nullptr}},
-    {((1 << 3) | 6), 1, X86InstInfo{"XTEST",    TYPE_INST, FLAGS_NONE, 0, nullptr}},
-    {((1 << 3) | 7), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+    {((1 << 3) | 0), 1, X86InstInfo{"XGETBV",   TYPE_INST,    FLAGS_NONE, 0}},
+    {((1 << 3) | 1), 1, X86InstInfo{"XSETBV",   TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((1 << 3) | 2), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((1 << 3) | 3), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((1 << 3) | 4), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((1 << 3) | 5), 1, X86InstInfo{"XEND",     TYPE_INST, FLAGS_NONE, 0}},
+    {((1 << 3) | 6), 1, X86InstInfo{"XTEST",    TYPE_INST, FLAGS_NONE, 0}},
+    {((1 << 3) | 7), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
 
     // REG /3
-    {((2 << 3) | 0), 1, X86InstInfo{"VMRUN",    TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((2 << 3) | 1), 1, X86InstInfo{"VMMCALL",  TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((2 << 3) | 2), 1, X86InstInfo{"VMLOAD",   TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((2 << 3) | 3), 1, X86InstInfo{"VMSAVE",   TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((2 << 3) | 4), 1, X86InstInfo{"STGI",     TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((2 << 3) | 5), 1, X86InstInfo{"CLGI",     TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((2 << 3) | 6), 1, X86InstInfo{"SKINIT",   TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((2 << 3) | 7), 1, X86InstInfo{"INVLPGA",  TYPE_INST,    FLAGS_NONE, 0, nullptr}},
+    {((2 << 3) | 0), 1, X86InstInfo{"VMRUN",    TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((2 << 3) | 1), 1, X86InstInfo{"VMMCALL",  TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((2 << 3) | 2), 1, X86InstInfo{"VMLOAD",   TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((2 << 3) | 3), 1, X86InstInfo{"VMSAVE",   TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((2 << 3) | 4), 1, X86InstInfo{"STGI",     TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((2 << 3) | 5), 1, X86InstInfo{"CLGI",     TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((2 << 3) | 6), 1, X86InstInfo{"SKINIT",   TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((2 << 3) | 7), 1, X86InstInfo{"INVLPGA",  TYPE_INST,    FLAGS_NONE, 0}},
 
     // REG /7
-    {((3 << 3) | 0), 1, X86InstInfo{"SWAPGS",   TYPE_INST,    FLAGS_NONE, 0, nullptr}},
-    {((3 << 3) | 1), 1, X86InstInfo{"RDTSCP",   TYPE_INST,    FLAGS_NONE, 0, nullptr}},
-    {((3 << 3) | 2), 1, X86InstInfo{"MONITORX", TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((3 << 3) | 3), 1, X86InstInfo{"MWAITX",   TYPE_PRIV,    FLAGS_NONE, 0, nullptr}},
-    {((3 << 3) | 4), 1, X86InstInfo{"CLZERO",   TYPE_INST,    GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SF_SRC_RAX | FLAGS_DEBUG_MEM_ACCESS, 0, nullptr}},
-    {((3 << 3) | 5), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((3 << 3) | 6), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {((3 << 3) | 7), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+    {((3 << 3) | 0), 1, X86InstInfo{"SWAPGS",   TYPE_INST,    FLAGS_NONE, 0}},
+    {((3 << 3) | 1), 1, X86InstInfo{"RDTSCP",   TYPE_INST,    FLAGS_NONE, 0}},
+    {((3 << 3) | 2), 1, X86InstInfo{"MONITORX", TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((3 << 3) | 3), 1, X86InstInfo{"MWAITX",   TYPE_PRIV,    FLAGS_NONE, 0}},
+    {((3 << 3) | 4), 1, X86InstInfo{"CLZERO",   TYPE_INST,    GenFlagsSameSize(SIZE_64BITDEF) | FLAGS_SF_SRC_RAX | FLAGS_DEBUG_MEM_ACCESS, 0}},
+    {((3 << 3) | 5), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((3 << 3) | 6), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
+    {((3 << 3) | 7), 1, X86InstInfo{"",         TYPE_INVALID, FLAGS_NONE, 0}},
   };
 
-  GenerateTable(&Table.at(0), SecondaryModRMExtensionOpTable, std::size(SecondaryModRMExtensionOpTable));
+  GenerateTable(Table.data(), SecondaryModRMExtensionOpTable, std::size(SecondaryModRMExtensionOpTable));
 
   IR::InstallToTable(Table, IR::OpDispatch_SecondaryModRMTables);
   return Table;

FEXCore/Source/Interface/Core/X86Tables/X86Tables.h

@@ -6,18 +6,20 @@ $end_info$
 */
 
 #pragma once
-#include <FEXCore/Core/Context.h>
 
 #include <FEXCore/Utils/LogManager.h>
 
+#include <array>
+#include <cstddef>
+#include <cstdint>
+#include <cstring>
+#include <type_traits>
+
 namespace FEXCore::IR {
-///< Forward declaration of OpDispatchBuilder
 class OpDispatchBuilder;
 }
 
 namespace FEXCore::X86Tables {
-
-///< Forward declaration of X86InstInfo
 struct X86InstInfo;
 
 namespace DecodeFlags {
@@ -191,6 +193,7 @@ struct DecodedInst {
   X86InstInfo const* TableInfo;
 
   uint32_t Flags;
+  uint8_t OPRaw;
   uint16_t OP;
 
   uint8_t ModRM;
@@ -199,6 +202,7 @@ struct DecodedInst {
   uint8_t LastEscapePrefix;
   bool DecodedModRM;
   bool DecodedSIB;
+  bool ForceTSO;
 };
 
 union ModRMDecoded {
@@ -231,6 +235,9 @@ enum InstType {
   TYPE_X87 = TYPE_INST,
   TYPE_INVALID,
   TYPE_COPY_OTHER,
+  // Changes `X86InstInfo::OpcodeDispatcher` member to use the `Indirect` version.
+  // Points to a 2 member array of X86InstInfo to choose instruction description based on executing bitness.
+  TYPE_ARCH_DISPATCHER,
 
   // Must be in order
   // Groups 1, 1a, 2, 3, 4, 5, 11 are for the primary op table
@@ -430,12 +437,17 @@ constexpr uint8_t OpToIndex(uint8_t Op) {
 using DecodedOp = DecodedInst const*;
 using OpDispatchPtr = void (IR::OpDispatchBuilder::*)(DecodedOp);
 
+union OpDispatchPtrWrapper {
+  OpDispatchPtr OpDispatch;
+  const struct X86InstInfo *Indirect;
+};
+
 struct X86InstInfo {
   char const *Name;
   InstType Type;
   InstFlags::InstFlagType Flags; ///< Must be larger than InstFlags enum
   uint8_t MoreBytes;
-  OpDispatchPtr OpcodeDispatcher;
+  OpDispatchPtrWrapper OpcodeDispatcher;
 
   bool operator==(const X86InstInfo &b) const {
     if (strcmp(Name, b.Name) != 0 ||
@@ -477,23 +489,27 @@ constexpr size_t MAX_VEX_TABLE_SIZE = (1 << 13);
 // group select (3 bits for now) | ModRM opcode (3 bits)
 constexpr size_t MAX_VEX_GROUP_TABLE_SIZE = (1 << 7);
 
-extern std::array<X86InstInfo, MAX_PRIMARY_TABLE_SIZE> BaseOps;
-extern std::array<X86InstInfo, MAX_SECOND_TABLE_SIZE> SecondBaseOps;
-extern std::array<X86InstInfo, MAX_REP_MOD_TABLE_SIZE> RepModOps;
-extern std::array<X86InstInfo, MAX_REPNE_MOD_TABLE_SIZE> RepNEModOps;
-extern std::array<X86InstInfo, MAX_OPSIZE_MOD_TABLE_SIZE> OpSizeModOps;
+extern const std::array<X86InstInfo, MAX_PRIMARY_TABLE_SIZE> BaseOps;
+extern const std::array<X86InstInfo, MAX_SECOND_TABLE_SIZE> SecondBaseOps;
+extern const std::array<X86InstInfo, MAX_REP_MOD_TABLE_SIZE> RepModOps;
+extern const std::array<X86InstInfo, MAX_REPNE_MOD_TABLE_SIZE> RepNEModOps;
+extern const std::array<X86InstInfo, MAX_OPSIZE_MOD_TABLE_SIZE> OpSizeModOps;
 
-extern std::array<X86InstInfo, MAX_INST_GROUP_TABLE_SIZE> PrimaryInstGroupOps;
-extern std::array<X86InstInfo, MAX_INST_SECOND_GROUP_TABLE_SIZE> SecondInstGroupOps;
-extern std::array<X86InstInfo, MAX_SECOND_MODRM_TABLE_SIZE> SecondModRMTableOps;
-extern std::array<X86InstInfo, MAX_X87_TABLE_SIZE> X87Ops;
-extern std::array<X86InstInfo, MAX_3DNOW_TABLE_SIZE> DDDNowOps;
-extern std::array<X86InstInfo, MAX_0F_38_TABLE_SIZE> H0F38TableOps;
-extern std::array<X86InstInfo, MAX_0F_3A_TABLE_SIZE> H0F3ATableOps;
+extern const std::array<X86InstInfo, MAX_INST_GROUP_TABLE_SIZE> PrimaryInstGroupOps;
+extern const std::array<X86InstInfo, MAX_INST_SECOND_GROUP_TABLE_SIZE> SecondInstGroupOps;
+extern const std::array<X86InstInfo, MAX_SECOND_MODRM_TABLE_SIZE> SecondModRMTableOps;
+extern const std::array<X86InstInfo, MAX_X87_TABLE_SIZE> X87F80Ops;
+extern const std::array<X86InstInfo, MAX_X87_TABLE_SIZE> X87F64Ops;
+extern const std::array<X86InstInfo, MAX_3DNOW_TABLE_SIZE> DDDNowOps;
+extern const std::array<X86InstInfo, MAX_0F_38_TABLE_SIZE> H0F38TableOps;
+extern const std::array<X86InstInfo, MAX_0F_3A_TABLE_SIZE> H0F3ATableOps;
 
 // VEX
-extern std::array<X86InstInfo, MAX_VEX_TABLE_SIZE> VEXTableOps;
-extern std::array<X86InstInfo, MAX_VEX_GROUP_TABLE_SIZE> VEXTableGroupOps;
+extern const std::array<X86InstInfo, MAX_VEX_TABLE_SIZE> VEXTableOps;
+extern const std::array<X86InstInfo, MAX_VEX_GROUP_TABLE_SIZE> VEXTableGroupOps;
+
+extern const std::array<X86InstInfo, MAX_VEX_TABLE_SIZE> VEXTableOps_AVX128;
+extern const std::array<X86InstInfo, MAX_VEX_GROUP_TABLE_SIZE> VEXTableGroupOps_AVX128;
 
 template <typename OpcodeType>
 struct X86TablesInfoStruct {
@@ -514,7 +530,7 @@ constexpr static inline void GenerateTable(X86InstInfo *FinalTable, X86TablesInf
       if (FinalTable[OpNum + i].Type != TYPE_UNKNOWN) {
         LOGMAN_MSG_A_FMT("Duplicate Entry {}->{}", FinalTable[OpNum + i].Name, Info.Name);
       }
-      if (FinalTable[OpNum + i].OpcodeDispatcher) {
+      if (FinalTable[OpNum + i].OpcodeDispatcher.OpDispatch) {
         LOGMAN_MSG_A_FMT("Already installed an OpcodeDispatcher for 0x{:x}", OpNum + i);
       }
       FinalTable[OpNum + i] = Info;
@@ -523,7 +539,7 @@ constexpr static inline void GenerateTable(X86InstInfo *FinalTable, X86TablesInf
 };
 
 template<typename OpcodeType>
-constexpr static inline void GenerateTableWithCopy(X86InstInfo *FinalTable, X86TablesInfoStruct<OpcodeType> const *LocalTable, size_t TableSize, X86InstInfo *OtherLocal) {
+constexpr static inline void GenerateTableWithCopy(X86InstInfo *FinalTable, X86TablesInfoStruct<OpcodeType> const *LocalTable, size_t TableSize, const X86InstInfo *OtherLocal) {
   for (size_t j = 0; j < TableSize; ++j) {
     X86TablesInfoStruct<OpcodeType> const &Op = LocalTable[j];
     auto OpNum = Op.first;
@@ -587,8 +603,6 @@ constexpr static inline void GenerateX87Table(X86InstInfo *FinalTable, X86Tables
   }
 };
 
-void InitializeInfoTables(Context::OperatingMode Mode);
-
 }
 
 template <>

FEXCore/Source/Interface/Core/X86Tables/X87Tables.cpp

@@ -6,234 +6,760 @@ $end_info$
 */
 
 #include "Interface/Core/X86Tables/X86Tables.h"
+#include "Interface/Core/OpcodeDispatcher.h"
 
