Merge pull request #3756 from Sonicadvance1/fix_vmovhlps

Fix VMOVLHPS instruction
FEX-Emu · Jun 25, 2024 · 3a310b8 · 3a310b8
2 parents 7ff9622 + 3e8d780
commit 3a310b8
Show file tree

Hide file tree

Showing 5 changed files with 68 additions and 17 deletions.
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp
@@ -748,17 +748,26 @@ void OpDispatchBuilder::AVX128_MOVQ(OpcodeArgs) {
 void OpDispatchBuilder::AVX128_VMOVLP(OpcodeArgs) {
   auto Src1 = AVX128_LoadSource_WithOpSize(Op, Op->Src[0], Op->Flags, false);
 
-  if (Op->Dest.IsGPR()) {
-    auto Src2 = AVX128_LoadSource_WithOpSize(Op, Op->Src[1], Op->Flags, false);
-
+  if (!Op->Dest.IsGPR()) {
+    ///< VMOVLPS/PD mem64, xmm1
+    StoreResult_WithOpSize(FPRClass, Op, Op->Dest, Src1.Low, OpSize::i64Bit, OpSize::i64Bit);
+  } else if (!Op->Src[1].IsGPR()) {
+    ///< VMOVLPS/PD xmm1, xmm2, mem64
     // Bits[63:0] come from Src2[63:0]
     // Bits[127:64] come from Src1[127:64]
-    Ref Result_Low = _VInsElement(OpSize::i128Bit, OpSize::i64Bit, 1, 1, Src2.Low, Src1.Low);
+    auto Src2 = LoadSource_WithOpSize(FPRClass, Op, Op->Src[1], OpSize::i64Bit, Op->Flags);
+    Ref Result_Low = _VInsElement(OpSize::i128Bit, OpSize::i64Bit, 1, 1, Src2, Src1.Low);
     Ref ZeroVector = LoadZeroVector(OpSize::i128Bit);
 
     AVX128_StoreResult_WithOpSize(Op, Op->Dest, RefPair {.Low = Result_Low, .High = ZeroVector});
   } else {
-    StoreResult_WithOpSize(FPRClass, Op, Op->Dest, Src1.Low, OpSize::i64Bit, OpSize::i64Bit);
+    ///< VMOVHLPS/PD xmm1, xmm2, xmm3
+    auto Src2 = AVX128_LoadSource_WithOpSize(Op, Op->Src[1], Op->Flags, false);
+
+    Ref Result_Low = _VInsElement(OpSize::i128Bit, OpSize::i64Bit, 0, 1, Src1.Low, Src2.Low);
+    Ref ZeroVector = LoadZeroVector(OpSize::i128Bit);
+
+    AVX128_StoreResult_WithOpSize(Op, Op->Dest, RefPair {.Low = Result_Low, .High = ZeroVector});
   }
 }
 

diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp
@@ -132,15 +132,23 @@ void OpDispatchBuilder::MOVLPOp(OpcodeArgs) {
 }
 
 void OpDispatchBuilder::VMOVLPOp(OpcodeArgs) {
-  if (Op->Dest.IsGPR()) {
-    Ref Src1 = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, {.Align = 16});
+  Ref Src1 = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, {.Align = 16});
+
+  if (!Op->Dest.IsGPR()) {
+    ///< VMOVLPS/PD mem64, xmm1
+    StoreResult_WithOpSize(FPRClass, Op, Op->Dest, Src1, 8, 8);
+  } else if (!Op->Src[1].IsGPR()) {
+    ///< VMOVLPS/PD xmm1, xmm2, mem64
+    // Bits[63:0] come from Src2[63:0]
+    // Bits[127:64] come from Src1[127:64]
     Ref Src2 = LoadSource(FPRClass, Op, Op->Src[1], Op->Flags, {.Align = 8});
-    Ref Result = _VInsElement(16, 8, 0, 0, Src1, Src2);
-
+    Ref Result = _VInsElement(OpSize::i128Bit, OpSize::i64Bit, 1, 1, Src2, Src1);
     StoreResult(FPRClass, Op, Result, -1);
   } else {
-    Ref Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, {.Align = 8});
-    StoreResult_WithOpSize(FPRClass, Op, Op->Dest, Src, 8, 8);
+    ///< VMOVHLPS/PD xmm1, xmm2, xmm3
+    Ref Src2 = LoadSource(FPRClass, Op, Op->Src[1], Op->Flags, {.Align = 16});
+    Ref Result = _VInsElement(OpSize::i128Bit, OpSize::i64Bit, 0, 1, Src1, Src2);
+    StoreResult(FPRClass, Op, Result, -1);
   }
 }
 

diff --git a/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp b/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp
@@ -27,8 +27,8 @@ std::array<X86InstInfo, MAX_VEX_TABLE_SIZE> VEXTableOps = []() consteval {
     {OPD(1, 0b10, 0x11), 1, X86InstInfo{"VMOVSS",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}},
     {OPD(1, 0b11, 0x11), 1, X86InstInfo{"VMOVSD",    TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_DST | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}},
 
