Merge pull request #3101 from alyssarosenzweig/opt/dec

Optimize out carry invert for DEC

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

@@ -205,7 +205,7 @@ DEF_OP(SubNZCV) {
       if (Result == 0) {
         NZCV |= 1U << 30;
       }
-      if (__builtin_usub_overflow(Src1, Src2, &Result)) {
+      if (__builtin_usub_overflow(Src1, Src2, &Result) ^ !(Op->InvertCarry)) {
         NZCV |= 1U << 29;
       }
       if (__builtin_ssub_overflow(Src1, Src2, &ResultSigned)) {
@@ -222,7 +222,7 @@ DEF_OP(SubNZCV) {
       if (Result == 0) {
         NZCV |= 1U << 30;
       }
-      if (__builtin_usubl_overflow(Src1, Src2, &Result)) {
+      if (__builtin_usubl_overflow(Src1, Src2, &Result) ^ !(Op->InvertCarry)) {
         NZCV |= 1U << 29;
       }
       if (__builtin_ssubl_overflow(Src1, Src2, &ResultSigned)) {

FEXCore/Source/Interface/Core/JIT/Arm64/ALUOps.cpp

@@ -154,13 +154,15 @@ DEF_OP(SubNZCV) {
   // TODO: Optimize this out
   mrs(Dst, ARMEmitter::SystemRegister::NZCV);
 
-  // The carry flag produced by arm64 subs is inverted compared to the x86 carry
-  // flag. Invert it now.
-  //
-  // TODO: Once we optimize out the mrs, this will become a cfinv operation, but
-  // that's only available with Feat_FlagM. For now the portable way is to flip
-  // bit 29 (carry) manually.
-  eor(ARMEmitter::Size::i32Bit, Dst, Dst, 1u << 29);
+  if (Op->InvertCarry) {
+    // The carry flag produced by arm64 subs is inverted compared to the x86 carry
+    // flag. Invert it now.
+    //
+    // TODO: Once we optimize out the mrs, this will become a cfinv operation, but
+    // that's only available with Feat_FlagM. For now the portable way is to flip
+    // bit 29 (carry) manually.
+    eor(ARMEmitter::Size::i32Bit, Dst, Dst, 1u << 29);
+  }
 }
 
 DEF_OP(Neg) {

FEXCore/Source/Interface/Core/JIT/x86_64/ALUOps.cpp

@@ -262,6 +262,10 @@ DEF_OP(SubNZCV) {
     break;
   }
 
+  if (!Op->InvertCarry) {
+    cmc();
+  }
+
   mov(TMP1, 0);
   mov(TMP2, 0);
   mov(TMP3, 0);

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

@@ -533,7 +533,8 @@ void OpDispatchBuilder::CalculateFlags_SUB(uint8_t SrcSize, OrderedNode *Res, Or
 
   // TODO: Could do this path for small sources if we have FEAT_FlagM
   if (SrcSize >= 4) {
-    SetNZCV(_SubNZCV(OpSize, Src1, Src2));
+    // We only bother inverting CF if we're actually going to update CF.
+    SetNZCV(_SubNZCV(OpSize, Src1, Src2, UpdateCF));
   } else {
     // SF/ZF
     SetNZ_ZeroCV(SrcSize, Res);

FEXCore/Source/Interface/IR/IR.json

@@ -901,9 +901,9 @@
           "Size == FEXCore::IR::OpSize::i32Bit || Size == FEXCore::IR::OpSize::i64Bit"
         ]
       },
-      "GPR = SubNZCV OpSize:$Size, GPR:$Src1, GPR:$Src2": {
+      "GPR = SubNZCV OpSize:$Size, GPR:$Src1, GPR:$Src2, u8:$InvertCarry": {
         "Desc": ["Return NZCV for the difference of two GPRs. ",
-                 "Note: Carry flag uses x86 definition, inverted from arm64.",
+                 "If InvertCarry is nonzero, carry flag uses x86 definition, inverted from arm64.",
                  ""],
         "DestSize": "4",
         "EmitValidation": [

unittests/InstructionCountCI/PrimaryGroup.json

@@ -3479,7 +3479,7 @@
       ]
     },
     "dec eax": {
-      "ExpectedInstructionCount": 12,
+      "ExpectedInstructionCount": 11,
       "Optimal": "No",
       "Comment": "GROUP4 0xfe /1",
       "ExpectedArm64ASM": [
@@ -3492,13 +3492,12 @@
         "ubfx w21, w21, #29, #1",
         "cmp w20, #0x1 (1)",
         "mrs x20, nzcv",
-        "eor w20, w20, #0x20000000",
         "bfi w20, w21, #29, #1",
         "str w20, [x28, #728]"
       ]
     },
     "dec rax": {
-      "ExpectedInstructionCount": 12,
+      "ExpectedInstructionCount": 11,
       "Optimal": "Yes",
       "Comment": "GROUP4 0xfe /1",
       "ExpectedArm64ASM": [
@@ -3511,7 +3510,6 @@
         "ubfx w21, w21, #29, #1",
         "cmp x20, #0x1 (1)",
         "mrs x20, nzcv",
-        "eor w20, w20, #0x20000000",
         "bfi w20, w21, #29, #1",
         "str w20, [x28, #728]"
       ]

unittests/InstructionCountCI/Primary_32Bit.json

@@ -360,7 +360,7 @@
       ]
     },
     "dec eax": {
-      "ExpectedInstructionCount": 12,
+      "ExpectedInstructionCount": 11,
       "Optimal": "No",
       "Comment": "0x48",
       "ExpectedArm64ASM": [
@@ -373,7 +373,6 @@
         "ubfx w21, w21, #29, #1",
         "cmp w20, #0x1 (1)",
         "mrs x20, nzcv",
-        "eor w20, w20, #0x20000000",
         "bfi w20, w21, #29, #1",
         "str w20, [x28, #728]"
       ]