 #include <iterator>
 
 namespace FEXCore::X86Tables {
 using namespace InstFlags;
-std::array<X86InstInfo, MAX_X87_TABLE_SIZE> X87Ops = []() consteval {
-  std::array<X86InstInfo, MAX_X87_TABLE_SIZE> Table{};
+using namespace IR;
+// Top bit indicating if it needs to be repeated with {0x40, 0x80} or'd in
+// All OPDReg versions need it
+#define OPDReg(op, reg) ((1 << 15) | ((op - 0xD8) << 8) | (reg << 3))
+#define OPD(op, modrmop) (((op - 0xD8) << 8) | modrmop)
+constexpr std::array<DispatchTableEntry, 133> X87F64OpTable = {{
+  {OPDReg(0xD8, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADDF64, OpSize::i32Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMULF64, OpSize::i32Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 2) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::i32Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xD8, 3) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::i32Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xD8, 4) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::i32Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 5) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::i32Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 6) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::i32Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 7) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::i32Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPD(0xD8, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADDF64, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xC8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMULF64, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xD0), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+  {OPD(0xD8, 0xD8), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+  {OPD(0xD8, 0xE0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xE8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xF0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xF8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD9, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64, OpSize::i32Bit>},
+
+  // 1 = Invalid
+
+  {OPDReg(0xD9, 2) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::i32Bit>},
+
+  {OPDReg(0xD9, 3) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::i32Bit>},
+
+  {OPDReg(0xD9, 4) | 0x00, 8, &OpDispatchBuilder::X87LDENVF64},
+
+  {OPDReg(0xD9, 5) | 0x00, 8, &OpDispatchBuilder::X87FLDCWF64},
+
+  {OPDReg(0xD9, 6) | 0x00, 8, &OpDispatchBuilder::X87FNSTENV},
+
+  {OPDReg(0xD9, 7) | 0x00, 8, &OpDispatchBuilder::X87FSTCW},
+
+  {OPD(0xD9, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDFromStack>},
+  {OPD(0xD9, 0xC8), 8, &OpDispatchBuilder::FXCH},
+  {OPD(0xD9, 0xD0), 1, &OpDispatchBuilder::NOPOp}, // FNOP
+  // D1 = Invalid
+  // D8 = Invalid
+  {OPD(0xD9, 0xE0), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80STACKCHANGESIGN, false>},
+  {OPD(0xD9, 0xE1), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80STACKABS, false>},
+  // E2 = Invalid
+  {OPD(0xD9, 0xE4), 1, &OpDispatchBuilder::FTSTF64},
+  {OPD(0xD9, 0xE5), 1, &OpDispatchBuilder::X87FXAM},
+  // E6 = Invalid
+  {OPD(0xD9, 0xE8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64_Const, 0x3FF0000000000000>}, // 1.0
+  {OPD(0xD9, 0xE9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64_Const, 0x400A934F0979A372>}, // log2l(10)
+  {OPD(0xD9, 0xEA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64_Const, 0x3FF71547652B82FE>}, // log2l(e)
+  {OPD(0xD9, 0xEB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64_Const, 0x400921FB54442D18>}, // pi
+  {OPD(0xD9, 0xEC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64_Const, 0x3FD34413509F79FF>}, // log10l(2)
+  {OPD(0xD9, 0xED), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64_Const, 0x3FE62E42FEFA39EF>}, // log(2)
+  {OPD(0xD9, 0xEE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64_Const, 0>},                  // 0.0
+
+  // EF = Invalid
+  {OPD(0xD9, 0xF0), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80F2XM1STACK, false>},
+  {OPD(0xD9, 0xF1), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87FYL2X, false>},
+  {OPD(0xD9, 0xF2), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80PTANSTACK, true>},
+  {OPD(0xD9, 0xF3), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80ATANSTACK, false>},
+  {OPD(0xD9, 0xF4), 1, &OpDispatchBuilder::X87FXTRACTF64},
+  {OPD(0xD9, 0xF5), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80FPREM1STACK, true>},
+  {OPD(0xD9, 0xF6), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87ModifySTP, false>},
+  {OPD(0xD9, 0xF7), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87ModifySTP, true>},
+  {OPD(0xD9, 0xF8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80FPREMSTACK, true>},
+  {OPD(0xD9, 0xF9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87FYL2X, true>},
+  {OPD(0xD9, 0xFA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80SQRTSTACK, false>},
+  {OPD(0xD9, 0xFB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80SINCOSSTACK, true>},
+  {OPD(0xD9, 0xFC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80ROUNDSTACK, false>},
+  {OPD(0xD9, 0xFD), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80SCALESTACK, false>},
+  {OPD(0xD9, 0xFE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80SINSTACK, true>},
+  {OPD(0xD9, 0xFF), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80COSSTACK, true>},
+
+  {OPDReg(0xDA, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADDF64, OpSize::i32Bit, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMULF64, OpSize::i32Bit, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 2) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::i32Bit, true, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDA, 3) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::i32Bit, true, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDA, 4) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::i32Bit, true, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 5) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::i32Bit, true, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 6) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::i32Bit, true, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 7) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::i32Bit, true, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPD(0xDA, 0xC0), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDA, 0xC8), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDA, 0xD0), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDA, 0xD8), 8, &OpDispatchBuilder::X87FCMOV},
+  // E0 = Invalid
+  // E8 = Invalid
+  {OPD(0xDA, 0xE9), 1,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, true>},
+  // EA = Invalid
+  // F0 = Invalid
+  // F8 = Invalid
+
+  {OPDReg(0xDB, 0) | 0x00, 8, &OpDispatchBuilder::FILDF64},
+
+  {OPDReg(0xDB, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FISTF64, true>},
+
+  {OPDReg(0xDB, 2) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FISTF64, false>},
+
+  {OPDReg(0xDB, 3) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FISTF64, false>},
+
+  // 4 = Invalid
+
+  {OPDReg(0xDB, 5) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64, OpSize::f80Bit>},
+
+  // 6 = Invalid
+
+  {OPDReg(0xDB, 7) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::f80Bit>},
+
+
+  {OPD(0xDB, 0xC0), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDB, 0xC8), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDB, 0xD0), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDB, 0xD8), 8, &OpDispatchBuilder::X87FCMOV},
+  // E0 = Invalid
+  {OPD(0xDB, 0xE2), 1, &OpDispatchBuilder::FNCLEX},
+  {OPD(0xDB, 0xE3), 1, &OpDispatchBuilder::FNINIT},
+  // E4 = Invalid
+  {OPD(0xDB, 0xE8), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_RFLAGS, false>},
+  {OPD(0xDB, 0xF0), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_RFLAGS, false>},
+
+  // F8 = Invalid
+
+  {OPDReg(0xDC, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADDF64, OpSize::i64Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMULF64, OpSize::i64Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 2) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::i64Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDC, 3) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::i64Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDC, 4) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::i64Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 5) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::i64Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 6) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::i64Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 7) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::i64Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPD(0xDC, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADDF64, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xC8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMULF64, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xE0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xE8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xF0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xF8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_STI>},
+
+  {OPDReg(0xDD, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDF64, OpSize::i64Bit>},
+
+  {OPDReg(0xDD, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FISTF64, true>},
+
+  {OPDReg(0xDD, 2) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::i64Bit>},
+
+  {OPDReg(0xDD, 3) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::i64Bit>},
+
+  {OPDReg(0xDD, 4) | 0x00, 8, &OpDispatchBuilder::X87FRSTOR},
+
+  // 5 = Invalid
+  {OPDReg(0xDD, 6) | 0x00, 8, &OpDispatchBuilder::X87FNSAVE},
+
+  {OPDReg(0xDD, 7) | 0x00, 8, &OpDispatchBuilder::X87FNSTSW},
+
+  {OPD(0xDD, 0xC0), 8, &OpDispatchBuilder::X87FFREE},
+  {OPD(0xDD, 0xD0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSTToStack>}, // register-register from regular X87
+  {OPD(0xDD, 0xD8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSTToStack>}, //^
+
+  {OPD(0xDD, 0xE0), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+  {OPD(0xDD, 0xE8), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDE, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADDF64, OpSize::i16Bit, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMULF64, OpSize::i16Bit, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 2) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::i16Bit, true, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDE, 3) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::i16Bit, true, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDE, 4) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::i16Bit, true, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 5) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::i16Bit, true, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 6) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::i16Bit, true, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 7) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::i16Bit, true, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPD(0xDE, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADDF64, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xC8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMULF64, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xD9), 1,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, true>},
+  {OPD(0xDE, 0xE0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xE8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUBF64, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xF0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xF8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIVF64, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_STI>},
+
+  {OPDReg(0xDF, 0) | 0x00, 8, &OpDispatchBuilder::FILDF64},
+
+  {OPDReg(0xDF, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FISTF64, true>},
+
+  {OPDReg(0xDF, 2) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FISTF64, false>},
+
+  {OPDReg(0xDF, 3) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FISTF64, false>},
+
+  {OPDReg(0xDF, 4) | 0x00, 8, &OpDispatchBuilder::FBLDF64},
+
+  {OPDReg(0xDF, 5) | 0x00, 8, &OpDispatchBuilder::FILDF64},
+
+  {OPDReg(0xDF, 6) | 0x00, 8, &OpDispatchBuilder::FBSTPF64},
+
+  {OPDReg(0xDF, 7) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FISTF64, false>},
+
+  // XXX: This should also set the x87 tag bits to empty
+  // We don't support this currently, so just pop the stack
+  {OPD(0xDF, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87ModifySTP, true>},
+
+  {OPD(0xDF, 0xE0), 8, &OpDispatchBuilder::X87FNSTSW},
+  {OPD(0xDF, 0xE8), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_RFLAGS, false>},
+  {OPD(0xDF, 0xF0), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMIF64, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_RFLAGS, false>},
+}};
+
+constexpr std::array<DispatchTableEntry, 133> X87F80OpTable = {{
+  {OPDReg(0xD8, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADD, OpSize::i32Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMUL, OpSize::i32Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 2) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::i32Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xD8, 3) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::i32Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xD8, 4) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::i32Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 5) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::i32Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 6) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::i32Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD8, 7) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::i32Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPD(0xD8, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADD, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xC8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMUL, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xD0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+  {OPD(0xD8, 0xD8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+  {OPD(0xD8, 0xE0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xE8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xF0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+  {OPD(0xD8, 0xF8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xD9, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD, OpSize::i32Bit>},
+
+  // 1 = Invalid
+
+  {OPDReg(0xD9, 2) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::i32Bit>},
+
+  {OPDReg(0xD9, 3) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::i32Bit>},
+
+  {OPDReg(0xD9, 4) | 0x00, 8, &OpDispatchBuilder::X87LDENV},
+
+  {OPDReg(0xD9, 5) | 0x00, 8, &OpDispatchBuilder::X87FLDCW}, // XXX: stubbed FLDCW
+
+  {OPDReg(0xD9, 6) | 0x00, 8, &OpDispatchBuilder::X87FNSTENV},
+
+  {OPDReg(0xD9, 7) | 0x00, 8, &OpDispatchBuilder::X87FSTCW},
+
+  {OPD(0xD9, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLDFromStack>},
+  {OPD(0xD9, 0xC8), 8, &OpDispatchBuilder::FXCH},
+  {OPD(0xD9, 0xD0), 1, &OpDispatchBuilder::NOPOp}, // FNOP
+  // D1 = Invalid
+  // D8 = Invalid
+  {OPD(0xD9, 0xE0), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80STACKCHANGESIGN, false>},
+  {OPD(0xD9, 0xE1), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80STACKABS, false>},
+  // E2 = Invalid
+  {OPD(0xD9, 0xE4), 1, &OpDispatchBuilder::FTST},
+  {OPD(0xD9, 0xE5), 1, &OpDispatchBuilder::X87FXAM},
+  // E6 = Invalid
+  {OPD(0xD9, 0xE8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD_Const, NamedVectorConstant::NAMED_VECTOR_X87_ONE>},     // 1.0
+  {OPD(0xD9, 0xE9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD_Const, NamedVectorConstant::NAMED_VECTOR_X87_LOG2_10>}, // log2l(10)
+  {OPD(0xD9, 0xEA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD_Const, NamedVectorConstant::NAMED_VECTOR_X87_LOG2_E>}, // log2l(e)
+  {OPD(0xD9, 0xEB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD_Const, NamedVectorConstant::NAMED_VECTOR_X87_PI>},     // pi
+  {OPD(0xD9, 0xEC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD_Const, NamedVectorConstant::NAMED_VECTOR_X87_LOG10_2>}, // log10l(2)
+  {OPD(0xD9, 0xED), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD_Const, NamedVectorConstant::NAMED_VECTOR_X87_LOG_2>},   // log(2)
+  {OPD(0xD9, 0xEE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD_Const, NamedVectorConstant::NAMED_VECTOR_ZERO>},        // 0.0
+
+  // EF = Invalid
+  {OPD(0xD9, 0xF0), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80F2XM1STACK, false>},
+  {OPD(0xD9, 0xF1), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87FYL2X, false>},
+  {OPD(0xD9, 0xF2), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80PTANSTACK, true>},
+  {OPD(0xD9, 0xF3), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80ATANSTACK, false>},
+  {OPD(0xD9, 0xF4), 1, &OpDispatchBuilder::X87FXTRACT},
+  {OPD(0xD9, 0xF5), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80FPREM1STACK, true>},
+  {OPD(0xD9, 0xF6), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87ModifySTP, false>},
+  {OPD(0xD9, 0xF7), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87ModifySTP, true>},
+  {OPD(0xD9, 0xF8), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80FPREMSTACK, true>},
+  {OPD(0xD9, 0xF9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87FYL2X, true>},
+  {OPD(0xD9, 0xFA), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80SQRTSTACK, false>},
+  {OPD(0xD9, 0xFB), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80SINCOSSTACK, true>},
+  {OPD(0xD9, 0xFC), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80ROUNDSTACK, false>},
+  {OPD(0xD9, 0xFD), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80SCALESTACK, false>},
+  {OPD(0xD9, 0xFE), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80SINSTACK, true>},
+  {OPD(0xD9, 0xFF), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87OpHelper, OP_F80COSSTACK, true>},
+
+  {OPDReg(0xDA, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADD, OpSize::i32Bit, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMUL, OpSize::i32Bit, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 2) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::i32Bit, true, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDA, 3) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::i32Bit, true, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDA, 4) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::i32Bit, true, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 5) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::i32Bit, true, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 6) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::i32Bit, true, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDA, 7) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::i32Bit, true, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPD(0xDA, 0xC0), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDA, 0xC8), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDA, 0xD0), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDA, 0xD8), 8, &OpDispatchBuilder::X87FCMOV},
+  // E0 = Invalid
+  // E8 = Invalid
+  {OPD(0xDA, 0xE9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, true>},
+  // EA = Invalid
+  // F0 = Invalid
+  // F8 = Invalid
+
+  {OPDReg(0xDB, 0) | 0x00, 8, &OpDispatchBuilder::FILD},
+
+  {OPDReg(0xDB, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FIST, true>},
+
+  {OPDReg(0xDB, 2) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FIST, false>},
+
+  {OPDReg(0xDB, 3) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FIST, false>},
+
+  // 4 = Invalid
+
+  {OPDReg(0xDB, 5) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD, OpSize::f80Bit>},
+
+  // 6 = Invalid
+
+  {OPDReg(0xDB, 7) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::f80Bit>},
+
+
+  {OPD(0xDB, 0xC0), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDB, 0xC8), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDB, 0xD0), 8, &OpDispatchBuilder::X87FCMOV},
+  {OPD(0xDB, 0xD8), 8, &OpDispatchBuilder::X87FCMOV},
+  // E0 = Invalid
+  {OPD(0xDB, 0xE2), 1, &OpDispatchBuilder::FNCLEX},
+  {OPD(0xDB, 0xE3), 1, &OpDispatchBuilder::FNINIT},
+  // E4 = Invalid
+  {OPD(0xDB, 0xE8), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_RFLAGS, false>},
+  {OPD(0xDB, 0xF0), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_RFLAGS, false>},
+
+  // F8 = Invalid
+
+  {OPDReg(0xDC, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADD, OpSize::i64Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMUL, OpSize::i64Bit, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 2) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::i64Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDC, 3) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::i64Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDC, 4) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::i64Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 5) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::i64Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 6) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::i64Bit, false, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDC, 7) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::i64Bit, false, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPD(0xDC, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADD, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xC8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMUL, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xE0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xE8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xF0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDC, 0xF8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_STI>},
+
+  {OPDReg(0xDD, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FLD, OpSize::i64Bit>},
+
+  {OPDReg(0xDD, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FIST, true>},
+
+  {OPDReg(0xDD, 2) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::i64Bit>},
+
+  {OPDReg(0xDD, 3) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FST, OpSize::i64Bit>},
+
+  {OPDReg(0xDD, 4) | 0x00, 8, &OpDispatchBuilder::X87FRSTOR},
+
+  // 5 = Invalid
+  {OPDReg(0xDD, 6) | 0x00, 8, &OpDispatchBuilder::X87FNSAVE},
+
+  {OPDReg(0xDD, 7) | 0x00, 8, &OpDispatchBuilder::X87FNSTSW},
+
+  {OPD(0xDD, 0xC0), 8, &OpDispatchBuilder::X87FFREE},
+  {OPD(0xDD, 0xD0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSTToStack>},
+  {OPD(0xDD, 0xD8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSTToStack>},
+
+  {OPD(0xDD, 0xE0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+  {OPD(0xDD, 0xE8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDE, 0) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADD, OpSize::i16Bit, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMUL, OpSize::i16Bit, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 2) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::i16Bit, true, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDE, 3) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::i16Bit, true, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, false>},
+
+  {OPDReg(0xDE, 4) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::i16Bit, true, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 5) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::i16Bit, true, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 6) | 0x00, 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::i16Bit, true, false, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPDReg(0xDE, 7) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::i16Bit, true, true, OpDispatchBuilder::OpResult::RES_ST0>},
+
+  {OPD(0xDE, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FADD, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xC8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FMUL, OpSize::f80Bit, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xD9), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_X87, true>},
+  {OPD(0xDE, 0xE0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xE8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FSUB, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xF0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::f80Bit, false, true, OpDispatchBuilder::OpResult::RES_STI>},
+  {OPD(0xDE, 0xF8), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FDIV, OpSize::f80Bit, false, false, OpDispatchBuilder::OpResult::RES_STI>},
+
+  {OPDReg(0xDF, 0) | 0x00, 8, &OpDispatchBuilder::FILD},
+
+  {OPDReg(0xDF, 1) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FIST, true>},
+
+  {OPDReg(0xDF, 2) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FIST, false>},
+
+  {OPDReg(0xDF, 3) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FIST, false>},
+
+  {OPDReg(0xDF, 4) | 0x00, 8, &OpDispatchBuilder::FBLD},
+
+  {OPDReg(0xDF, 5) | 0x00, 8, &OpDispatchBuilder::FILD},
+
+  {OPDReg(0xDF, 6) | 0x00, 8, &OpDispatchBuilder::FBSTP},
+
+  {OPDReg(0xDF, 7) | 0x00, 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::FIST, false>},
+
+  // XXX: This should also set the x87 tag bits to empty
+  // We don't support this currently, so just pop the stack
+  {OPD(0xDF, 0xC0), 8, &OpDispatchBuilder::Bind<&OpDispatchBuilder::X87ModifySTP, true>},
+
+  {OPD(0xDF, 0xE0), 8, &OpDispatchBuilder::X87FNSTSW},
+  {OPD(0xDF, 0xE8), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_RFLAGS, false>},
+  {OPD(0xDF, 0xF0), 8,
+   &OpDispatchBuilder::Bind<&OpDispatchBuilder::FCOMI, OpSize::f80Bit, false, OpDispatchBuilder::FCOMIFlags::FLAGS_RFLAGS, false>},
+}};
+#undef OPD
+#undef OPDReg
+
+auto GenerateX87TableLambda = [](const auto DispatchTable) consteval {
 #define OPD(op, modrmop) (((op - 0xD8) << 8) | modrmop)
 #define OPDReg(op, reg) (((op - 0xD8) << 8) | (reg << 3))
+ std::array<X86InstInfo, MAX_X87_TABLE_SIZE> Table{};
   constexpr U16U8InfoStruct X87OpTable[] = {
     // 0xD8
-    {OPDReg(0xD8, 0), 1, X86InstInfo{"FADD",  TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD8, 1), 1, X86InstInfo{"FMUL",  TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD8, 2), 1, X86InstInfo{"FCOM",  TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD8, 3), 1, X86InstInfo{"FCOMP", TYPE_X87, FLAGS_MODRM | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xD8, 4), 1, X86InstInfo{"FSUB",  TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD8, 5), 1, X86InstInfo{"FSUBR", TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD8, 6), 1, X86InstInfo{"FDIV",  TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD8, 7), 1, X86InstInfo{"FDIVR", TYPE_X87, FLAGS_MODRM, 0, nullptr}},
+    {OPDReg(0xD8, 0), 1, X86InstInfo{"FADD",  TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xD8, 1), 1, X86InstInfo{"FMUL",  TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xD8, 2), 1, X86InstInfo{"FCOM",  TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xD8, 3), 1, X86InstInfo{"FCOMP", TYPE_X87, FLAGS_MODRM | FLAGS_POP, 0}},
+    {OPDReg(0xD8, 4), 1, X86InstInfo{"FSUB",  TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xD8, 5), 1, X86InstInfo{"FSUBR", TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xD8, 6), 1, X86InstInfo{"FDIV",  TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xD8, 7), 1, X86InstInfo{"FDIVR", TYPE_X87, FLAGS_MODRM, 0}},
       //  / 0
-      {OPD(0xD8, 0xC0), 8, X86InstInfo{"FADD", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD8, 0xC0), 8, X86InstInfo{"FADD", TYPE_X87, FLAGS_NONE, 0}},
       //  / 1
-      {OPD(0xD8, 0xC8), 8, X86InstInfo{"FMUL", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD8, 0xC8), 8, X86InstInfo{"FMUL", TYPE_X87, FLAGS_NONE, 0}},
       //  / 2
-      {OPD(0xD8, 0xD0), 8, X86InstInfo{"FCOM", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD8, 0xD0), 8, X86InstInfo{"FCOM", TYPE_X87, FLAGS_NONE, 0}},
       //  / 3
-      {OPD(0xD8, 0xD8), 8, X86InstInfo{"FCOMP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xD8, 0xD8), 8, X86InstInfo{"FCOMP", TYPE_X87, FLAGS_POP, 0}},
       //  / 4
-      {OPD(0xD8, 0xE0), 8, X86InstInfo{"FSUB", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD8, 0xE0), 8, X86InstInfo{"FSUB", TYPE_X87, FLAGS_NONE, 0}},
       //  / 5
-      {OPD(0xD8, 0xE8), 8, X86InstInfo{"FSUBR", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD8, 0xE8), 8, X86InstInfo{"FSUBR", TYPE_X87, FLAGS_NONE, 0}},
       //  / 6
-      {OPD(0xD8, 0xF0), 8, X86InstInfo{"FDIV", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD8, 0xF0), 8, X86InstInfo{"FDIV", TYPE_X87, FLAGS_NONE, 0}},
       //  / 7
-      {OPD(0xD8, 0xF8), 8, X86InstInfo{"FDIVR", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD8, 0xF8), 8, X86InstInfo{"FDIVR", TYPE_X87, FLAGS_NONE, 0}},
     // 0xD9
-    {OPDReg(0xD9, 0), 1, X86InstInfo{"FLD",     TYPE_INST, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD9, 1), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {OPDReg(0xD9, 2), 1, X86InstInfo{"FST",     TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPDReg(0xD9, 3), 1, X86InstInfo{"FSTP",    TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xD9, 4), 1, X86InstInfo{"FLDENV",  TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD9, 5), 1, X86InstInfo{"FLDCW",   TYPE_X87, GenFlagsSameSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xD9, 6), 1, X86InstInfo{"FNSTENV", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPDReg(0xD9, 7), 1, X86InstInfo{"FNSTCW",  TYPE_INST, GenFlagsSameSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
+    {OPDReg(0xD9, 0), 1, X86InstInfo{"FLD",     TYPE_INST, FLAGS_MODRM, 0}},
+    {OPDReg(0xD9, 1), 1, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0}},
+    {OPDReg(0xD9, 2), 1, X86InstInfo{"FST",     TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPDReg(0xD9, 3), 1, X86InstInfo{"FSTP",    TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
+    {OPDReg(0xD9, 4), 1, X86InstInfo{"FLDENV",  TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xD9, 5), 1, X86InstInfo{"FLDCW",   TYPE_X87, GenFlagsSameSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xD9, 6), 1, X86InstInfo{"FNSTENV", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPDReg(0xD9, 7), 1, X86InstInfo{"FNSTCW",  TYPE_INST, GenFlagsSameSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
       //  / 0
-      {OPD(0xD9, 0xC0), 8, X86InstInfo{"FLD",   TYPE_INST, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD9, 0xC0), 8, X86InstInfo{"FLD",   TYPE_INST, FLAGS_NONE, 0}},
       //  / 1
-      {OPD(0xD9, 0xC8), 8, X86InstInfo{"FXCH",  TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD9, 0xC8), 8, X86InstInfo{"FXCH",  TYPE_X87, FLAGS_NONE, 0}},
       //  / 2
-      {OPD(0xD9, 0xD0), 1, X86InstInfo{"FNOP",  TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xD1), 7, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD9, 0xD0), 1, X86InstInfo{"FNOP",  TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xD1), 7, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 3
-      {OPD(0xD9, 0xD8), 8, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD9, 0xD8), 8, X86InstInfo{"",      TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 4
-      {OPD(0xD9, 0xE0), 1, X86InstInfo{"FCHS", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xE1), 1, X86InstInfo{"FABS", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xE2), 2, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xE4), 1, X86InstInfo{"FTST", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xE5), 1, X86InstInfo{"FXAM", TYPE_INST,  FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xE6), 2, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD9, 0xE0), 1, X86InstInfo{"FCHS", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xE1), 1, X86InstInfo{"FABS", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xE2), 2, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xE4), 1, X86InstInfo{"FTST", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xE5), 1, X86InstInfo{"FXAM", TYPE_INST,  FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xE6), 2, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 5
-      {OPD(0xD9, 0xE8), 1, X86InstInfo{"FLD1", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xE9), 1, X86InstInfo{"FLDL2T", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xEA), 1, X86InstInfo{"FLDL2E", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xEB), 1, X86InstInfo{"FLDPI", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xEC), 1, X86InstInfo{"FLDLG2", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xED), 1, X86InstInfo{"FLDLN2", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xEE), 1, X86InstInfo{"FLDZ", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xEF), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD9, 0xE8), 1, X86InstInfo{"FLD1", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xE9), 1, X86InstInfo{"FLDL2T", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xEA), 1, X86InstInfo{"FLDL2E", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xEB), 1, X86InstInfo{"FLDPI", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xEC), 1, X86InstInfo{"FLDLG2", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xED), 1, X86InstInfo{"FLDLN2", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xEE), 1, X86InstInfo{"FLDZ", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xEF), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 6
-      {OPD(0xD9, 0xF0), 1, X86InstInfo{"F2XM1", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xF1), 1, X86InstInfo{"FYL2X", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xF2), 1, X86InstInfo{"FPTAN", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xF3), 1, X86InstInfo{"FPATAN", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xF4), 1, X86InstInfo{"FXTRACT", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xF5), 1, X86InstInfo{"FPREM1", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xF6), 1, X86InstInfo{"FDECSTP", TYPE_X87, FLAGS_POP, 0, nullptr}},
-      {OPD(0xD9, 0xF7), 1, X86InstInfo{"FINCSTP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xD9, 0xF0), 1, X86InstInfo{"F2XM1", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xF1), 1, X86InstInfo{"FYL2X", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xF2), 1, X86InstInfo{"FPTAN", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xF3), 1, X86InstInfo{"FPATAN", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xF4), 1, X86InstInfo{"FXTRACT", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xF5), 1, X86InstInfo{"FPREM1", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xF6), 1, X86InstInfo{"FDECSTP", TYPE_X87, FLAGS_POP, 0}},
+      {OPD(0xD9, 0xF7), 1, X86InstInfo{"FINCSTP", TYPE_X87, FLAGS_POP, 0}},
       //  / 7
-      {OPD(0xD9, 0xF8), 1, X86InstInfo{"FPREM", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xF9), 1, X86InstInfo{"FYL2XP1", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xFA), 1, X86InstInfo{"FSQRT", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xFB), 1, X86InstInfo{"FSINCOS", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xFC), 1, X86InstInfo{"FRNDINT", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xFD), 1, X86InstInfo{"FSCALE", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xFE), 1, X86InstInfo{"FSIN", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xD9, 0xFF), 1, X86InstInfo{"FCOS", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xD9, 0xF8), 1, X86InstInfo{"FPREM", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xF9), 1, X86InstInfo{"FYL2XP1", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xFA), 1, X86InstInfo{"FSQRT", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xFB), 1, X86InstInfo{"FSINCOS", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xFC), 1, X86InstInfo{"FRNDINT", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xFD), 1, X86InstInfo{"FSCALE", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xFE), 1, X86InstInfo{"FSIN", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xD9, 0xFF), 1, X86InstInfo{"FCOS", TYPE_X87, FLAGS_NONE, 0}},
     // 0xDA
-    {OPDReg(0xDA, 0), 1, X86InstInfo{"FIADD", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDA, 1), 1, X86InstInfo{"FIMUL", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDA, 2), 1, X86InstInfo{"FICOM", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDA, 3), 1, X86InstInfo{"FICOMP", TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDA, 4), 1, X86InstInfo{"FISUB", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDA, 5), 1, X86InstInfo{"FISUBR", TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDA, 6), 1, X86InstInfo{"FIDIV", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDA, 7), 1, X86InstInfo{"FIDIVR", TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0, nullptr}},
+    {OPDReg(0xDA, 0), 1, X86InstInfo{"FIADD", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDA, 1), 1, X86InstInfo{"FIMUL", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDA, 2), 1, X86InstInfo{"FICOM", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDA, 3), 1, X86InstInfo{"FICOMP", TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_POP, 0}},
+    {OPDReg(0xDA, 4), 1, X86InstInfo{"FISUB", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDA, 5), 1, X86InstInfo{"FISUBR", TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDA, 6), 1, X86InstInfo{"FIDIV", TYPE_X87,  GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDA, 7), 1, X86InstInfo{"FIDIVR", TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0}},
       //  / 0
-      {OPD(0xDA, 0xC0), 8, X86InstInfo{"FCMOVB", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDA, 0xC0), 8, X86InstInfo{"FCMOVB", TYPE_X87, FLAGS_NONE, 0}},
       //  / 1
-      {OPD(0xDA, 0xC8), 8, X86InstInfo{"FCMOVE", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDA, 0xC8), 8, X86InstInfo{"FCMOVE", TYPE_X87, FLAGS_NONE, 0}},
       //  / 2
-      {OPD(0xDA, 0xD0), 8, X86InstInfo{"FCMOVBE", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDA, 0xD0), 8, X86InstInfo{"FCMOVBE", TYPE_X87, FLAGS_NONE, 0}},
       //  / 3
-      {OPD(0xDA, 0xD8), 8, X86InstInfo{"FCMOVU", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDA, 0xD8), 8, X86InstInfo{"FCMOVU", TYPE_X87, FLAGS_NONE, 0}},
       //  / 4
-      {OPD(0xDA, 0xE0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDA, 0xE0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 5
-      {OPD(0xDA, 0xE8), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xDA, 0xE9), 1, X86InstInfo{"FUCOMPP", TYPE_X87, FLAGS_POP, 0, nullptr}},
-      {OPD(0xDA, 0xEA), 6, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDA, 0xE8), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
+      {OPD(0xDA, 0xE9), 1, X86InstInfo{"FUCOMPP", TYPE_X87, FLAGS_POP, 0}},
+      {OPD(0xDA, 0xEA), 6, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 6
-      {OPD(0xDA, 0xF0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDA, 0xF0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 7
-      {OPD(0xDA, 0xF8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDA, 0xF8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
     // 0xDB
-    {OPDReg(0xDB, 0), 1, X86InstInfo{"FILD",   TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDB, 1), 1, X86InstInfo{"FISTTP", TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDB, 2), 1, X86InstInfo{"FIST",   TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPDReg(0xDB, 3), 1, X86InstInfo{"FISTP",  TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDB, 4), 1, X86InstInfo{"",       TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {OPDReg(0xDB, 5), 1, X86InstInfo{"FLD",    TYPE_X87,    FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDB, 6), 1, X86InstInfo{"",       TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {OPDReg(0xDB, 7), 1, X86InstInfo{"FSTP",   TYPE_X87,   FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
+    {OPDReg(0xDB, 0), 1, X86InstInfo{"FILD",   TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDB, 1), 1, X86InstInfo{"FISTTP", TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
+    {OPDReg(0xDB, 2), 1, X86InstInfo{"FIST",   TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPDReg(0xDB, 3), 1, X86InstInfo{"FISTP",  TYPE_X87, GenFlagsSrcSize(SIZE_32BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
+    {OPDReg(0xDB, 4), 1, X86InstInfo{"",       TYPE_INVALID, FLAGS_NONE, 0}},
+    {OPDReg(0xDB, 5), 1, X86InstInfo{"FLD",    TYPE_X87,    FLAGS_MODRM, 0}},
+    {OPDReg(0xDB, 6), 1, X86InstInfo{"",       TYPE_INVALID, FLAGS_NONE, 0}},
+    {OPDReg(0xDB, 7), 1, X86InstInfo{"FSTP",   TYPE_X87,   FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
       //  / 0
-      {OPD(0xDB, 0xC0), 8, X86InstInfo{"FCMOVNB", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDB, 0xC0), 8, X86InstInfo{"FCMOVNB", TYPE_X87, FLAGS_NONE, 0}},
       //  / 1
-      {OPD(0xDB, 0xC8), 8, X86InstInfo{"FCMOVNE", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDB, 0xC8), 8, X86InstInfo{"FCMOVNE", TYPE_X87, FLAGS_NONE, 0}},
       //  / 2
-      {OPD(0xDB, 0xD0), 8, X86InstInfo{"FCMOVNBE", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDB, 0xD0), 8, X86InstInfo{"FCMOVNBE", TYPE_X87, FLAGS_NONE, 0}},
       //  / 3
-      {OPD(0xDB, 0xD8), 8, X86InstInfo{"FCMOVNU", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDB, 0xD8), 8, X86InstInfo{"FCMOVNU", TYPE_X87, FLAGS_NONE, 0}},
       //  / 4
-      {OPD(0xDB, 0xE0), 2, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xDB, 0xE2), 1, X86InstInfo{"FNCLEX", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xDB, 0xE3), 1, X86InstInfo{"FNINIT", TYPE_X87, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xDB, 0xE4), 4, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDB, 0xE0), 2, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
+      {OPD(0xDB, 0xE2), 1, X86InstInfo{"FNCLEX", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xDB, 0xE3), 1, X86InstInfo{"FNINIT", TYPE_X87, FLAGS_NONE, 0}},
+      {OPD(0xDB, 0xE4), 4, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 5
-      {OPD(0xDB, 0xE8), 8, X86InstInfo{"FUCOMI", TYPE_INST, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDB, 0xE8), 8, X86InstInfo{"FUCOMI", TYPE_INST, FLAGS_NONE, 0}},
       //  / 6
-      {OPD(0xDB, 0xF0), 8, X86InstInfo{"FCOMI", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDB, 0xF0), 8, X86InstInfo{"FCOMI", TYPE_X87, FLAGS_NONE, 0}},
       //  / 7
-      {OPD(0xDB, 0xF8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDB, 0xF8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
     // 0xDC
-    {OPDReg(0xDC, 0), 1, X86InstInfo{"FADD", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0, nullptr}},
-    {OPDReg(0xDC, 1), 1, X86InstInfo{"FMUL", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0, nullptr}},
-    {OPDReg(0xDC, 2), 1, X86InstInfo{"FCOM", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0, nullptr}},
-    {OPDReg(0xDC, 3), 1, X86InstInfo{"FCOMP", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDC, 4), 1, X86InstInfo{"FSUB", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0, nullptr}},
-    {OPDReg(0xDC, 5), 1, X86InstInfo{"FSUBR", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0, nullptr}},
-    {OPDReg(0xDC, 6), 1, X86InstInfo{"FDIV", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0, nullptr}},
-    {OPDReg(0xDC, 7), 1, X86InstInfo{"FDIVR", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0, nullptr}},
+    {OPDReg(0xDC, 0), 1, X86InstInfo{"FADD", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0}},
+    {OPDReg(0xDC, 1), 1, X86InstInfo{"FMUL", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0}},
+    {OPDReg(0xDC, 2), 1, X86InstInfo{"FCOM", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0}},
+    {OPDReg(0xDC, 3), 1, X86InstInfo{"FCOMP", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS | FLAGS_POP, 0}},
+    {OPDReg(0xDC, 4), 1, X86InstInfo{"FSUB", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0}},
+    {OPDReg(0xDC, 5), 1, X86InstInfo{"FSUBR", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0}},
+    {OPDReg(0xDC, 6), 1, X86InstInfo{"FDIV", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0}},
+    {OPDReg(0xDC, 7), 1, X86InstInfo{"FDIVR", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0}},
       //  / 0
-      {OPD(0xDC, 0xC0), 8, X86InstInfo{"FADD", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDC, 0xC0), 8, X86InstInfo{"FADD", TYPE_X87, FLAGS_NONE, 0}},
       //  / 1
-      {OPD(0xDC, 0xC8), 8, X86InstInfo{"FMUL", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDC, 0xC8), 8, X86InstInfo{"FMUL", TYPE_X87, FLAGS_NONE, 0}},
       //  / 2
-      {OPD(0xDC, 0xD0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDC, 0xD0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 3
-      {OPD(0xDC, 0xD8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDC, 0xD8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 4
-      {OPD(0xDC, 0xE0), 8, X86InstInfo{"FSUBR", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDC, 0xE0), 8, X86InstInfo{"FSUBR", TYPE_X87, FLAGS_NONE, 0}},
       //  / 5
-      {OPD(0xDC, 0xE8), 8, X86InstInfo{"FSUB", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDC, 0xE8), 8, X86InstInfo{"FSUB", TYPE_X87, FLAGS_NONE, 0}},
       //  / 6
-      {OPD(0xDC, 0xF0), 8, X86InstInfo{"FDIVR", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDC, 0xF0), 8, X86InstInfo{"FDIVR", TYPE_X87, FLAGS_NONE, 0}},
       //  / 7
-      {OPD(0xDC, 0xF8), 8, X86InstInfo{"FDIV", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDC, 0xF8), 8, X86InstInfo{"FDIV", TYPE_X87, FLAGS_NONE, 0}},
     // 0xDD
-    {OPDReg(0xDD, 0), 1, X86InstInfo{"FLD", TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDD, 1), 1, X86InstInfo{"FISTTP", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDD, 2), 1, X86InstInfo{"FST", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPDReg(0xDD, 3), 1, X86InstInfo{"FSTP", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDD, 4), 1, X86InstInfo{"FRSTOR", TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDD, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-    {OPDReg(0xDD, 6), 1, X86InstInfo{"FNSAVE", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPDReg(0xDD, 7), 1, X86InstInfo{"FNSTSW", TYPE_X87, GenFlagsSameSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
+    {OPDReg(0xDD, 0), 1, X86InstInfo{"FLD", TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xDD, 1), 1, X86InstInfo{"FISTTP", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
+    {OPDReg(0xDD, 2), 1, X86InstInfo{"FST", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPDReg(0xDD, 3), 1, X86InstInfo{"FSTP", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
+    {OPDReg(0xDD, 4), 1, X86InstInfo{"FRSTOR", TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xDD, 5), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
+    {OPDReg(0xDD, 6), 1, X86InstInfo{"FNSAVE", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPDReg(0xDD, 7), 1, X86InstInfo{"FNSTSW", TYPE_X87, GenFlagsSameSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
       //  / 0
-      {OPD(0xDD, 0xC0), 8, X86InstInfo{"FFREE", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDD, 0xC0), 8, X86InstInfo{"FFREE", TYPE_X87, FLAGS_NONE, 0}},
       //  / 1
-      {OPD(0xDD, 0xC8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDD, 0xC8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 2
-      {OPD(0xDD, 0xD0), 8, X86InstInfo{"FST", TYPE_INST, FLAGS_SF_MOD_DST, 0, nullptr}},
+      {OPD(0xDD, 0xD0), 8, X86InstInfo{"FST", TYPE_INST, FLAGS_SF_MOD_DST, 0}},
       //  / 3
-      {OPD(0xDD, 0xD8), 8, X86InstInfo{"FSTP", TYPE_X87, FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
+      {OPD(0xDD, 0xD8), 8, X86InstInfo{"FSTP", TYPE_X87, FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
       //  / 4
-      {OPD(0xDD, 0xE0), 8, X86InstInfo{"FUCOM", TYPE_X87, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDD, 0xE0), 8, X86InstInfo{"FUCOM", TYPE_X87, FLAGS_NONE, 0}},
       //  / 5
-      {OPD(0xDD, 0xE8), 8, X86InstInfo{"FUCOMP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xDD, 0xE8), 8, X86InstInfo{"FUCOMP", TYPE_X87, FLAGS_POP, 0}},
       //  / 6
-      {OPD(0xDD, 0xF0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDD, 0xF0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 7
-      {OPD(0xDD, 0xF8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDD, 0xF8), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
     // 0xDE
-    {OPDReg(0xDE, 0), 1, X86InstInfo{"FIADD", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDE, 1), 1, X86InstInfo{"FIMUL", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDE, 2), 1, X86InstInfo{"FICOM", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDE, 3), 1, X86InstInfo{"FICOMP", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDE, 4), 1, X86InstInfo{"FISUB", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDE, 5), 1, X86InstInfo{"FISUBR", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDE, 6), 1, X86InstInfo{"FIDIV", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDE, 7), 1, X86InstInfo{"FIDIVR", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
+    {OPDReg(0xDE, 0), 1, X86InstInfo{"FIADD", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDE, 1), 1, X86InstInfo{"FIMUL", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDE, 2), 1, X86InstInfo{"FICOM", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDE, 3), 1, X86InstInfo{"FICOMP", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_POP, 0}},
+    {OPDReg(0xDE, 4), 1, X86InstInfo{"FISUB", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDE, 5), 1, X86InstInfo{"FISUBR", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDE, 6), 1, X86InstInfo{"FIDIV", TYPE_X87,  GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDE, 7), 1, X86InstInfo{"FIDIVR", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
       //  / 0
-      {OPD(0xDE, 0xC0), 8, X86InstInfo{"FADDP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xDE, 0xC0), 8, X86InstInfo{"FADDP", TYPE_X87, FLAGS_POP, 0}},
       //  / 1
-      {OPD(0xDE, 0xC8), 8, X86InstInfo{"FMULP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xDE, 0xC8), 8, X86InstInfo{"FMULP", TYPE_X87, FLAGS_POP, 0}},
       //  / 2
-      {OPD(0xDE, 0xD0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDE, 0xD0), 8, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 3
-      {OPD(0xDE, 0xD8), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
-      {OPD(0xDE, 0xD9), 1, X86InstInfo{"FCOMPP", TYPE_X87, FLAGS_POP, 0, nullptr}},
-      {OPD(0xDE, 0xDA), 6, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDE, 0xD8), 1, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
+      {OPD(0xDE, 0xD9), 1, X86InstInfo{"FCOMPP", TYPE_X87, FLAGS_POP, 0}},
+      {OPD(0xDE, 0xDA), 6, X86InstInfo{"", TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 4
-      {OPD(0xDE, 0xE0), 8, X86InstInfo{"FSUBRP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xDE, 0xE0), 8, X86InstInfo{"FSUBRP", TYPE_X87, FLAGS_POP, 0}},
       //  / 5
-      {OPD(0xDE, 0xE8), 8, X86InstInfo{"FSUBP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xDE, 0xE8), 8, X86InstInfo{"FSUBP", TYPE_X87, FLAGS_POP, 0}},
       //  / 6
-      {OPD(0xDE, 0xF0), 8, X86InstInfo{"FDIVRP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xDE, 0xF0), 8, X86InstInfo{"FDIVRP", TYPE_X87, FLAGS_POP, 0}},
       //  / 7
-      {OPD(0xDE, 0xF8), 8, X86InstInfo{"FDIVP", TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xDE, 0xF8), 8, X86InstInfo{"FDIVP", TYPE_X87, FLAGS_POP, 0}},
     // 0xDF
-    {OPDReg(0xDF, 0), 1, X86InstInfo{"FILD", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDF, 1), 1, X86InstInfo{"FISTTP", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDF, 2), 1, X86InstInfo{"FIST",   TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 0, nullptr}},
-    {OPDReg(0xDF, 3), 1, X86InstInfo{"FISTP",  TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDF, 4), 1, X86InstInfo{"FBLD", TYPE_X87, FLAGS_MODRM, 0, nullptr}},
-    {OPDReg(0xDF, 5), 1, X86InstInfo{"FILD", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0, nullptr}},
-    {OPDReg(0xDF, 6), 1, X86InstInfo{"FBSTP", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
-    {OPDReg(0xDF, 7), 1, X86InstInfo{"FISTP", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS | FLAGS_SF_MOD_DST | FLAGS_POP, 0, nullptr}},
+    {OPDReg(0xDF, 0), 1, X86InstInfo{"FILD", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM, 0}},
+    {OPDReg(0xDF, 1), 1, X86InstInfo{"FISTTP", TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
+    {OPDReg(0xDF, 2), 1, X86InstInfo{"FIST",   TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST, 0}},
+    {OPDReg(0xDF, 3), 1, X86InstInfo{"FISTP",  TYPE_X87, GenFlagsSrcSize(SIZE_16BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
+    {OPDReg(0xDF, 4), 1, X86InstInfo{"FBLD", TYPE_X87, FLAGS_MODRM, 0}},
+    {OPDReg(0xDF, 5), 1, X86InstInfo{"FILD", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS, 0}},
+    {OPDReg(0xDF, 6), 1, X86InstInfo{"FBSTP", TYPE_X87, FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
+    {OPDReg(0xDF, 7), 1, X86InstInfo{"FISTP", TYPE_X87, GenFlagsSrcSize(SIZE_64BIT) | FLAGS_MODRM | FLAGS_X87_FLAGS | FLAGS_SF_MOD_DST | FLAGS_POP, 0}},
       //  / 0
       //  This instruction is a bit special. This is an undocumented(Almost) x87 instruction.
       //  https://en.wikipedia.org/wiki/X86_instruction_listings#Undocumented_x87_instructions
@@ -243,28 +769,52 @@ std::array<X86InstInfo, MAX_X87_TABLE_SIZE> X87Ops = []() consteval {
       //  At some point the Nvidia OpenGL binary driver uses this instruction.
       //  GCC may also end up emitting this instruction in some rare edge case!
       //  Almost all x86 CPUs implement this, and it is expected to be around
-      {OPD(0xDF, 0xC0), 8, X86InstInfo{"FFREEP",  TYPE_X87, FLAGS_POP, 0, nullptr}},
+      {OPD(0xDF, 0xC0), 8, X86InstInfo{"FFREEP",  TYPE_X87, FLAGS_POP, 0}},
       //  / 1
-      {OPD(0xDF, 0xC8), 8, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDF, 0xC8), 8, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 2
-      {OPD(0xDF, 0xD0), 8, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDF, 0xD0), 8, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 3
-      {OPD(0xDF, 0xD8), 8, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDF, 0xD8), 8, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 4
-      {OPD(0xDF, 0xE0), 1, X86InstInfo{"FNSTSW",  TYPE_INST, GenFlagsSameSize(SIZE_16BIT) | FLAGS_SF_DST_RAX, 0, nullptr}},
-      {OPD(0xDF, 0xE1), 7, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDF, 0xE0), 1, X86InstInfo{"FNSTSW",  TYPE_INST, GenFlagsSameSize(SIZE_16BIT) | FLAGS_SF_DST_RAX, 0}},
+      {OPD(0xDF, 0xE1), 7, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0}},
       //  / 5
-      {OPD(0xDF, 0xE8), 8, X86InstInfo{"FUCOMIP", TYPE_INST,    FLAGS_POP, 0, nullptr}},
+      {OPD(0xDF, 0xE8), 8, X86InstInfo{"FUCOMIP", TYPE_INST,    FLAGS_POP, 0}},
       //  / 6
-      {OPD(0xDF, 0xF0), 8, X86InstInfo{"FCOMIP",  TYPE_X87,   FLAGS_POP, 0, nullptr}},
+      {OPD(0xDF, 0xF0), 8, X86InstInfo{"FCOMIP",  TYPE_X87,   FLAGS_POP, 0}},
       //  / 7
-      {OPD(0xDF, 0xF8), 8, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0, nullptr}},
+      {OPD(0xDF, 0xF8), 8, X86InstInfo{"",        TYPE_INVALID, FLAGS_NONE, 0}},
   };
 #undef OPD
 #undef OPDReg
 
