[wasm-simd][ia32] Consolidate SSE/AVX opcodes for SIMD unops

Bug: v8:11217
Change-Id: Ic58b0ac90fa227cadc35829bd1e5629f6749020a
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2616102
Commit-Queue: Deepti Gandluri <gdeepti@chromium.org>
Reviewed-by: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#72083}

src/compiler/backend/ia32/code-generator-ia32.cc

@@ -2435,36 +2435,32 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       __ Addps(dst, dst, kScratchDoubleReg);  // add hi and lo, may round.
       break;
     }
-    case kSSEF32x4Abs: {
+    case kIA32F32x4Abs: {
       XMMRegister dst = i.OutputSimd128Register();
-      DCHECK_EQ(i.InputSimd128Register(0), dst);
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ psrld(kScratchDoubleReg, 1);
-      __ andps(dst, kScratchDoubleReg);
-      break;
+      XMMRegister src = i.InputSimd128Register(0);
+      if (dst == src) {
+        __ Pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
+        __ Psrld(kScratchDoubleReg, kScratchDoubleReg, 1);
+        __ Andps(dst, kScratchDoubleReg);
+      } else {
+        __ Pcmpeqd(dst, dst);
+        __ Psrld(dst, dst, 1);
+        __ Andps(dst, src);
       }
-    case kAVXF32x4Abs: {
-      CpuFeatureScope avx_scope(tasm(), AVX);
-      __ vpcmpeqd(kScratchDoubleReg, kScratchDoubleReg, kScratchDoubleReg);
-      __ vpsrld(kScratchDoubleReg, kScratchDoubleReg, 1);
-      __ vandps(i.OutputSimd128Register(), kScratchDoubleReg,
-                i.InputOperand(0));
       break;
     }
-    case kSSEF32x4Neg: {
+    case kIA32F32x4Neg: {
       XMMRegister dst = i.OutputSimd128Register();
-      DCHECK_EQ(dst, i.InputSimd128Register(0));
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ pslld(kScratchDoubleReg, 31);
-      __ xorps(dst, kScratchDoubleReg);
-      break;
+      XMMRegister src = i.InputSimd128Register(0);
+      if (dst == src) {
+        __ Pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
+        __ Pslld(kScratchDoubleReg, kScratchDoubleReg, 31);
+        __ Xorps(dst, kScratchDoubleReg);
+      } else {
+        __ Pcmpeqd(dst, dst);
+        __ Pslld(dst, dst, 31);
+        __ Xorps(dst, src);
       }
-    case kAVXF32x4Neg: {
-      CpuFeatureScope avx_scope(tasm(), AVX);
-      __ vpcmpeqd(kScratchDoubleReg, kScratchDoubleReg, kScratchDoubleReg);
-      __ vpslld(kScratchDoubleReg, kScratchDoubleReg, 31);
-      __ vxorps(i.OutputSimd128Register(), kScratchDoubleReg,
-                i.InputOperand(0));
       break;
     }
     case kIA32F32x4Sqrt: {
@@ -3824,17 +3820,16 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       __ Pcmpeqd(dst, dst);
       break;
     }
-    case kSSES128Not: {
+    case kIA32S128Not: {
       XMMRegister dst = i.OutputSimd128Register();
-      DCHECK_EQ(dst, i.InputSimd128Register(0));
-      __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ pxor(dst, kScratchDoubleReg);
-      break;
+      XMMRegister src = i.InputSimd128Register(0);
+      if (dst == src) {
+        __ Pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
+        __ Pxor(dst, kScratchDoubleReg);
+      } else {
+        __ Pcmpeqd(dst, dst);
+        __ Pxor(dst, src);
       }
-    case kAVXS128Not: {
-      CpuFeatureScope avx_scope(tasm(), AVX);
-      __ vpcmpeqd(kScratchDoubleReg, kScratchDoubleReg, kScratchDoubleReg);
-      __ vpxor(i.OutputSimd128Register(), kScratchDoubleReg, i.InputOperand(0));
       break;
     }
     case kSSES128And: {

