OpcodeDispatcher: Handle VMOVAPS

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

@@ -6986,6 +6986,9 @@ constexpr uint16_t PF_F2 = 3;
 
 #define OPD(map_select, pp, opcode) (((map_select - 1) << 10) | (pp << 8) | (opcode))
   constexpr std::tuple<uint16_t, uint8_t, FEXCore::X86Tables::OpDispatchPtr> VEXTable[] = {
+    {OPD(1, 0b00, 0x28), 1, &OpDispatchBuilder::VMOVAPSOp},
+    {OPD(1, 0b00, 0x29), 1, &OpDispatchBuilder::VMOVAPSOp},
+
     {OPD(1, 0b01, 0x6E), 2, &OpDispatchBuilder::UnimplementedOp},
 
     {OPD(1, 0b10, 0x6F), 1, &OpDispatchBuilder::UnimplementedOp},

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

@@ -404,6 +404,9 @@ public:
   // ADX Ops
   void ADXOp(OpcodeArgs);
 
+  // AVX Ops
+  void VMOVAPSOp(OpcodeArgs);
+
   // X87 Ops
   template<size_t width>
   void FLD(OpcodeArgs);

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

@@ -38,6 +38,18 @@ void OpDispatchBuilder::MOVAPSOp(OpcodeArgs) {
   StoreResult(FPRClass, Op, Src, -1);
 }
 
+void OpDispatchBuilder::VMOVAPSOp(OpcodeArgs) {
+  OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
+  const auto Is128BitDest = GetDstSize(Op) == Core::CPUState::XMM_SSE_REG_SIZE;
+
+  if (Op->Dest.IsGPR() && Is128BitDest) {
+    // Perform 32 byte store to clear the upper lane.
+    StoreResult_WithOpSize(FPRClass, Op, Op->Dest, Src, 32, -1);
+  } else {
+    StoreResult(FPRClass, Op, Src, -1);
+  }
+}
+
 void OpDispatchBuilder::MOVUPSOp(OpcodeArgs) {
   OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, 1);
   StoreResult(FPRClass, Op, Src, 1);

External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp

@@ -112,10 +112,10 @@ void InitializeVEXTables() {
     // This table doesn't state which VEX.pp is for which instruction
     // XXX: Confirm all the above encoding opcodes
 
-    {OPD(1, 0b00, 0x28), 1, X86InstInfo{"VMOVAPS",   TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
+    {OPD(1, 0b00, 0x28), 1, X86InstInfo{"VMOVAPS",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
     {OPD(1, 0b01, 0x28), 1, X86InstInfo{"VMOVAPD",   TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
 
-    {OPD(1, 0b00, 0x29), 1, X86InstInfo{"VMOVAPS",   TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
+    {OPD(1, 0b00, 0x29), 1, X86InstInfo{"VMOVAPS",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_XMM_FLAGS, 0, nullptr}},
     {OPD(1, 0b01, 0x29), 1, X86InstInfo{"VMOVAPD",   TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
 
     {OPD(1, 0b10, 0x2A), 1, X86InstInfo{"VCVTSI2SS",   TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},

unittests/ASM/VEX/vmovaps.asm

@@ -0,0 +1,26 @@
+%ifdef CONFIG
+{
+  "HostFeatures": ["AVX"],
+  "RegData": {
+      "XMM1": ["0xFFFFFFFFFFFFFFFF", "0xFFFFFFFFFFFFFFFF", "0xFFFFFFFFFFFFFFFF", "0xFFFFFFFFFFFFFFFF"],
+      "XMM2": ["0xFFFFFFFFFFFFFFFF", "0xFFFFFFFFFFFFFFFF", "0x0000000000000000", "0x0000000000000000"],
+      "XMM3": ["0xCCCCCCCCCCCCCCCC", "0xCCCCCCCCCCCCCCCC", "0xCCCCCCCCCCCCCCCC", "0xCCCCCCCCCCCCCCCC"]
+  }
+}
+%endif
+
+lea rdx, [rel .data]
+
+; Load inputs
+vmovaps ymm1, [rdx]
+vmovaps xmm2, [rdx]
+vmovaps ymm3, [rdx + 32]
+
+hlt
+
+align 32
+.data:
+db 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF
+db 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF
+db 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC
+db 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC

unittests/ASM/VEX/vmovaps_mem.asm

@@ -0,0 +1,49 @@
+%ifdef CONFIG
+{
+  "HostFeatures": ["AVX"],
+  "RegData": {
+    "RAX": "0x6162636465666768",
+    "RBX": "0x7172737475767778",
+    "XMM0": ["0x4142434445464748", "0x5152535455565758", "0x6162636465666768", "0x7172737475767778"],
+    "XMM1": ["0x4142434445464748", "0x5152535455565758", "0x6162636465666768", "0x7172737475767778"],
+    "XMM2": ["0xCCCCCCCCCCCCCCCC", "0xDDDDDDDDDDDDDDDD", "0x0000000000000000", "0x0000000000000000"]
+  },
+  "MemoryRegions": {
+    "0x100000000": "4096"
+  }
+}
+%endif
+
+mov rdx, 0xe0000000
+
+mov rax, 0x4142434445464748
+mov [rdx + 8 * 0], rax
+mov rax, 0x5152535455565758
+mov [rdx + 8 * 1], rax
+mov rax, 0x6162636465666768
+mov [rdx + 8 * 2], rax
+mov rax, 0x7172737475767778
+mov [rdx + 8 * 3], rax
+
+mov rax, 0xCCCCCCCCCCCCCCCC
+mov [rdx + 8 * 4], rax
+mov rax, 0xDDDDDDDDDDDDDDDD
+mov [rdx + 8 * 5], rax
+mov rax, 0xEEEEEEEEEEEEEEEE
+mov [rdx + 8 * 6], rax
+mov rax, 0xFFFFFFFFFFFFFFFF
+mov [rdx + 8 * 7], rax
+
+; Test truncation
+vmovaps ymm2, [rdx + 8 * 4]
+vmovaps xmm2, [rdx + 8 * 4]
+
+; Test memory overwrite
+vmovaps ymm0, [rdx]
+vmovaps [rdx + 8 * 4], ymm0
+vmovaps ymm1, ymm0
+
+mov rax, [rdx + 8 * 6]
+mov rbx, [rdx + 8 * 7]
+
+hlt