-  GenerateX87Table(&Table.at(0), X87OpTable, std::size(X87OpTable));
+  auto InstallToX87Table = [](auto& FinalTable, auto& LocalTable) {
+    for (auto Op : LocalTable) {
+      auto OpNum = Op.Op;
+      bool Repeat = (OpNum & 0x8000) != 0;
+      OpNum = OpNum & 0x7FF;
+      auto Dispatcher = Op.Ptr;
+      for (uint8_t i = 0; i < Op.Count; ++i) {
+        LOGMAN_THROW_A_FMT(FinalTable[OpNum + i].OpcodeDispatcher.OpDispatch == nullptr, "Duplicate Entry");
+
+        FinalTable[OpNum + i].OpcodeDispatcher.OpDispatch = Dispatcher;
+
+        // Flag to indicate if we need to repeat this op in {0x40, 0x80} ranges
+        if (Repeat) {
+          FinalTable[(OpNum | 0x40) + i].OpcodeDispatcher.OpDispatch = Dispatcher;
+          FinalTable[(OpNum | 0x80) + i].OpcodeDispatcher.OpDispatch = Dispatcher;
+        }
+      }
+    }
+  };
+
+  GenerateX87Table(Table.data(), X87OpTable, std::size(X87OpTable));
+  InstallToX87Table(Table, DispatchTable);
   return Table;
-}();
+};
+
+constexpr std::array<X86InstInfo, MAX_X87_TABLE_SIZE> X87F80Ops = GenerateX87TableLambda(X87F80OpTable);
+constexpr std::array<X86InstInfo, MAX_X87_TABLE_SIZE> X87F64Ops = GenerateX87TableLambda(X87F64OpTable);
 