-    {OPD(1, 0b00, 0x12), 1, X86InstInfo{"VMOVLPS",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY | FLAGS_XMM_FLAGS | FLAGS_VEX_1ST_SRC, 0, nullptr}},
-    {OPD(1, 0b01, 0x12), 1, X86InstInfo{"VMOVLPD",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY | FLAGS_XMM_FLAGS | FLAGS_VEX_1ST_SRC, 0, nullptr}},
+    {OPD(1, 0b00, 0x12), 1, X86InstInfo{"VMOVLPS",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_VEX_1ST_SRC, 0, nullptr}},
+    {OPD(1, 0b01, 0x12), 1, X86InstInfo{"VMOVLPD",   TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_VEX_1ST_SRC, 0, nullptr}},
     {OPD(1, 0b10, 0x12), 1, X86InstInfo{"VMOVSLDUP", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
     {OPD(1, 0b11, 0x12), 1, X86InstInfo{"VMOVDDUP",  TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
 

diff --git a/unittests/ASM/VEX/vmovhlps.asm b/unittests/ASM/VEX/vmovhlps.asm
@@ -0,0 +1,34 @@
+%ifdef CONFIG
+{
+  "HostFeatures": ["AVX"],
+  "RegData": {
+      "XMM1": ["0x4150f0e342241b6c", "0xdddddddddddddddd", "0x0000000000000000", "0x0000000000000000"],
+      "XMM2": ["0xCCCCCCCCCCCCCCCC", "0xDDDDDDDDDDDDDDDD", "0xEEEEEEEEEEEEEEEE", "0xFFFFFFFFFFFFFFFF"],
+      "XMM3": ["0x4150f0e342241b6c", "0xdddddddddddddddd", "0x0000000000000000", "0x0000000000000000"],
+      "XMM5": ["0x428b029f42a63326", "0x4150f0e342241b6c", "0x41aff21340ab4706", "0x40aa5bea411ac802"],
+      "XMM6": ["0x428b029f42a63326", "0x4150f0e342241b6c", "0x41aff21340ab4706", "0x40aa5bea411ac802"]
+  }
+}
+%endif
+
+; Load inputs
+vmovapd ymm1, [rel .data]
+vmovapd ymm2, [rel .data + 32]
+vmovapd ymm5, [rel .data_random]
+vmovapd ymm6, [rel .data_random]
+
+vmovhlps xmm1, xmm2, xmm5
+vmovhlps xmm3, xmm1, xmm5
+
+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, 0xDD, 0xDD, 0xDD, 0xDD, 0xDD, 0xDD, 0xDD, 0xDD
+db 0xEE, 0xEE, 0xEE, 0xEE, 0xEE, 0xEE, 0xEE, 0xEE, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF
+
+.data_random:
+dd 83.0999,69.50512,41.02678,13.05881,5.35242,21.9932,9.67383,5.32372,29.02872,66.50151,19.30764,91.3633,40.45086,50.96153,32.64489,23.97574,90.64316,24.22547,98.9394,91.21715,90.80143,99.48407,64.97245,74.39838,35.22761,25.35321,5.8732,90.19956,33.03133,52.02952,58.38554,10.17531,47.84703,84.04831,90.02965,65.81329,96.27991,6.64479,25.58971,95.00694,88.1929,37.16964,49.52602,10.27223,77.70605,20.21439,9.8056,41.29389,15.4071,57.54286,9.61117,55.54302,52.90745,4.88086,72.52882,3.0201,56.55091,71.22749,61.84736,88.74295,47.72641,24.17404,33.70564,96.71303
+
diff --git a/unittests/InstructionCountCI/VEX_map1.json b/unittests/InstructionCountCI/VEX_map1.json
@@ -216,8 +216,8 @@
       ],
       "ExpectedArm64ASM": [
         "ldr q2, [x4]",
-        "mov v16.16b, v17.16b",
-        "mov v16.d[0], v2.d[0]"
+        "mov v16.16b, v2.16b",
+        "mov v16.d[1], v17.d[1]"
       ]
     },
     "vmovlpd xmm0, xmm1, [rax]": {
@@ -228,8 +228,8 @@
       ],
       "ExpectedArm64ASM": [
         "ldr q2, [x4]",
-        "mov v16.16b, v17.16b",
-        "mov v16.d[0], v2.d[0]"
+        "mov v16.16b, v2.16b",
+        "mov v16.d[1], v17.d[1]"
       ]
     },
     "vmovsldup xmm0, [rax]": {