 }

FEXCore/Source/Interface/GDBJIT/GDBJIT.cpp

@@ -90,7 +90,7 @@ void GDBJITRegister(FEXCore::IR::AOTIRCacheEntry* Entry, uintptr_t VAFileStart,
     mem += info->nlines * sizeof(gdb_line_mapping);
 
     if (info->nlines) {
-      memcpy(lines, &Lines.at(0), info->nlines * sizeof(gdb_line_mapping));
+      memcpy(lines, Lines.data(), info->nlines * sizeof(gdb_line_mapping));
     }
 
     auto entry = new jit_code_entry {0, 0, 0, 0};

FEXCore/Source/Interface/IR/AOTIR.cpp

@@ -2,311 +2,23 @@
 #include "FEXHeaderUtils/Filesystem.h"
 #include "Interface/Context/Context.h"
 #include "Interface/IR/AOTIR.h"
-#include "Interface/IR/IntrusiveIRList.h"
-#include "Interface/IR/RegisterAllocationData.h"
 
 #include <FEXCore/Utils/Allocator.h>
 #include <FEXCore/HLE/SyscallHandler.h>
 #include <FEXCore/fextl/fmt.h>
-#include <FEXCore/fextl/string.h>
 
 #include <Interface/Core/LookupCache.h>
 #include <Interface/GDBJIT/GDBJIT.h>
 
-#include <cstddef>
-#include <cstdint>
-#include <mutex>
-#include <sys/stat.h>
-#include <unistd.h>
 #include <xxhash.h>
 
-
 namespace FEXCore::IR {
-AOTIRInlineEntry* AOTIRInlineIndex::GetInlineEntry(uint64_t DataOffset) {
-  uintptr_t This = (uintptr_t)this;
-
-  return (AOTIRInlineEntry*)(This + DataBase + DataOffset);
-}
-
-AOTIRInlineEntry* AOTIRInlineIndex::Find(uint64_t GuestStart) {
-  ssize_t l = 0;
-  ssize_t r = Count - 1;
-
-  while (l <= r) {
-    size_t m = l + (r - l) / 2;
-
-    if (Entries[m].GuestStart == GuestStart) {
-      return GetInlineEntry(Entries[m].DataOffset);
-    } else if (Entries[m].GuestStart < GuestStart) {
-      l = m + 1;
-    } else {
-      r = m - 1;
-    }
-  }
-
-  return nullptr;
-}
-
-IR::IRListView* AOTIRInlineEntry::GetIRData() {
-  return (IR::IRListView*)InlineData;
-}
-
-void AOTIRCaptureCacheEntry::AppendAOTIRCaptureCache(uint64_t GuestRIP, uint64_t Start, uint64_t Length, uint64_t Hash,
-                                                     const FEXCore::IR::IRListView& IRList) {
-  auto Inserted = Index.emplace(GuestRIP, Stream->Offset());
-
-  if (Inserted.second) {
-    AOTIRInlineEntry entry {
-      .GuestHash = Hash,
-      .GuestLength = Length,
-    };
-    Stream->Write((const char*)&entry, sizeof(entry));
-
-    // IRData (inline)
-    IRList.Serialize(*Stream);
-  }
-}
-
-static bool readAll(int fd, void* data, size_t size) {
-  int rv = read(fd, data, size);
-
-  if (rv != size) {
-    return false;
-  } else {
-    return true;
-  }
-}
-
-static bool LoadAOTIRCache(AOTIRCacheEntry* Entry, int streamfd) {
-#ifndef _WIN32
-  uint64_t tag;
-
-  if (!readAll(streamfd, (char*)&tag, sizeof(tag)) || tag != FEXCore::IR::AOTIR_COOKIE) {
-    return false;
-  }
-
-  fextl::string Module;
-  uint64_t ModSize;
-  uint64_t IndexSize;
-
-  lseek(streamfd, -sizeof(ModSize), SEEK_END);
-
-  if (!readAll(streamfd, (char*)&ModSize, sizeof(ModSize))) {
-    return false;
-  }
-
-  Module.resize(ModSize);
-
-  lseek(streamfd, -sizeof(ModSize) - ModSize, SEEK_END);
-
-  if (!readAll(streamfd, (char*)&Module[0], Module.size())) {
-    return false;
-  }
-
-  if (Entry->FileId != Module) {
-    return false;
-  }
-
-  lseek(streamfd, -sizeof(ModSize) - ModSize - sizeof(IndexSize), SEEK_END);
-
-  if (!readAll(streamfd, (char*)&IndexSize, sizeof(IndexSize))) {
-    return false;
-  }
-
-  struct stat fileinfo;
-  if (fstat(streamfd, &fileinfo) < 0) {
-    return false;
-  }
-  size_t Size = (fileinfo.st_size + 4095) & ~4095;
-
-  size_t IndexOffset = fileinfo.st_size - IndexSize - sizeof(ModSize) - ModSize - sizeof(IndexSize);
-
-  void* FilePtr = FEXCore::Allocator::mmap(nullptr, Size, PROT_READ, MAP_SHARED, streamfd, 0);
-
-  if (FilePtr == MAP_FAILED) {
-    return false;
-  }
-
-  auto Array = (AOTIRInlineIndex*)((char*)FilePtr + IndexOffset);
-
-  LOGMAN_THROW_A_FMT(Entry->Array == nullptr && Entry->FilePtr == nullptr, "Entry must not be initialized here");
-  Entry->Array = Array;
-  Entry->FilePtr = FilePtr;
-  Entry->Size = Size;
-
-  LogMan::Msg::DFmt("AOTIR: Module {} has {} functions", Module, Array->Count);
-
-  return true;
-#else
-  return false;
-#endif
-}
-
-void AOTIRCaptureCache::FinalizeAOTIRCache() {
-  AOTIRCaptureCacheWriteoutQueue_Flush();
-
-  std::unique_lock lk(AOTIRCacheLock);
-
-  for (auto& [String, Entry] : AOTIRCaptureCacheMap) {
-    if (!Entry.Stream) {
-      continue;
-    }
-
-    const auto ModSize = String.size();
-    auto& stream = Entry.Stream;
-
-    // pad to 32 bytes
-    constexpr char Zero = 0;
-    while (stream->Offset() & 31) {
-      stream->Write(&Zero, 1);
-    }
-
-    AOTIRInlineIndex index {
-      .Count = Entry.Index.size(),
-      .DataBase = -stream->Offset(),
-    };
-    stream->Write((const char*)&index, sizeof(index));
-
-    for (const auto& [GuestStart, DataOffset] : Entry.Index) {
-      AOTIRInlineIndexEntry entry {
-        .GuestStart = GuestStart,
-        .DataOffset = DataOffset,
-      };
-
-      stream->Write((const char*)&entry, sizeof(entry));
-    }
-
-    // End of file header
-    const auto IndexSize = sizeof(AOTIRInlineIndex) + index.Count * sizeof(FEXCore::IR::AOTIRInlineIndexEntry);
-    stream->Write((const char*)&IndexSize, sizeof(IndexSize));
-    stream->Write(String.c_str(), ModSize);
-    stream->Write((const char*)&ModSize, sizeof(ModSize));
-
-    // Close the stream
-    stream->Close();
-
-    // Rename the file to atomically update the cache with the temporary file
-    AOTIRRenamer(String);
-  }
-}
-
-void AOTIRCaptureCache::AOTIRCaptureCacheWriteoutQueue_Flush() {
-  {
-    std::shared_lock lk {AOTIRCaptureCacheWriteoutLock};
-    if (AOTIRCaptureCacheWriteoutQueue.size() == 0) {
-      AOTIRCaptureCacheWriteoutFlusing.store(false);
-      return;
-    }
-  }
-
-  for (;;) {
-    // This code is tricky to refactor so it doesn't allocate memory through glibc.
-    // The moved std::function object deallocates memory at the end of scope.
-    FEXCore::Allocator::YesIKnowImNotSupposedToUseTheGlibcAllocator glibc;
-
-    AOTIRCaptureCacheWriteoutLock.lock();
-    WriteOutFn fn = std::move(AOTIRCaptureCacheWriteoutQueue.front());
-    bool MaybeEmpty = false;
-    AOTIRCaptureCacheWriteoutQueue.pop();
-    MaybeEmpty = AOTIRCaptureCacheWriteoutQueue.size() == 0;
-    AOTIRCaptureCacheWriteoutLock.unlock();
-
-    fn();
-    if (MaybeEmpty) {
-      std::shared_lock lk {AOTIRCaptureCacheWriteoutLock};
-      if (AOTIRCaptureCacheWriteoutQueue.size() == 0) {
-        AOTIRCaptureCacheWriteoutFlusing.store(false);
-        return;
-      }
-    }
-  }
-
-  LOGMAN_MSG_A_FMT("Must never get here");
-}
-
-void AOTIRCaptureCache::AOTIRCaptureCacheWriteoutQueue_Append(const WriteOutFn& fn) {
-  bool Flush = false;
-
-  {
-    std::unique_lock lk {AOTIRCaptureCacheWriteoutLock};
-    AOTIRCaptureCacheWriteoutQueue.push(fn);
-    if (AOTIRCaptureCacheWriteoutQueue.size() > 10000) {
-      Flush = true;
-    }
-  }
-
-  bool test_val = false;
-  if (Flush && AOTIRCaptureCacheWriteoutFlusing.compare_exchange_strong(test_val, true)) {
-    AOTIRCaptureCacheWriteoutQueue_Flush();
-  }
-}
-
-void AOTIRCaptureCache::WriteFilesWithCode(const Context::AOTIRCodeFileWriterFn& Writer) {
-  std::shared_lock lk(AOTIRCacheLock);
-  for (const auto& Entry : AOTIRCache) {
-    if (Entry.second.ContainsCode) {
-      Writer(Entry.second.FileId, Entry.second.Filename);
-    }
-  }
-}
-
-// IRStorageBase with memory owned by IR cache
-class IRInlineStorage : public IRStorageBase {
-  AOTIRInlineEntry& entry;
-public:
-  IRInlineStorage(AOTIRInlineEntry& entry)
-    : entry(entry) {}
-  IRListView GetIRView() override {
-    return entry.GetIRData();
-  }
-};
-
-
-std::optional<AOTIRCaptureCache::PreGenerateIRFetchResult>
-AOTIRCaptureCache::PreGenerateIRFetch(FEXCore::Core::InternalThreadState* Thread, uint64_t GuestRIP) {
-  auto AOTIRCacheEntry = CTX->SyscallHandler->LookupAOTIRCacheEntry(Thread, GuestRIP);
-
-  PreGenerateIRFetchResult Result {};
-
-  if (AOTIRCacheEntry.Entry) {
-    AOTIRCacheEntry.Entry->ContainsCode = true;
-
-    if (CTX->Config.AOTIRLoad()) {
-      auto Mod = AOTIRCacheEntry.Entry->Array;
-
-      if (Mod != nullptr) {
-        auto AOTEntry = Mod->Find(GuestRIP - AOTIRCacheEntry.VAFileStart);
-
-        if (AOTEntry) {
-          // verify hash
-          auto MappedStart = GuestRIP;
-          auto hash = XXH3_64bits((void*)MappedStart, AOTEntry->GuestLength);
-          if (hash == AOTEntry->GuestHash) {
-            Result.IR = fextl::make_unique<IRInlineStorage>(*AOTEntry);
-            // LogMan::Msg::DFmt("using {} + {:x} -> {:x}\n", file->second.fileid, AOTEntry->first, GuestRIP);
-            Result.DebugData = new FEXCore::Core::DebugData();
-            Result.StartAddr = MappedStart;
-            Result.Length = AOTEntry->GuestLength;
-            return Result;
-          } else {
-            LogMan::Msg::IFmt("AOTIR: hash check failed {:x}\n", MappedStart);
-          }
-        } else {
-          // LogMan::Msg::IFmt("AOTIR: Failed to find {:x}, {:x}, {}\n", GuestRIP, GuestRIP - file->second.Start + file->second.Offset, file->second.fileid);
-        }
-      }
-    }
-  }
-
-  return std::nullopt;
-}
 
 bool AOTIRCaptureCache::PostCompileCode(FEXCore::Core::InternalThreadState* Thread, void* CodePtr, uint64_t GuestRIP, uint64_t StartAddr,
-                                        uint64_t Length, fextl::unique_ptr<FEXCore::IR::IRStorageBase> IR,
-                                        FEXCore::Core::DebugData* DebugData, bool GeneratedIR) {
+                                        uint64_t Length, FEXCore::Core::DebugData* DebugData) {
 
   // Both generated ir and LibraryJITName need a named region lookup
-  if (GeneratedIR || CTX->Config.LibraryJITNaming() || CTX->Config.GDBSymbols()) {
+  if (CTX->Config.LibraryJITNaming() || CTX->Config.GDBSymbols()) {
 
     auto AOTIRCacheEntry = CTX->SyscallHandler->LookupAOTIRCacheEntry(Thread, GuestRIP);
 
@@ -318,44 +30,7 @@ bool AOTIRCaptureCache::PostCompileCode(FEXCore::Core::InternalThreadState* Thre
       if (CTX->Config.GDBSymbols()) {
         GDBJITRegister(AOTIRCacheEntry.Entry, AOTIRCacheEntry.VAFileStart, GuestRIP, (uintptr_t)CodePtr, DebugData);
       }
-
-      // Add to AOT cache if aot generation is enabled
-      if (GeneratedIR && (CTX->Config.AOTIRCapture() || CTX->Config.AOTIRGenerate())) {
-
-        auto hash = XXH3_64bits((void*)StartAddr, Length);
-
-        auto LocalRIP = GuestRIP - AOTIRCacheEntry.VAFileStart;
-        auto LocalStartAddr = StartAddr - AOTIRCacheEntry.VAFileStart;
-        const auto& FileId = AOTIRCacheEntry.Entry->FileId;
-
-        // The lambda is converted to std::function. This is tricky to refactor so it doesn't allocate memory through glibc.
-        // NOTE: unique_ptr must be passed as a raw pointer since std::function requires lambda captures to be copyable
-        FEXCore::Allocator::YesIKnowImNotSupposedToUseTheGlibcAllocator glibc;
-        AOTIRCaptureCacheWriteoutQueue_Append([this, LocalRIP, LocalStartAddr, Length, hash, IRRaw = IR.release(), FileId]() {
-          fextl::unique_ptr<FEXCore::IR::IRStorageBase> IR(IRRaw);
-
-          // It is guaranteed via AOTIRCaptureCacheWriteoutLock and AOTIRCaptureCacheWriteoutFlusing that this will not run concurrently
-          // Memory coherency is guaranteed via AOTIRCaptureCacheWriteoutLock
-
-          auto* AotFile = &AOTIRCaptureCacheMap[FileId];
-
-          if (!AotFile->Stream) {
-            AotFile->Stream = AOTIRWriter(FileId);
-            uint64_t tag = FEXCore::IR::AOTIR_COOKIE;
-            AotFile->Stream->Write(&tag, sizeof(tag));
-          }
-          AotFile->AppendAOTIRCaptureCache(LocalRIP, LocalStartAddr, Length, hash, IR->GetIRView());
-        });
-
-        if (CTX->Config.AOTIRGenerate()) {
-          // cleanup memory and early exit here -- we're not running the application
-          Thread->CPUBackend->ClearCache();
-          return true;
-        }
-      }
     }
-
-    // Insert to caches if we generated IR
   }
 
   return false;
@@ -376,31 +51,10 @@ AOTIRCacheEntry* AOTIRCaptureCache::LoadAOTIRCacheEntry(const fextl::string& fil
     auto Inserted = AOTIRCache.insert({fileid, AOTIRCacheEntry {.FileId = fileid, .Filename = filename}});
     auto Entry = &(Inserted.first->second);
 
-    LOGMAN_THROW_A_FMT(Entry->Array == nullptr, "Duplicate LoadAOTIRCacheEntry");
-
-    if (CTX->Config.AOTIRLoad && AOTIRLoader) {
-      auto streamfd = AOTIRLoader(fileid);
-      if (streamfd != -1) {
-        FEXCore::IR::LoadAOTIRCache(Entry, streamfd);
-        close(streamfd);
-      }
-    }
     return Entry;
   }
 
   return nullptr;
 }
 
-void AOTIRCaptureCache::UnloadAOTIRCacheEntry(AOTIRCacheEntry* Entry) {
-#ifndef _WIN32
-  LOGMAN_THROW_A_FMT(Entry != nullptr, "Removing not existing entry");
-
-  if (Entry->Array) {
-    FEXCore::Allocator::munmap(Entry->FilePtr, Entry->Size);
-    Entry->Array = nullptr;
-    Entry->FilePtr = nullptr;
-    Entry->Size = 0;
-  }
-#endif
-}
 } // namespace FEXCore::IR

FEXCore/Source/Interface/IR/AOTIR.h

@@ -1,27 +1,21 @@
 // SPDX-License-Identifier: MIT
 #pragma once
 
-#include "Interface/IR/RegisterAllocationData.h"
-#include "Interface/IR/IntrusiveIRList.h"
-
-#include <FEXCore/Config/Config.h>
-#include <FEXCore/fextl/map.h>
-#include <FEXCore/fextl/string.h>
-#include <FEXCore/fextl/queue.h>
-#include <FEXCore/fextl/unordered_map.h>
-
-#include <atomic>
-#include <cstdint>
-#include <functional>
-#include <memory>
-#include <shared_mutex>
+#include <FEXCore/Utils/AllocatorHooks.h>
 #include <FEXCore/HLE/SourcecodeResolver.h>
+#include <FEXCore/fextl/string.h>
+#include <FEXCore/fextl/unordered_map.h>
+#include <FEXCore/fextl/vector.h>
+
+#include <cstdint>
+#include <shared_mutex>
 
 namespace FEXCore::CPU {
 union Relocation;
 } // namespace FEXCore::CPU
 
 namespace FEXCore::Core {
+struct InternalThreadState;
 struct DebugDataSubblock {
   uint32_t HostCodeOffset;
   uint32_t HostCodeSize;
@@ -50,123 +44,29 @@ class ContextImpl;
 }
 
 namespace FEXCore::IR {
-class IRListView;
-
-constexpr auto COOKIE_VERSION = [](const char CookieText[4], uint32_t Version) {
-  uint64_t Cookie = Version;
-  Cookie <<= 32;
-
-  // Make the cookie text be the lower bits
-  Cookie |= CookieText[3];
-  Cookie <<= 8;
-  Cookie |= CookieText[2];
-  Cookie <<= 8;
-  Cookie |= CookieText[1];
-  Cookie <<= 8;
-  Cookie |= CookieText[0];
-
-  return Cookie;
-};
-constexpr static uint32_t AOTIR_VERSION = 0x0000'00004;
-constexpr static uint64_t AOTIR_COOKIE = COOKIE_VERSION("FEXI", AOTIR_VERSION);
-
-struct AOTIRInlineEntry {
-  uint64_t GuestHash;
-  uint64_t GuestLength;
-
-  /* IRData */
-  uint8_t InlineData[0];
-
-  IR::IRListView* GetIRData();
-};
-
-struct AOTIRInlineIndexEntry {
-  uint64_t GuestStart;
-  uint64_t DataOffset;
-};
-
-struct AOTIRInlineIndex {
-  uint64_t Count;
-  uint64_t DataBase;
-  AOTIRInlineIndexEntry Entries[0];
-
-  AOTIRInlineEntry* Find(uint64_t GuestStart);
-  AOTIRInlineEntry* GetInlineEntry(uint64_t DataOffset);
-};
-
-struct AOTIRCaptureCacheEntry {
-  fextl::unique_ptr<FEXCore::Context::AOTIRWriter> Stream;
-  fextl::map<uint64_t, uint64_t> Index;
-
-  void AppendAOTIRCaptureCache(uint64_t GuestRIP, uint64_t Start, uint64_t Length, uint64_t Hash, const FEXCore::IR::IRListView& IRList);
-};
 
 struct AOTIRCacheEntry {
-  AOTIRInlineIndex* Array;
-  void* FilePtr;
-  size_t Size;
   fextl::unique_ptr<FEXCore::HLE::SourcecodeMap> SourcecodeMap;
   fextl::string FileId;
   fextl::string Filename;
-  bool ContainsCode;
 };
 
-using AOTCacheType = fextl::unordered_map<fextl::string, FEXCore::IR::AOTIRCacheEntry>;
-
 class AOTIRCaptureCache final {
 public:
-  using WriteOutFn = std::function<void()>;
-
   AOTIRCaptureCache(FEXCore::Context::ContextImpl* ctx)
     : CTX {ctx} {}
 
-  void FinalizeAOTIRCache();
-  void AOTIRCaptureCacheWriteoutQueue_Flush();
-  void AOTIRCaptureCacheWriteoutQueue_Append(const WriteOutFn& fn);
-  void WriteFilesWithCode(const Context::AOTIRCodeFileWriterFn& Writer);
-
-  struct PreGenerateIRFetchResult {
-    fextl::unique_ptr<IRStorageBase> IR;
-    FEXCore::Core::DebugData* DebugData {};
-    uint64_t StartAddr {};
-    uint64_t Length {};
-  };
-  [[nodiscard]]
-  std::optional<PreGenerateIRFetchResult> PreGenerateIRFetch(FEXCore::Core::InternalThreadState* Thread, uint64_t GuestRIP);
-
   bool PostCompileCode(FEXCore::Core::InternalThreadState* Thread, void* CodePtr, uint64_t GuestRIP, uint64_t StartAddr, uint64_t Length,
-                       fextl::unique_ptr<FEXCore::IR::IRStorageBase> IR, FEXCore::Core::DebugData* DebugData, bool GeneratedIR);
+                       FEXCore::Core::DebugData* DebugData);
 
   AOTIRCacheEntry* LoadAOTIRCacheEntry(const fextl::string& filename);
-  void UnloadAOTIRCacheEntry(AOTIRCacheEntry* Entry);
-
-  // Callbacks
-  void SetAOTIRLoader(Context::AOTIRLoaderCBFn CacheReader) {
-    AOTIRLoader = std::move(CacheReader);
-  }
-
-  void SetAOTIRWriter(Context::AOTIRWriterCBFn CacheWriter) {
-    AOTIRWriter = std::move(CacheWriter);
-  }
-
-  void SetAOTIRRenamer(Context::AOTIRRenamerCBFn CacheRenamer) {
-    AOTIRRenamer = std::move(CacheRenamer);
-  }
 
 private:
   FEXCore::Context::ContextImpl* CTX;
 
   std::shared_mutex AOTIRCacheLock;
-  std::shared_mutex AOTIRCaptureCacheWriteoutLock;
-  std::atomic<bool> AOTIRCaptureCacheWriteoutFlusing;
 
-  fextl::queue<WriteOutFn> AOTIRCaptureCacheWriteoutQueue;
-
-  FEXCore::IR::AOTCacheType AOTIRCache;
-
-  Context::AOTIRLoaderCBFn AOTIRLoader;
-  Context::AOTIRWriterCBFn AOTIRWriter;
-  Context::AOTIRRenamerCBFn AOTIRRenamer;
-  fextl::unordered_map<fextl::string, FEXCore::IR::AOTIRCaptureCacheEntry> AOTIRCaptureCacheMap;
+  fextl::unordered_map<fextl::string, FEXCore::IR::AOTIRCacheEntry> AOTIRCache;
 };
+
 } // namespace FEXCore::IR

FEXCore/Source/Interface/IR/IR.h

@@ -1,12 +1,15 @@
 // SPDX-License-Identifier: MIT
 #pragma once
 
+#include <FEXCore/Utils/EnumUtils.h>
 #include <FEXCore/Utils/ThreadPoolAllocator.h>
 #include <FEXCore/IR/IR.h>
 
 #include <FEXCore/fextl/memory.h>
 #include <FEXCore/fextl/sstream.h>
 
+#include <array>
+
 namespace FEXCore::IR {
 
 class OrderedNode;
@@ -621,11 +624,7 @@ enum class ShiftType : uint8_t {
   ROR,
 };
 
-enum class BranchHint : uint8_t {
-  None = 0,
-  Call,
-  Return,
-};
+enum class BranchHint : uint8_t { None = 0, Call, Return, CheckTF };
 
 
 // Converts a size stored as an integer in to an OpSize enum.
@@ -774,7 +773,7 @@ inline NodeID NodeWrapperBase<Type>::ID() const {
   return NodeID(NodeOffset / sizeof(IR::OrderedNode));
 }
 
-bool IsFragmentExit(FEXCore::IR::IROps Op);
+[[nodiscard]]
 bool IsBlockExit(FEXCore::IR::IROps Op);
 
 void Dump(fextl::stringstream* out, const IRListView* IR);

FEXCore/Source/Interface/IR/IR.json

@@ -159,7 +159,8 @@
     "NamedVectorConstant": "FEXCore::IR::NamedVectorConstant",
     "IndexNamedVectorConstant": "FEXCore::IR::IndexNamedVectorConstant",
     "ShiftType": "FEXCore::IR::ShiftType",
-    "BranchHint": "FEXCore::IR::BranchHint"
+    "BranchHint": "FEXCore::IR::BranchHint",
+    "Array16": "std::array<uint8_t, 16>"
   },
   "Ops": {
     "Misc": {
@@ -200,10 +201,9 @@
         "HasSideEffects": true
       },
 
-      "GPR = ValidateCode u64:$CodeOriginalLow, u64:$CodeOriginalhigh, i64:$Offset, u8:$CodeLength": {
+      "GPR = ValidateCode Array16:$CodeOriginal, GPR:$Address, u8:$CodeLength": {
         "HasSideEffects": true,
         "HasDest": true,
-        "ImplicitFlagClobber": true,
         "DestSize": "OpSize::i64Bit"
       },
 
@@ -291,6 +291,14 @@
         "Desc": [
           "Implement a low power wait attempting to sleep until RDTSC >= Upper:Lower.",
           "Will spuriously wake up."
+	]
+      },
+      "MonoBackpatcherWrite OpSize:$Size, GPR:$Value, GPR:$Addr": {
+        "HasSideEffects": true,
+        "Desc": [ "Writes and invalidates the target address with the invalidation mutex locked. This is a fault-avoiding",
+                  "replacement for the atomic SMC writes used in the mono callsite backpatcher." ],
+        "EmitValidation": [
+          "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
         ]
       }
     },
@@ -300,12 +308,14 @@
         "RAOverride": "0"
       },
       "CondJump SSA:$Cmp1, SSA:$Cmp2, SSA:$TrueBlock, SSA:$FalseBlock, CondClass:$Cond{{COND_NEQ}}, OpSize:$CompareSize{OpSize::iInvalid}, i1:$FromNZCV{false}": {
+        "Inline": ["", "AddSub"],
         "HasSideEffects": true,
         "RAOverride": "2"
       },
       "ExitFunction OpSize:#Size, GPR:$NewRIP, BranchHint:$Hint, GPR:$CallReturnAddress, SSA:$CallReturnBlock": {
         "Desc": ["Exits the current JIT function with a target RIP"
                 ],
+        "Inline": ["Any"],
         "HasSideEffects": true,
         "DestSize": "Size",
         "RAOverride": "2"
@@ -440,10 +450,10 @@
                  "Zero Extends if value's type is too small",
                  "Truncates if value's type is too large"
                 ],
+        "Inline": ["Zero", ""],
         "HasSideEffects": true,
         "DestSize": "ByteSize",
         "EmitValidation": [
-          "WalkFindRegClass($Value) == $Class",
           "($Class == GPRClass && (#ByteSize == IR::OpSize::i8Bit || #ByteSize == IR::OpSize::i16Bit || #ByteSize == IR::OpSize::i32Bit || #ByteSize == IR::OpSize::i64Bit)) || $Class == FPRClass",
           "($Class == FPRClass && (#ByteSize == IR::OpSize::i8Bit || #ByteSize == IR::OpSize::i16Bit || #ByteSize == IR::OpSize::i32Bit || #ByteSize == IR::OpSize::i64Bit || #ByteSize == IR::OpSize::i128Bit || #ByteSize == IR::OpSize::i256Bit)) || $Class == GPRClass",
           "!($Offset >= offsetof(Core::CPUState, gregs[0]) && $Offset < offsetof(Core::CPUState, gregs[16])) && \"Can't StoreContext to GPR\"",
@@ -535,6 +545,7 @@
       },
 
       "SSA = LoadMem RegisterClass:$Class, OpSize:#Size, GPR:$Addr, GPR:$Offset, OpSize:$Align, MemOffsetType:$OffsetType, u8:$OffsetScale": {
+        "Inline": ["", "Mem"],
         "DestSize": "Size"
       },
 
@@ -549,11 +560,9 @@
                   "Zero Extends if value's type is too small",
                   "Truncates if value's type is too large"
                 ],
+        "Inline": ["Zero", "", "Mem"],
         "HasSideEffects": true,
-        "DestSize": "Size",
-        "EmitValidation": [
-          "WalkFindRegClass($Value) == $Class"
-        ]
+        "DestSize": "Size"
       },
 
       "StoreMemPair RegisterClass:$Class, OpSize:#Size, SSA:$Value1, SSA:$Value2, GPR:$Addr, u32:$Offset": {
@@ -561,11 +570,9 @@
                   "Zero Extends if value's type is too small",
                   "Truncates if value's type is too large"
                 ],
+        "Inline": ["Zero", "Zero"],
         "HasSideEffects": true,
-        "DestSize": "Size",
-        "EmitValidation": [
-          "WalkFindRegClass($Value1) == $Class"
-        ]
+        "DestSize": "Size"
       },
 
       "StoreMemX87SVEOptPredicate OpSize:#RegisterSize, OpSize:#ElementSize, FPR:$Value, GPR:$Addr": {
@@ -585,6 +592,7 @@
       "SSA = LoadMemTSO RegisterClass:$Class, OpSize:#Size, GPR:$Addr, GPR:$Offset, OpSize:$Align, MemOffsetType:$OffsetType, u8:$OffsetScale": {
         "Desc": ["Does a x86 TSO compatible load from memory. Offset must be Invalid()."
                 ],
+        "Inline": ["", "Memtso"],
         "DestSize": "Size",
         "DynamicDispatch": true
       },
@@ -592,12 +600,10 @@
       "StoreMemTSO RegisterClass:$Class, OpSize:#Size, SSA:$Value, GPR:$Addr, GPR:$Offset, OpSize:$Align, MemOffsetType:$OffsetType, u8:$OffsetScale": {
         "Desc": ["Does a x86 TSO compatible store to memory. Offset must be Invalid()."
                 ],
+        "Inline": ["Zero", "", "Memtso"],
         "HasSideEffects": true,
         "DestSize": "Size",
-        "DynamicDispatch": true,
-        "EmitValidation": [
-          "WalkFindRegClass($Value) == $Class"
-        ]
+        "DynamicDispatch": true
       },
 
       "FPR = VLoadVectorMasked OpSize:#RegisterSize, OpSize:#ElementSize, FPR:$Mask, GPR:$Addr, GPR:$Offset, MemOffsetType:$OffsetType, u8:$OffsetScale": {
@@ -708,6 +714,7 @@
         "Desc": ["Duplicates behaviour of x86 STOS repeat",
                  "Returns the final address that gets generated without the prefix appended."
                 ],
+        "Inline": ["", "", "Zero", "", "Any"],
         "HasSideEffects": true,
         "DestSize": "OpSize::i64Bit"
       },
@@ -715,6 +722,7 @@
         "Desc": ["Duplicates behaviour of x86 MOVS repeat",
                  "Returns the final addresses after they have been incremented or decremented"
                 ],
+        "Inline": ["", "", "", "Any"],
         "HasSideEffects": true,
         "DestSize": "OpSize::i64Bit"
       },
@@ -752,9 +760,8 @@
       "Prefetch i1:$ForStore, i1:$Stream, i8:$CacheLevel, GPR:$Addr, GPR:$Offset, MemOffsetType:$OffsetType, u8:$OffsetScale": {
         "Desc": ["Does a cacheline prefetch operation"
                 ],
-        "EmitValidation": [
-          "CacheLevel > 0 && CacheLevel < 4"
-        ],
+        "Inline": ["", "Mem"],
+        "EmitValidation": ["CacheLevel > 0 && CacheLevel < 4"],
         "HasSideEffects": true,
         "DestSize": "OpSize::i64Bit"
       },
@@ -1072,6 +1079,7 @@
         "Desc": [ "Integer Add",
                   "Will truncate to 64 or 32bits"
                 ],
+        "Inline": ["", "LargeAddSub"],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
@@ -1081,6 +1089,7 @@
         "Desc": [ "Integer Add with carry",
                   "Will truncate to 64 or 32bits"
                 ],
+        "Inline": ["Zero", ""],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
@@ -1108,6 +1117,7 @@
       },
       "GPR = AddWithFlags OpSize:#Size, GPR:$Src1, GPR:$Src2": {
         "Desc": [ "Integer add. Truncates and sets NZCV per AddNZCV"],
+        "Inline": ["", "LargeAddSub"],
         "DestSize": "Size",
         "HasSideEffects": true,
         "EmitValidation": [
@@ -1116,6 +1126,7 @@
       },
       "AddNZCV OpSize:#Size, GPR:$Src1, GPR:$Src2": {
         "Desc": ["Set NZCV for the sum of two GPRs"],
+        "Inline": ["", "LargeAddSub"],
         "HasSideEffects": true,
         "DestSize": "Size"
       },
@@ -1142,10 +1153,12 @@
       },
       "RmifNZCV GPR:$Src, u8:$Rotate, u8:$Mask": {
         "Desc": ["Rotate, mask, and insert into NZCV on FlagM platforms"],
+        "Inline": ["Zero", ""],
         "HasSideEffects": true
       },
       "CondAddNZCV OpSize:#Size, GPR:$Src1, GPR:$Src2, CondClass:$Cond, u8:$FalseNZCV": {
         "Desc": ["If condition is true, set NZCV per sum of GPRs, else force NZCV to a constant."],
+        "Inline": ["Zero", "AddSub"],
         "HasSideEffects": true,
         "DestSize": "Size",
         "EmitValidation": [
@@ -1154,6 +1167,7 @@
       },
       "CondSubNZCV OpSize:#Size, GPR:$Src1, GPR:$Src2, CondClass:$Cond, u8:$FalseNZCV": {
         "Desc": ["If condition is true, set NZCV per difference of GPRs, else force NZCV to a constant."],
+        "Inline": ["Zero", "AddSub"],
         "HasSideEffects": true,
         "DestSize": "Size",
         "EmitValidation": [
@@ -1162,6 +1176,7 @@
       },
       "GPR = AdcWithFlags OpSize:#Size, GPR:$Src1, GPR:$Src2": {
         "Desc": ["Adds and set NZCV for the sum of two GPRs and carry-in given as NZCV"],
+        "Inline": ["Zero", ""],
         "HasSideEffects": true,
         "DestSize": "Size",
         "EmitValidation": [
@@ -1211,6 +1226,7 @@
         "Desc": [ "Integer Sub",
                   "Will truncate to 64 or 32bits"
                 ],
+        "Inline": ["SubtractZero", "LargeAddSub"],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
@@ -1228,6 +1244,7 @@
       },
       "GPR = SubWithFlags OpSize:#Size, GPR:$Src1, GPR:$Src2": {
         "Desc": [ "Integer Sub. Truncates and sets NZCV per SubNZCV"],
+        "Inline": ["SubtractZero", "LargeAddSub"],
         "DestSize": "Size",
         "HasSideEffects": true,
         "EmitValidation": [
@@ -1244,6 +1261,7 @@
         "Desc": ["Set NZCV for the difference of two GPRs. ",
                  "Carry flag uses arm64 definition, inverted x86.",
                  ""],
+        "Inline": ["Zero", "LargeAddSub"],
         "DestSize": "Size",
         "HasSideEffects": true
       },
@@ -1251,6 +1269,7 @@
         "Desc": ["Integer binary or"
                 ],
         "DestSize": "Size",
+        "Inline": ["", "Logical"],
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
         ]
@@ -1280,8 +1299,8 @@
         ]
       },
       "GPR = Xor OpSize:#Size, GPR:$Src1, GPR:$Src2": {
-        "Desc": ["Integer binary exclusive or"
-                ],
+        "Desc": ["Integer binary exclusive or"],
+        "Inline": ["", "Logical"],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
@@ -1302,8 +1321,15 @@
         ]
       },
       "GPR = And OpSize:#Size, GPR:$Src1, GPR:$Src2": {
-        "Desc": ["Integer binary and"
-                ],
+        "Desc": ["Integer binary and"],
+        "Inline": ["", "Logical"],
+        "DestSize": "Size",
+        "EmitValidation": [
+          "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
+        ]
+      },
+      "GPR = AndShift OpSize:#Size, GPR:$Src1, GPR:$Src2, ShiftType:$Shift{ShiftType::LSL}, u8:$ShiftAmount{0}": {
+        "Desc": [ "Integer binary and with shifted register"],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
@@ -1312,6 +1338,7 @@
       "GPR = AndWithFlags OpSize:#Size, GPR:$Src1, GPR:$Src2": {
         "Desc": ["Integer binary and"
                 ],
+        "Inline": ["", "Logical"],
         "DestSize": "Size",
         "TiedSource": 0,
         "HasSideEffects": true
@@ -1319,12 +1346,14 @@
       "GPR = Andn OpSize:#Size, GPR:$Src1, GPR:$Src2": {
         "Desc": ["Integer binary AND NOT. Performs the equivalent of Src1 & ~Src2"],
         "DestSize": "Size",
+        "Inline": ["", "Logical"],
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
         ]
       },
       "TestNZ OpSize:#Size, GPR:$Src1, GPR:$Src2": {
         "Desc": ["Set NZCV for the binary AND of two GPRs, setting N and Z accordingly and zeroing C and V"],
+        "Inline": ["", "Logical"],
         "DestSize": "Size",
         "HasSideEffects": true
       },
@@ -1334,24 +1363,24 @@
         "HasSideEffects": true
       },
       "GPR = Lshl OpSize:#Size, GPR:$Src1, GPR:$Src2": {
-        "Desc": ["Integer logical shift left"
-                ],
+        "Desc": ["Integer logical shift left"],
+        "Inline": ["", "Any"],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
         ]
       },
       "GPR = Lshr OpSize:#Size, GPR:$Src1, GPR:$Src2": {
-        "Desc": ["Integer logical shift right"
-                ],
+        "Desc": ["Integer logical shift right"],
+        "Inline": ["", "Any"],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
         ]
       },
       "GPR = Ashr OpSize:#Size, GPR:$Src1, GPR:$Src2": {
-        "Desc": ["Integer arithmetic shift right"
-                ],
+        "Desc": ["Integer arithmetic shift right"],
+        "Inline": ["", "Any"],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
@@ -1368,8 +1397,8 @@
         "HasSideEffects": true
       },
       "GPR = Ror OpSize:#Size, GPR:$Src1, GPR:$Src2": {
-        "Desc": ["Integer rotate right"
-                ],
+        "Desc": ["Integer rotate right"],
+        "Inline": ["", "Any"],
         "DestSize": "Size",
         "EmitValidation": [
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
@@ -1505,12 +1534,12 @@
                  "op:",
                  "Dest = Cmp1 <Cond> Cmp2 ? TrueVal : FalseVal"
                 ],
+        "Inline": ["", "AddSub", "", ""],
         "DestSize": "ResultSize",
         "ImplicitFlagClobber": true,
         "EmitValidation": [
           "CompareSize == FEXCore::IR::OpSize::i32Bit || CompareSize == FEXCore::IR::OpSize::i64Bit || CompareSize == FEXCore::IR::OpSize::i128Bit",
-          "ResultSize == FEXCore::IR::OpSize::i32Bit || ResultSize == FEXCore::IR::OpSize::i64Bit",
-          "WalkFindRegClass($Cmp1) == WalkFindRegClass($Cmp2)"
+          "ResultSize == FEXCore::IR::OpSize::i32Bit || ResultSize == FEXCore::IR::OpSize::i64Bit"
         ]
       },
       "GPR = Extr OpSize:#Size, GPR:$Upper, GPR:$Lower, u8:$LSB": {

FEXCore/Source/Interface/IR/IRDumper.cpp

@@ -10,6 +10,7 @@ $end_info$
 #include "Interface/IR/RegisterAllocationData.h"
 
 #include <FEXCore/IR/IR.h>
+#include <FEXCore/fextl/fmt.h>
 #include <FEXCore/fextl/sstream.h>
 
 #include <algorithm>
@@ -30,14 +31,11 @@ namespace FEXCore::IR {
 #include <FEXCore/IR/IRDefines.inc>
 
 static void PrintArg(fextl::stringstream* out, const IRListView*, const SHA256Sum& Arg) {
-  *out << "sha256:";
-  for (auto byte : Arg.data) {
-    *out << std::hex << std::setfill('0') << std::setw(2) << (unsigned int)byte;
-  }
+  *out << fextl::fmt::format("sha256:{:02x}", fmt::join(Arg.data, ""));
 }
 
 static void PrintArg(fextl::stringstream* out, const IRListView*, uint64_t Arg) {
-  *out << "#0x" << std::hex << Arg << std::dec;
+  *out << fextl::fmt::format("#{:#x}", Arg);
 }
 
 static void PrintArg(fextl::stringstream* out, const IRListView*, CondClassType Arg) {
@@ -225,6 +223,31 @@ static void PrintArg(fextl::stringstream* out, const IRListView*, FEXCore::IR::N
   }();
 }
 
+static void PrintArg(fextl::stringstream* out, const IRListView*, IndexNamedVectorConstant Arg) {
+  *out << [Arg] {
+    // clang-format off
+    switch (Arg) {
+    case IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_PSHUFLW:
+      return "pshuflw";
+    case IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_PSHUFHW:
+      return "pshufhw";
+    case IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_PSHUFD:
+      return "pshufd";
+    case IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_SHUFPS:
+      return "shufps";
+    case IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_DPPS_MASK:
+      return "dpps_mask";
+    case IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_DPPD_MASK:
+      return "dppd_mask";
+    case IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_PBLENDW:
+      return "pblendw";
+    default:
+      return "<Unknown Indexed Named Vector Constant>";
+    }
+    // clang-format on
+  }();
+}
+
 static void PrintArg(fextl::stringstream* out, const IRListView*, FEXCore::IR::OpSize Arg) {
   switch (Arg) {
   case OpSize::i8Bit: *out << "i8"; break;
@@ -267,7 +290,7 @@ static void PrintArg(fextl::stringstream* out, const IRListView*, FEXCore::IR::S
   }
 }
 
-static void PrintArg(fextl::stringstream* out, [[maybe_unused]] const IRListView* IR, FEXCore::IR::BranchHint Arg) {
+static void PrintArg(fextl::stringstream* out, const IRListView*, FEXCore::IR::BranchHint Arg) {
   switch (Arg) {
   case BranchHint::None: *out << "None"; break;
   case BranchHint::Call: *out << "Call"; break;
@@ -276,6 +299,10 @@ static void PrintArg(fextl::stringstream* out, [[maybe_unused]] const IRListView
   }
 }
 
+static void PrintArg(fextl::stringstream* out, const IRListView*, const std::array<uint8_t, 0x10>& Arg) {
+  *out << fextl::fmt::format("{:02x}", fmt::join(Arg, ""));
+}
+
 void Dump(fextl::stringstream* out, const IRListView* IR) {
   auto HeaderOp = IR->GetHeader();
 
@@ -288,12 +315,7 @@ void Dump(fextl::stringstream* out, const IRListView* IR) {
 
   ++CurrentIndent;
   AddIndent();
-  *out << "(%0) "
-       << "IRHeader ";
-  *out << "%" << HeaderOp->Blocks.ID() << ", ";
-  *out << "#" << std::dec << HeaderOp->OriginalRIP << ", ";
-  *out << "#" << std::dec << HeaderOp->BlockCount << ", ";
-  *out << "#" << std::dec << HeaderOp->NumHostInstructions << std::endl;
+  *out << fextl::fmt::format("(%0) IRHeader %{}, #{:#x}, #{}, #{}\n", HeaderOp->Blocks.ID(), HeaderOp->OriginalRIP, HeaderOp->BlockCount, HeaderOp->NumHostInstructions);
 
   for (auto [BlockNode, BlockHeader] : IR->GetBlocks()) {
     {

FEXCore/Source/Interface/IR/IREmitter.cpp

@@ -12,13 +12,12 @@ $end_info$
 #include <FEXCore/Utils/EnumUtils.h>
 #include <FEXCore/Utils/LogManager.h>
 
-#include <array>
-#include <stdint.h>
-#include <string.h>
+#include <cstdint>
+#include <cstring>
 
 namespace FEXCore::IR {
 
-bool IsFragmentExit(FEXCore::IR::IROps Op) {
+static bool IsFragmentExit(FEXCore::IR::IROps Op) {
   switch (Op) {
   case OP_EXITFUNCTION:
   case OP_BREAK: return true;

FEXCore/Source/Interface/IR/IREmitter.h

@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: MIT
 #pragma once
 
+#include "CodeEmitter/Emitter.h"
 #include "Interface/IR/IR.h"
 #include "Interface/IR/IntrusiveIRList.h"
 
@@ -23,8 +24,9 @@ class IREmitter {
   friend class FEXCore::IR::PassManager;
 
 public:
-  IREmitter(FEXCore::Utils::IntrusivePooledAllocator& ThreadAllocator)
-    : DualListData {ThreadAllocator, 8 * 1024 * 1024} {
+  IREmitter(FEXCore::Utils::IntrusivePooledAllocator& ThreadAllocator, bool SupportsTSOImm9)
+    : DualListData {ThreadAllocator, 8 * 1024 * 1024}
+    , SupportsTSOImm9(SupportsTSOImm9) {
     ReownOrClaimBuffer();
     ResetWorkingList();
   }
@@ -51,6 +53,56 @@ public:
 
   FEXCore::IR::RegisterClassType WalkFindRegClass(Ref Node);
 
+  // These inlining helpers are used by IRDefines.inc so define first.
+  Ref InlineMem(OpSize Size, Ref Offset, MemOffsetType OffsetType, uint8_t& OffsetScale, bool TSO = false) {
+    uint64_t Imm {};
+    if (OffsetType != MEM_OFFSET_SXTX || !IsValueConstant(WrapNode(Offset), &Imm)) {
+      return Offset;
+    }
+
+    // The immediate may be scaled in the IR, we need to correct for that.
+    Imm *= OffsetScale;
+
+    // Signed immediate unscaled 9-bit range for both regular and LRCPC2 ops.
+    bool IsSIMM9 = ((int64_t)Imm >= -256) && ((int64_t)Imm <= 255);
+    IsSIMM9 &= (SupportsTSOImm9 || !TSO);
+
+    // Extended offsets for regular loadstore only.
+    LOGMAN_THROW_A_FMT(Size >= IR::OpSize::i8Bit && Size <= IR::OpSize::i256Bit, "Must be sized");
+
+    bool IsExtended = (Imm & (IR::OpSizeToSize(Size) - 1)) == 0 && Imm / IR::OpSizeToSize(Size) <= 4095;
+    IsExtended &= !TSO;
+
+    if (IsSIMM9 || IsExtended) {
+      OffsetScale = 1;
+      return _InlineConstant(Imm);
+    } else {
+      return Offset;
+    }
+  }
+
+#define DEF_INLINE(Type, Variable, Filter)                          \
+  Ref Inline##Type(OpSize Size, Ref Source) {                       \
+    uint64_t Variable;                                              \
+    if (IsValueConstant(WrapNode(Source), &Variable) && (Filter)) { \
+      return _InlineConstant(Variable);                             \
+    } else {                                                        \
+      return Source;                                                \
+    }                                                               \
+  }
+
+  DEF_INLINE(Any, _, true)
+  DEF_INLINE(Zero, X, X == 0)
+  DEF_INLINE(AddSub, X, ARMEmitter::IsImmAddSub(X))
+  DEF_INLINE(LargeAddSub, X, ARMEmitter::IsImmAddSub(X) && Size >= OpSize::i32Bit);
+  DEF_INLINE(Logical, X, ARMEmitter::Emitter::IsImmLogical(X, std::max((int)IR::OpSizeAsBits(Size), 32)));
+
+  Ref InlineSubtractZero(OpSize Size, Ref Src1, Ref Src2) {
+    // Only inline a zero if we won't inline the other source.
+    return IsValueConstant(WrapNode(Src2)) ? Src1 : InlineZero(Size, Src1);
+  }
+#undef DEF_INLINE
+
 // These handlers add cost to the constructor and destructor
 // If it becomes an issue then blow them away
 // GCC also generates some pretty atrocious code around these
@@ -82,6 +134,66 @@ public:
     return _StoreMem(Class, Size, Value, Addr, Invalid(), Align, MEM_OFFSET_SXTX, 1);
   }
 
+  IRPair<IROp_Select> Select01(FEXCore::IR::OpSize CompareSize, CondClassType Cond, OrderedNode* Cmp1, OrderedNode* Cmp2) {
+    return _Select(OpSize::i64Bit, CompareSize, Cond, Cmp1, Cmp2, _InlineConstant(1), _InlineConstant(0));
+  }
+
+  IRPair<IROp_Select> To01(FEXCore::IR::OpSize CompareSize, OrderedNode* Cmp1) {
+    return Select01(CompareSize, CondClassType {COND_NEQ}, Cmp1, Constant(0));
+  }
+
+  IRPair<IROp_NZCVSelect> _NZCVSelect01(CondClassType Cond) {
+    return _NZCVSelect(OpSize::i64Bit, Cond, _InlineConstant(1), _InlineConstant(0));
+  }
+
+  Ref Addsub(IR::OpSize Size, IROps Op, IROps NegatedOp, Ref Src1, uint64_t Src2) {
+    // Sign-extend the constant
+    if (Size == OpSize::i32Bit) {
+      Src2 = (int64_t)(int32_t)Src2;
+    }
+
+    // Negative constants need to be negated to inline.
+    if (Src2 & (1ull << 63) && ARMEmitter::IsImmAddSub(-Src2)) {
+      Op = NegatedOp;
+      Src2 = -Src2;
+    }
+
+    auto Dest = _Add(Size, Src1, Constant(Src2));
+    Dest.first->Header.Op = Op;
+    return Dest;
+  }
+
+  Ref Add(IR::OpSize Size, Ref Src1, uint64_t Src2) {
+    return Addsub(Size, OP_ADD, OP_SUB, Src1, Src2);
+  }
+
+  Ref Sub(IR::OpSize Size, Ref Src1, uint64_t Src2) {
+    return Addsub(Size, OP_SUB, OP_ADD, Src1, Src2);
+  }
+
+  Ref AddWithFlags(IR::OpSize Size, Ref Src1, uint64_t Src2) {
+    return Addsub(Size, OP_ADDWITHFLAGS, OP_SUBWITHFLAGS, Src1, Src2);
+  }
+
+  Ref SubWithFlags(IR::OpSize Size, Ref Src1, uint64_t Src2) {
+    return Addsub(Size, OP_SUBWITHFLAGS, OP_ADDWITHFLAGS, Src1, Src2);
+  }
+
+#define DEF_ADDSUB(Op)                                \
+  Ref Op(IR::OpSize Size, Ref Src1, Ref Src2) {       \
+    uint64_t Constant;                                \
+    if (IsValueConstant(WrapNode(Src2), &Constant)) { \
+      return Op(Size, Src1, Constant);                \
+    } else {                                          \
+      return _##Op(Size, Src1, Src2);                 \
+    }                                                 \
+  }
+
+  DEF_ADDSUB(Add)
+  DEF_ADDSUB(Sub)
+  DEF_ADDSUB(AddWithFlags)
+  DEF_ADDSUB(SubWithFlags)
+
   int64_t Constants[32];
   Ref ConstantRefs[32];
   uint32_t NrConstants;
@@ -206,9 +318,6 @@ public:
   void Remove(Ref Node);
   void RemovePostRA(Ref Node);
 
-  void SetPackedRFLAG(bool Lower8, Ref Src);
-  Ref GetPackedRFLAG(bool Lower8);
-
   void CopyData(const IREmitter& rhs) {
     LOGMAN_THROW_A_FMT(rhs.DualListData.DataBackingSize() <= DualListData.DataBackingSize(), "Trying to take ownership of data that is too "
                                                                                              "large");
@@ -284,12 +393,10 @@ public:
    * @{ */
   /**  @} */
   void LinkCodeBlocks(Ref CodeNode, Ref Next) {
+    [[maybe_unused]] auto CurrentIROp = CodeNode->Op(DualListData.DataBegin())->CW<FEXCore::IR::IROp_CodeBlock>();
 #if defined(ASSERTIONS_ENABLED) && ASSERTIONS_ENABLED
-    FEXCore::IR::IROp_CodeBlock* CurrentIROp =
-#endif
-      CodeNode->Op(DualListData.DataBegin())->CW<FEXCore::IR::IROp_CodeBlock>();
-
     LOGMAN_THROW_A_FMT(CurrentIROp->Header.Op == IROps::OP_CODEBLOCK, "Invalid");
+#endif
 
     CodeNode->append(DualListData.ListBegin(), Next);
   }
@@ -348,6 +455,7 @@ protected:
   Ref CurrentCodeBlock {};
   fextl::vector<Ref> CodeBlocks;
   uint64_t Entry {};
+  bool SupportsTSOImm9 {};
 };
 
 } // namespace FEXCore::IR

FEXCore/Source/Interface/IR/IntrusiveIRList.h

@@ -3,17 +3,14 @@
 
 #include "Interface/IR/IR.h"
 
-#include <FEXCore/Core/Context.h>
 #include <FEXCore/Utils/Allocator.h>
 #include <FEXCore/Utils/LogManager.h>
 #include <FEXCore/Utils/ThreadPoolAllocator.h>
-#include <FEXCore/fextl/vector.h>
 
 #include <cstddef>
+#include <cstdint>
 #include <cstring>
-#include <tuple>
-#include <istream>
-#include <ostream>
+#include <type_traits>
 
 namespace FEXCore::IR {
 /**
@@ -156,44 +153,6 @@ public:
     , DataSize(DataSize_)
     , ListSize(ListSize_) {}
 
-  void Serialize(FEXCore::Context::AOTIRWriter& stream) const {
-    void* nul = nullptr;
-    // void *IRDataInternal;
-    stream.Write((const char*)&nul, sizeof(nul));
-    // void *ListDataInternal;
-    stream.Write((const char*)&nul, sizeof(nul));
-    // size_t DataSize;
-    stream.Write((const char*)&DataSize, sizeof(DataSize));
-    // size_t ListSize;
-    stream.Write((const char*)&ListSize, sizeof(ListSize));
-
-    // inline data
-    stream.Write((const char*)GetData(), DataSize);
-    stream.Write((const char*)GetListData(), ListSize);
-  }
-
-  void Serialize(uint8_t* ptr) const {
-    void* nul = nullptr;
-    // void *IRDataInternal;
-    memcpy(ptr, &nul, sizeof(nul));
-    ptr += sizeof(nul);
-    // void *ListDataInternal;
-    memcpy(ptr, &nul, sizeof(nul));
-    ptr += sizeof(nul);
-    // size_t DataSize;
-    memcpy(ptr, &DataSize, sizeof(DataSize));
-    ptr += sizeof(DataSize);
-    // size_t ListSize;
-    memcpy(ptr, &ListSize, sizeof(ListSize));
-    ptr += sizeof(ListSize);
-
-    // inline data
-    memcpy(ptr, (const void*)GetData(), DataSize);
-    ptr += DataSize;
-    memcpy(ptr, (const void*)GetListData(), ListSize);
-    ptr += ListSize;
-  }
-
   [[nodiscard]]
   size_t GetInlineSize() const {
     static_assert(sizeof(*this) == 32);
@@ -411,11 +370,4 @@ private:
   uint8_t InlineData[0];
 };
 
-class IRStorageBase {
-public:
-  virtual ~IRStorageBase() = default;
-
-  virtual IRListView GetIRView() = 0;
-};
-
 } // namespace FEXCore::IR

FEXCore/Source/Interface/IR/PassManager.cpp

@@ -71,7 +71,6 @@ void PassManager::AddDefaultPasses(FEXCore::Context::ContextImpl* ctx) {
 
   if (!DisablePasses()) {
     InsertPass(CreateX87StackOptimizationPass(ctx->HostFeatures, ctx->Config.Is64BitMode ? IR::OpSize::i64Bit : IR::OpSize::i32Bit));
-    InsertPass(CreateConstProp(ctx->HostFeatures.SupportsTSOImm9));
     InsertPass(CreateDeadFlagCalculationEliminination());
   }
 }

FEXCore/Source/Interface/IR/PassManager.h

@@ -43,7 +43,6 @@ protected:
 };
 
 class PassManager final {
-  friend class ConstProp;
 public:
   void AddDefaultPasses(FEXCore::Context::ContextImpl* ctx);
   void AddDefaultValidationPasses();

FEXCore/Source/Interface/IR/Passes.h

@@ -16,7 +16,6 @@ namespace FEXCore::IR {
 class Pass;
 class RegisterAllocationPass;
 
-fextl::unique_ptr<FEXCore::IR::Pass> CreateConstProp(bool SupportsTSOImm9);
 fextl::unique_ptr<FEXCore::IR::Pass> CreateDeadFlagCalculationEliminination();
 fextl::unique_ptr<FEXCore::IR::RegisterAllocationPass> CreateRegisterAllocationPass(const FEXCore::CPUIDEmu* CPUID);
 fextl::unique_ptr<FEXCore::IR::Pass> CreateX87StackOptimizationPass(const FEXCore::HostFeatures&, OpSize GPROpSize);

FEXCore/Source/Interface/IR/Passes/ConstProp.cpp

@@ -1,341 +0,0 @@
-// SPDX-License-Identifier: MIT
-/*
-$info$
-tags: ir|opts
-desc: ConstProp, ZExt elim, const pooling, fcmp reduction, const inlining
-$end_info$
-*/
-#include <CodeEmitter/Emitter.h>
-
-#include "Interface/IR/IREmitter.h"
-#include "Interface/IR/PassManager.h"
-
-#include <FEXCore/IR/IR.h>
-#include <FEXCore/HLE/SyscallHandler.h>
-#include <FEXCore/Utils/LogManager.h>
-#include <FEXCore/Utils/Profiler.h>
-#include <FEXCore/fextl/map.h>
-#include <FEXCore/fextl/unordered_map.h>
-
-#include <cstdint>
-#include <string.h>
-
-namespace FEXCore::IR {
-
-// aarch64 heuristics
-static bool IsImmLogical(uint64_t imm, unsigned width) {
-  if (width < 32) {
-    width = 32;
-  }
-  return ARMEmitter::Emitter::IsImmLogical(imm, width);
-}
-
-class ConstProp final : public FEXCore::IR::Pass {
-public:
-  explicit ConstProp(bool SupportsTSOImm9)
-    : SupportsTSOImm9 {SupportsTSOImm9} {}
-
-  void Run(IREmitter* IREmit) override;
-
-private:
-  void ConstantPropagation(IREmitter* IREmit, const IRListView& CurrentIR, Ref CodeNode, IROp_Header* IROp);
-
-  bool SupportsTSOImm9 {};
-
-  template<class F>
-  bool InlineIf(IREmitter* IREmit, const IRListView& CurrentIR, Ref CodeNode, IROp_Header* IROp, unsigned Index, F Filter) {
-    uint64_t Constant;
-    if (!IREmit->IsValueConstant(IROp->Args[Index], &Constant) || !Filter(Constant)) {
-      return false;
-    }
-
-    IREmit->SetWriteCursor(CurrentIR.GetNode(IROp->Args[Index]));
-    IREmit->ReplaceNodeArgument(CodeNode, Index, IREmit->_InlineConstant(Constant));
-    return true;
-  }
-
-  bool Inline(IREmitter* IREmit, const IRListView& CurrentIR, Ref CodeNode, IROp_Header* IROp, unsigned Index) {
-    return InlineIf(IREmit, CurrentIR, CodeNode, IROp, Index, [](uint64_t _) { return true; });
-  }
-
-  bool InlineIfZero(IREmitter* IREmit, const IRListView& CurrentIR, Ref CodeNode, IROp_Header* IROp, unsigned Index) {
-    return InlineIf(IREmit, CurrentIR, CodeNode, IROp, Index, [](uint64_t X) { return X == 0; });
-  }
-
-  bool InlineIfLargeAddSub(IREmitter* IREmit, const IRListView& CurrentIR, Ref CodeNode, IROp_Header* IROp, unsigned Index) {
-    // We don't allow 8/16-bit operations to have constants, since no
-    // constant would be in bounds after the JIT's 24/16 shift.
-    auto Filter = [&IROp](uint64_t X) {
-      return ARMEmitter::IsImmAddSub(X) && IROp->Size >= OpSize::i32Bit;
-    };
-
-    return InlineIf(IREmit, CurrentIR, CodeNode, IROp, Index, Filter);
-  }
-
-  void InlineMemImmediate(IREmitter* IREmit, const IRListView& IR, Ref CodeNode, IR::RegisterClassType RegisterClass, IROp_Header* IROp,
-                          OrderedNodeWrapper Offset, MemOffsetType OffsetType, const size_t Offset_Index, uint8_t& OffsetScale, bool TSO) {
-    uint64_t Imm {};
-    if (OffsetType != MEM_OFFSET_SXTX || !IREmit->IsValueConstant(Offset, &Imm)) {
-      return;
-    }
-
-    // The immediate may be scaled in the IR, we need to correct for that.
-    Imm *= OffsetScale;
-
-    // Signed immediate unscaled 9-bit range for both regular and LRCPC2 ops.
-    bool IsSIMM9 = ((int64_t)Imm >= -256) && ((int64_t)Imm <= 255);
-    IsSIMM9 &= (SupportsTSOImm9 || !TSO);
-
-    // Extended offsets for regular loadstore only.
-    LOGMAN_THROW_A_FMT(IROp->Size >= IR::OpSize::i8Bit && IROp->Size <= (RegisterClass == GPRClass ? IR::OpSize::i64Bit : IR::OpSize::i256Bit),
-                       "Invalid "
-                       "size");
-    bool IsExtended = (Imm & (IR::OpSizeToSize(IROp->Size) - 1)) == 0 && Imm / IR::OpSizeToSize(IROp->Size) <= 4095;
-    IsExtended &= !TSO;
-
-    if (IsSIMM9 || IsExtended) {
-      IREmit->SetWriteCursor(IR.GetNode(Offset));
-      IREmit->ReplaceNodeArgument(CodeNode, Offset_Index, IREmit->_InlineConstant(Imm));
-      OffsetScale = 1;
-    }
-  }
-};
-
-// constprop + some more per instruction logic
-void ConstProp::ConstantPropagation(IREmitter* IREmit, const IRListView& CurrentIR, Ref CodeNode, IROp_Header* IROp) {
-  switch (IROp->Op) {
-  case OP_ADD:
-  case OP_SUB:
-  case OP_ADDWITHFLAGS:
-  case OP_SUBWITHFLAGS: {
-    auto Op = IROp->C<IR::IROp_Add>();
-    uint64_t Constant1 {};
-    uint64_t Constant2 {};
-    bool IsConstant2 = IREmit->IsValueConstant(IROp->Args[1], &Constant2);
-
-    /* IsImmAddSub assumes the constants are sign-extended, take care of that
-     * here so we get the optimization for 32-bit adds too.
-     */
-    if (Op->Header.Size == OpSize::i32Bit) {
-      Constant1 = (int64_t)(int32_t)Constant1;
-      Constant2 = (int64_t)(int32_t)Constant2;
-    }
-
-    if (IsConstant2 && !ARMEmitter::IsImmAddSub(Constant2) && ARMEmitter::IsImmAddSub(-Constant2)) {
-      // If the second argument is constant, the immediate is not ImmAddSub, but when negated is.
-      // So, negate the operation to negate (and inline) the constant.
-      if (IROp->Op == OP_ADD) {
-        IROp->Op = OP_SUB;
-      } else if (IROp->Op == OP_SUB) {
-        IROp->Op = OP_ADD;
-      } else if (IROp->Op == OP_ADDWITHFLAGS) {
-        IROp->Op = OP_SUBWITHFLAGS;
-      } else if (IROp->Op == OP_SUBWITHFLAGS) {
-        IROp->Op = OP_ADDWITHFLAGS;
-      }
-
-      IREmit->SetWriteCursorBefore(CodeNode);
-
-      // Negate the constant.
-      auto NegConstant = IREmit->_Constant(-Constant2);
-
-      // Replace the second source with the negated constant.
-      IREmit->ReplaceNodeArgument(CodeNode, Op->Src2_Index, NegConstant);
-    }
-
-    if (!InlineIfLargeAddSub(IREmit, CurrentIR, CodeNode, IROp, 1) && (IROp->Op == OP_SUB || IROp->Op == OP_SUBWITHFLAGS)) {
-      // TODO: Generalize this
-      InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, 0);
-    }
-
-    break;
-  }
-  case OP_ADDNZCV: {
-    InlineIfLargeAddSub(IREmit, CurrentIR, CodeNode, IROp, 1);
-    break;
-  }
-  case OP_SUBNZCV: {
-    if (!InlineIfLargeAddSub(IREmit, CurrentIR, CodeNode, IROp, 1)) {
-      // TODO: Generalize this
-      InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, 0);
-    }
-    break;
-  }
-  case OP_AND:
-  case OP_OR:
-  case OP_XOR: {
-    InlineIf(IREmit, CurrentIR, CodeNode, IROp, 1, [&IROp](uint64_t X) { return IsImmLogical(X, IR::OpSizeAsBits(IROp->Size)); });
-    break;
-  }
-  case OP_ANDWITHFLAGS:
-  case OP_ANDN:
-  case OP_TESTNZ: {
-    InlineIf(IREmit, CurrentIR, CodeNode, IROp, 1, [&IROp](uint64_t X) { return IsImmLogical(X, IR::OpSizeAsBits(IROp->Size)); });
-    break;
-  }
-  case OP_ASHR:
-  case OP_ROR: {
-    Inline(IREmit, CurrentIR, CodeNode, IROp, 1);
-    break;
-  }
-  case OP_LSHL: {
-    Inline(IREmit, CurrentIR, CodeNode, IROp, 1);
-    break;
-  }
-  case OP_LSHR: {
-    Inline(IREmit, CurrentIR, CodeNode, IROp, 1);
-    break;
-  }
-  case OP_ADC:
-  case OP_ADCWITHFLAGS:
-  case OP_RMIFNZCV: {
-    InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, 0);
-    break;
-  }
-  case OP_STORECONTEXT: {
-    // For i128Bit, we won't see a normal Constant to inline, but as a special
-    // case we can replace with a 2x64-bit store which can use inline zeroes.
-    if (IROp->Size == OpSize::i128Bit) {
-      auto Op = IROp->C<IR::IROp_StoreContext>();
-      auto Header = IREmit->GetOpHeader(IROp->Args[0]);
-      const auto MAX_STP_OFFSET = (252 * 4);
-
-      if (Op->Offset <= MAX_STP_OFFSET && Header->Op == OP_LOADNAMEDVECTORCONSTANT) {
-        auto Const = Header->C<IR::IROp_LoadNamedVectorConstant>();
-
-        if (Const->Constant == IR::NamedVectorConstant::NAMED_VECTOR_ZERO) {
-          IREmit->SetWriteCursor(CodeNode);
-          Ref Zero = IREmit->_Constant(0);
-          Ref STP = IREmit->_StoreContextPair(IR::OpSize::i64Bit, GPRClass, Zero, Zero, Op->Offset);
-          IREmit->Remove(CodeNode);
-
-          // XXX: This works around InlineConstant not having an associated
-          // register class, else we'd just do InlineConstant above.
-          Ref InlineZero = IREmit->_InlineConstant(0);
-          IREmit->ReplaceNodeArgument(STP, 0, InlineZero);
-          IREmit->ReplaceNodeArgument(STP, 1, InlineZero);
-        }
-      }
-    } else {
-      InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, 0);
-    }
-
-    break;
-  }
-
-  case OP_CONDADDNZCV:
-  case OP_CONDSUBNZCV: {
-    InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, 0);
-    InlineIf(IREmit, CurrentIR, CodeNode, IROp, 1, ARMEmitter::IsImmAddSub);
-    break;
-  }
-  case OP_SELECT: {
-    InlineIf(IREmit, CurrentIR, CodeNode, IROp, 1, ARMEmitter::IsImmAddSub);
-
-    uint64_t AllOnes = IROp->Size == OpSize::i64Bit ? 0xffff'ffff'ffff'ffffull : 0xffff'ffffull;
-
-    uint64_t Constant2 {};
-    uint64_t Constant3 {};
-    if (IREmit->IsValueConstant(IROp->Args[2], &Constant2) && IREmit->IsValueConstant(IROp->Args[3], &Constant3) &&
-        (Constant2 == 1 || Constant2 == AllOnes) && Constant3 == 0) {
-      IREmit->SetWriteCursor(CurrentIR.GetNode(IROp->Args[2]));
-
-      IREmit->ReplaceNodeArgument(CodeNode, 2, IREmit->_InlineConstant(Constant2));
-      IREmit->ReplaceNodeArgument(CodeNode, 3, IREmit->_InlineConstant(Constant3));
-    }
-
-    break;
-  }
-  case OP_NZCVSELECT: {
-    // We always allow source 1 to be zero, but source 0 can only be a
-    // special 1/~0 constant if source 1 is 0.
-    if (InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, 1)) {
-      uint64_t AllOnes = IROp->Size == OpSize::i64Bit ? 0xffff'ffff'ffff'ffffull : 0xffff'ffffull;
-      InlineIf(IREmit, CurrentIR, CodeNode, IROp, 0, [&AllOnes](uint64_t X) { return X == 1 || X == AllOnes; });
-    }
-    break;
-  }
-  case OP_CONDJUMP: {
-    InlineIf(IREmit, CurrentIR, CodeNode, IROp, 1, ARMEmitter::IsImmAddSub);
-    break;
-  }
-  case OP_EXITFUNCTION: {
-    auto Op = IROp->C<IR::IROp_ExitFunction>();
-
-    if (!Inline(IREmit, CurrentIR, CodeNode, IROp, Op->NewRIP_Index)) {
-      auto NewRIP = IREmit->GetOpHeader(Op->NewRIP);
-      if (NewRIP->Op == OP_ENTRYPOINTOFFSET) {
-        auto EO = NewRIP->C<IR::IROp_EntrypointOffset>();
-        IREmit->SetWriteCursor(CurrentIR.GetNode(Op->NewRIP));
-
-        IREmit->ReplaceNodeArgument(CodeNode, 0, IREmit->_InlineEntrypointOffset(EO->Header.Size, EO->Offset));
-      }
-    }
-    break;
-  }
-
-  case OP_LOADMEM: {
-    auto Op = IROp->CW<IR::IROp_LoadMem>();
-    InlineMemImmediate(IREmit, CurrentIR, CodeNode, Op->Class, IROp, Op->Offset, Op->OffsetType, Op->Offset_Index, Op->OffsetScale, false);
-    break;
-  }
-  case OP_STOREMEM: {
-    auto Op = IROp->CW<IR::IROp_StoreMem>();
-    InlineMemImmediate(IREmit, CurrentIR, CodeNode, Op->Class, IROp, Op->Offset, Op->OffsetType, Op->Offset_Index, Op->OffsetScale, false);
-    InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, Op->Value_Index);
-    break;
-  }
-  case OP_PREFETCH: {
-    auto Op = IROp->CW<IR::IROp_Prefetch>();
-    InlineMemImmediate(IREmit, CurrentIR, CodeNode, GPRClass, IROp, Op->Offset, Op->OffsetType, Op->Offset_Index, Op->OffsetScale, false);
-    break;
-  }
-  case OP_LOADMEMTSO: {
-    auto Op = IROp->CW<IR::IROp_LoadMemTSO>();
-    InlineMemImmediate(IREmit, CurrentIR, CodeNode, Op->Class, IROp, Op->Offset, Op->OffsetType, Op->Offset_Index, Op->OffsetScale, true);
-    break;
-  }
-  case OP_STOREMEMTSO: {
-    auto Op = IROp->CW<IR::IROp_StoreMemTSO>();
-    InlineMemImmediate(IREmit, CurrentIR, CodeNode, Op->Class, IROp, Op->Offset, Op->OffsetType, Op->Offset_Index, Op->OffsetScale, true);
-    InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, Op->Value_Index);
-    break;
-  }
-  case OP_STOREMEMPAIR: {
-    auto Op = IROp->CW<IR::IROp_StoreMemPair>();
-    InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, Op->Value1_Index);
-    InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, Op->Value2_Index);
-    break;
-  }
-  case OP_MEMCPY: {
-    auto Op = IROp->CW<IR::IROp_MemCpy>();
-    Inline(IREmit, CurrentIR, CodeNode, IROp, Op->Direction_Index);
-    break;
-  }
-  case OP_MEMSET: {
-    auto Op = IROp->CW<IR::IROp_MemSet>();
-    Inline(IREmit, CurrentIR, CodeNode, IROp, Op->Direction_Index);
-    InlineIfZero(IREmit, CurrentIR, CodeNode, IROp, Op->Value_Index);
-    break;
-  }
-
-  default: break;
-  }
-}
-
-void ConstProp::Run(IREmitter* IREmit) {
-  FEXCORE_PROFILE_SCOPED("PassManager::ConstProp");
-
-  auto CurrentIR = IREmit->ViewIR();
-
-  for (auto [BlockNode, BlockIROp] : CurrentIR.GetBlocks()) {
-    for (auto [CodeNode, IROp] : CurrentIR.GetCode(BlockNode)) {
-      ConstantPropagation(IREmit, CurrentIR, CodeNode, IROp);
-    }
-  }
-}
-
-fextl::unique_ptr<FEXCore::IR::Pass> CreateConstProp(bool SupportsTSOImm9) {
-  return fextl::make_unique<ConstProp>(SupportsTSOImm9);
-}
-} // namespace FEXCore::IR

FEXCore/Source/Interface/IR/Passes/RegisterAllocationPass.cpp

@@ -148,7 +148,8 @@ private:
   };
 
   bool HasSource(IROp_Header* I, PhysicalRegister Reg) {
-    for (auto s = 0; s < IR::GetRAArgs(I->Op); ++s) {
+    int NumArgs = IR::GetRAArgs(I->Op);
+    for (int s = 0; s < NumArgs; ++s) {
       if (I->Args[s].IsImmediate()) {
         // When spilling for a destination, we'll see register sources
         if (PhysicalRegister(I->Args[s]) == Reg) {
@@ -240,7 +241,7 @@ private:
       }
       // End of iteration gunk
 
-      const uint8_t NumArgs = IR::GetRAArgs(IROp->Op);
+      const int NumArgs = IR::GetRAArgs(IROp->Op);
       for (int i = NumArgs - 1; i >= 0; --i) {
         auto V = IROp->Args[i];
         V.ClearKill();
@@ -277,7 +278,7 @@ private:
     // next-use has the /smallest/ unsigned IP.
     Ref Candidate = nullptr;
     uint32_t BestDistance = UINT32_MAX;
-    [[maybe_unused]] uint8_t BestReg = ~0;
+    uint8_t BestReg = ~0;
     uint32_t Allocated = ((1u << Class->Count) - 1) & ~Class->Available;
 
     foreach_bit(i, Allocated) {
@@ -392,22 +393,20 @@ private:
     }
 
     // Try to coalesce reserved pairs. Just a heuristic to remove some moves.
-    if (IROp->Op == OP_ALLOCATEGPR) {
-      if (IROp->C<IROp_AllocateGPR>()->ForPair) {
-        uint32_t Available = Classes[GPRClass].Available;
+    if (IROp->Op == OP_ALLOCATEGPR && IROp->C<IROp_AllocateGPR>()->ForPair) {
+      uint32_t Available = Classes[GPRClass].Available;
 
-        // Only choose base register R if R and R + 1 are both free
-        Available &= (Available >> 1);
+      // Only choose base register R if R and R + 1 are both free
+      Available &= (Available >> 1);
 
-        // Only consider aligned registers in the pair region
-        constexpr uint32_t EVEN_BITS = 0x55555555;
-        Available &= (EVEN_BITS & ((1u << PairRegs) - 1));
+      // Only consider aligned registers in the pair region
+      constexpr uint32_t EVEN_BITS = 0x55555555;
+      Available &= (EVEN_BITS & ((1u << PairRegs) - 1));
 
-        if (Available) {
-          unsigned Reg = std::countr_zero(Available);
-          SetReg(CodeNode, PhysicalRegister(GPRClass, Reg));
-          return;
-        }
+      if (Available) {
+        unsigned Reg = std::countr_zero(Available);
+        SetReg(CodeNode, PhysicalRegister(GPRClass, Reg));
+        return;
       }
     } else if (IROp->Op == OP_ALLOCATEGPRAFTER) {
       uint32_t Available = Classes[GPRClass].Available;
@@ -550,7 +549,8 @@ bool ConstrainedRAPass::TryPostRAMerge(Ref LastNode, Ref CodeNode, IROp_Header*
   if (PhysicalRegister(LastNode) == PhysicalRegister(CodeNode) && KillMove(LastOp, IROp, LastNode, CodeNode)) {
     LOGMAN_THROW_A_FMT(!PhysicalRegister(CodeNode).IsInvalid(), "invariant");
 
-    for (auto s = 0; s < IR::GetRAArgs(IROp->Op); ++s) {
+    int NumArgs = IR::GetRAArgs(IROp->Op);
+    for (int s = 0; s < NumArgs; ++s) {
       if (IROp->Args[s].IsImmediate() && PhysicalRegister(IROp->Args[s]) == PhysicalRegister(LastNode)) {
         IROp->Args[s].SetImmediate(PhysicalRegister(LastOp->Args[0]).Raw);
       }
@@ -623,7 +623,7 @@ void ConstrainedRAPass::Run(IREmitter* IREmit_) {
           }
         }
 
-        const uint8_t NumArgs = IR::GetRAArgs(IROp->Op);
+        const int NumArgs = IR::GetRAArgs(IROp->Op);
         for (int i = NumArgs - 1; i >= 0; --i) {
           const auto& Arg = IROp->Args[i];
           if (!Arg.IsInvalid()) {
@@ -698,8 +698,10 @@ void ConstrainedRAPass::Run(IREmitter* IREmit_) {
       // the register file simultaneously.
       //
       // Also update next-use info, again only relevant if we've spilled.
+      int NumArgs = IR::GetRAArgs(IROp->Op);
+
       if (AnySpilledBeforeThisInstruction) {
-        for (auto s = 0; s < IR::GetRAArgs(IROp->Op); ++s) {
+        for (int s = 0; s < NumArgs; ++s) {
           auto V = IROp->Args[s];
           V.ClearKill();
 
@@ -725,7 +727,7 @@ void ConstrainedRAPass::Run(IREmitter* IREmit_) {
         }
       }
 
-      for (auto s = 0; s < IR::GetRAArgs(IROp->Op); ++s) {
+      for (int s = 0; s < NumArgs; ++s) {
         if (IROp->Args[s].IsInvalid()) {
           continue;
         }

FEXCore/Source/Interface/IR/Passes/x87StackOptimizationPass.cpp

@@ -410,7 +410,7 @@ inline Ref X87StackOptimization::GetOffsetTopWithCache_Slow(uint8_t Offset) {
 
   auto* OffsetTop = GetTopWithCache_Slow();
   if (Offset != 0) {
-    OffsetTop = IREmit->_And(OpSize::i32Bit, IREmit->_Add(OpSize::i32Bit, OffsetTop, GetConstant(Offset)), GetConstant(7));
+    OffsetTop = IREmit->_And(OpSize::i32Bit, IREmit->Add(OpSize::i32Bit, OffsetTop, Offset), GetConstant(7));
     // GetTopWithCache_Slow already sets the cache so we don't need to set it here for offset == 0
     TopOffsetCache[Offset] = OffsetTop;
   }
@@ -542,7 +542,7 @@ void X87StackOptimization::HandleBinopStack(IROps Op64, bool VFOp64, IROps Op80,
 inline void X87StackOptimization::UpdateTopForPop_Slow() {
   // Pop the top of the x87 stack
   auto* TopOffset = GetTopWithCache_Slow();
-  TopOffset = IREmit->_Add(OpSize::i32Bit, TopOffset, GetConstant(1));
+  TopOffset = IREmit->Add(OpSize::i32Bit, TopOffset, 1);
   TopOffset = IREmit->_And(OpSize::i32Bit, TopOffset, GetConstant(7));
   SetTopWithCache_Slow(TopOffset);
 }
@@ -550,7 +550,7 @@ inline void X87StackOptimization::UpdateTopForPop_Slow() {
 inline void X87StackOptimization::UpdateTopForPush_Slow() {
   // Pop the top of the x87 stack
   auto* TopOffset = GetTopWithCache_Slow();
-  TopOffset = IREmit->_Sub(OpSize::i32Bit, TopOffset, GetConstant(1));
+  TopOffset = IREmit->Sub(OpSize::i32Bit, TopOffset, 1);
   TopOffset = IREmit->_And(OpSize::i32Bit, TopOffset, GetConstant(7));
   SetTopWithCache_Slow(TopOffset);
 }
@@ -569,12 +569,7 @@ Ref X87StackOptimization::SynchronizeStackValues() {
 
   if (TopOffset != 0) {
     auto* OrigTop = GetTopWithCache_Slow();
-    Ref NewTop {};
-    if (TopOffset > 0) {
-      NewTop = IREmit->_And(OpSize::i32Bit, IREmit->_Sub(OpSize::i32Bit, OrigTop, GetConstant(TopOffset)), GetConstant(0x7));
-    } else {
-      NewTop = IREmit->_And(OpSize::i32Bit, IREmit->_Add(OpSize::i32Bit, OrigTop, GetConstant(-TopOffset)), GetConstant(0x7));
-    }
+    Ref NewTop = IREmit->_And(OpSize::i32Bit, IREmit->Sub(OpSize::i32Bit, OrigTop, TopOffset), GetConstant(0x7));
     SetTopWithCache_Slow(NewTop);
   }
   StackData.TopOffset = 0;
@@ -832,7 +827,7 @@ void X87StackOptimization::Run(IREmitter* Emit) {
             auto* TopValue = GetTopWithCache_Slow();
             if (Offset != 0) {
               auto* Mask = GetConstant(7);
-              TopValue = IREmit->_And(OpSize::i32Bit, IREmit->_Add(OpSize::i32Bit, TopValue, GetConstant(Offset)), Mask);
+              TopValue = IREmit->_And(OpSize::i32Bit, IREmit->Add(OpSize::i32Bit, TopValue, Offset), Mask);
             }
             SetX87ValidTag(TopValue, false);
           } else {

FEXCore/Source/Interface/IR/RegisterAllocationData.h

@@ -1,9 +1,8 @@
 // SPDX-License-Identifier: MIT
 #pragma once
+
 #include "IR.h"
-#include <FEXCore/Core/Context.h>
-#include <FEXCore/Utils/Allocator.h>
-#include <cstring>
+#include <cstdint>
 
 namespace FEXCore::IR {
 

FEXCore/Source/Utils/Allocator.cpp

@@ -3,15 +3,20 @@
 #include <FEXCore/Utils/Allocator.h>
 #include <FEXCore/Utils/CompilerDefs.h>
 #include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/Utils/MathUtils.h>
 #include <FEXCore/Utils/TypeDefines.h>
 #include <FEXCore/fextl/fmt.h>
+#include <FEXCore/fextl/memory.h>
 #include <FEXCore/fextl/memory_resource.h>
 #include <FEXHeaderUtils/Syscalls.h>
 
 #include <algorithm>
 #include <array>
 #include <cctype>
+#include <cerrno>
 #include <charconv>
+#include <cstddef>
+#include <cstdint>
 #include <cstdio>
 #include <fcntl.h>
 #ifndef _WIN32
@@ -19,11 +24,6 @@
 #include <sys/user.h>
 #endif
 
-#include <errno.h>
-#include <memory>
-#include <stddef.h>
-#include <stdint.h>
-
 namespace fextl::pmr {
 static fextl::pmr::default_resource FEXDefaultResource;
 std::pmr::memory_resource* get_default_resource() {

FEXCore/Source/Utils/Allocator/64BitAllocator.cpp

@@ -164,11 +164,8 @@ private:
       FEXCore::AlignUp(LiveVMARegion::GetFEXManagedVMARegionSize(ReservedRegion->RegionSize), FEXCore::Utils::FEX_PAGE_SIZE);
     size_t SizePlusManagedData = UsedSize + SizeOfLiveRegion;
 
-    [[maybe_unused]] auto Res = mprotect(reinterpret_cast<void*>(ReservedRegion->Base), SizePlusManagedData, PROT_READ | PROT_WRITE);
-
-    if (Res == -1) {
-      LOGMAN_MSG_A_FMT("Couldn't mprotect region: {} '{}' Likely occurs when running out of memory or Maximum VMAs", errno, strerror(errno));
-    }
+    auto Res = mprotect(reinterpret_cast<void*>(ReservedRegion->Base), SizePlusManagedData, PROT_READ | PROT_WRITE);
+    LOGMAN_THROW_A_FMT(Res != -1, "Couldn't mprotect region: {} '{}' Likely occurs when running out of memory or Maximum VMAs", errno, strerror(errno));
 
     LiveVMARegion* LiveRange = new (reinterpret_cast<void*>(ReservedRegion->Base)) LiveVMARegion();
 

FEXCore/Source/Utils/Allocator/HostAllocator.h

@@ -1,13 +1,15 @@
 // SPDX-License-Identifier: MIT
 #pragma once
-#include <FEXCore/fextl/allocator.h>
+
 #include <FEXCore/fextl/memory.h>
 #include <FEXCore/fextl/vector.h>
-#include <FEXCore/Utils/Allocator.h>
 
 #include <cstddef>
 #include <sys/types.h>
 
+namespace FEXCore::Allocator {
+struct MemoryRegion;
+}
 namespace FEXCore::Core {
 struct InternalThreadState;
 }

FEXCore/Source/Utils/Allocator/IntrusiveArenaAllocator.h

@@ -28,13 +28,13 @@ public:
 
   template<class U, class... Args>
   U* new_construct(Args&&... args) {
-    void* Ptr = do_allocate(sizeof(U), std::alignment_of<U>::value);
+    void* Ptr = do_allocate(sizeof(U), alignof(U));
     return new (Ptr) U(args...);
   }
 
   template<class U, class... Args>
   U* new_construct(U* Class, Args&&... args) {
-    void* Ptr = do_allocate(sizeof(U), std::alignment_of<U>::value);
+    void* Ptr = do_allocate(sizeof(U), alignof(U));
     return new (Ptr) U(args...);
   }
 
@@ -91,13 +91,13 @@ public:
 
   template<class U, class... Args>
   U* new_construct(Args&&... args) {
-    void* Ptr = do_allocate(sizeof(U), std::alignment_of<U>::value);
+    void* Ptr = do_allocate(sizeof(U), alignof(U));
     return new (Ptr) U(args...);
   }
 
   template<class U, class... Args>
   U* new_construct(U* Class, Args&&... args) {
-    void* Ptr = do_allocate(sizeof(U), std::alignment_of<U>::value);
+    void* Ptr = do_allocate(sizeof(U), alignof(U));
     return new (Ptr) U(args...);
   }