src/compiler/backend/ia32/instruction-codes-ia32.h

@@ -163,10 +163,8 @@ namespace compiler {
   V(IA32Insertps)                  \
   V(IA32F32x4SConvertI32x4)        \
   V(IA32F32x4UConvertI32x4)        \
-  V(SSEF32x4Abs)                   \
-  V(AVXF32x4Abs)                   \
-  V(SSEF32x4Neg)                   \
-  V(AVXF32x4Neg)                   \
+  V(IA32F32x4Abs)                  \
+  V(IA32F32x4Neg)                  \
   V(IA32F32x4Sqrt)                 \
   V(IA32F32x4RecipApprox)          \
   V(IA32F32x4RecipSqrtApprox)      \
@@ -359,8 +357,7 @@ namespace compiler {
   V(IA32S128Const)                 \
   V(IA32S128Zero)                  \
   V(IA32S128AllOnes)               \
-  V(SSES128Not)                    \
-  V(AVXS128Not)                    \
+  V(IA32S128Not)                   \
   V(SSES128And)                    \
   V(AVXS128And)                    \
   V(SSES128Or)                     \

src/compiler/backend/ia32/instruction-scheduler-ia32.cc

@@ -145,10 +145,8 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kIA32Insertps:
     case kIA32F32x4SConvertI32x4:
     case kIA32F32x4UConvertI32x4:
-    case kSSEF32x4Abs:
-    case kAVXF32x4Abs:
-    case kSSEF32x4Neg:
-    case kAVXF32x4Neg:
+    case kIA32F32x4Abs:
+    case kIA32F32x4Neg:
     case kIA32F32x4Sqrt:
     case kIA32F32x4RecipApprox:
     case kIA32F32x4RecipSqrtApprox:
@@ -341,8 +339,7 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kIA32S128Const:
     case kIA32S128Zero:
     case kIA32S128AllOnes:
-    case kSSES128Not:
-    case kAVXS128Not:
+    case kIA32S128Not:
     case kSSES128And:
     case kAVXS128And:
     case kSSES128Or:

src/compiler/backend/ia32/instruction-selector-ia32.cc

@@ -2279,6 +2279,8 @@ void InstructionSelector::VisitWord32AtomicPairCompareExchange(Node* node) {
   V(I16x8ExtMulHighI8x16U)
 
 #define SIMD_UNOP_LIST(V)   \
+  V(F32x4Abs)               \
+  V(F32x4Neg)               \
   V(F32x4Sqrt)              \
   V(F32x4SConvertI32x4)     \
   V(F32x4RecipApprox)       \
@@ -2303,11 +2305,7 @@ void InstructionSelector::VisitWord32AtomicPairCompareExchange(Node* node) {
   V(I16x8Abs)               \
   V(I8x16Neg)               \
   V(I8x16Abs)               \
-  V(I8x16BitMask)
-
-#define SIMD_UNOP_PREFIX_LIST(V) \
-  V(F32x4Abs)                    \
-  V(F32x4Neg)                    \
+  V(I8x16BitMask)           \
   V(S128Not)
 
 #define SIMD_ANYTRUE_LIST(V) \
@@ -2612,25 +2610,6 @@ SIMD_UNOP_LIST(VISIT_SIMD_UNOP)
 #undef VISIT_SIMD_UNOP
 #undef SIMD_UNOP_LIST
 
-// TODO(v8:9198): SSE instructions that read 16 bytes from memory require the
-// operand to be 16-byte aligned. AVX instructions relax this requirement, but
-// might have reduced performance if the memory crosses cache line. But since we
-// have limited xmm registers, this might be okay to alleviate register
-// pressure.
-#define VISIT_SIMD_UNOP_PREFIX(Opcode)                                       \
-  void InstructionSelector::Visit##Opcode(Node* node) {                      \
-    IA32OperandGenerator g(this);                                            \
-    if (IsSupported(AVX)) {                                                  \
-      Emit(kAVX##Opcode, g.DefineAsRegister(node), g.Use(node->InputAt(0))); \
-    } else {                                                                 \
-      Emit(kSSE##Opcode, g.DefineSameAsFirst(node),                          \
-           g.UseRegister(node->InputAt(0)));                                 \
-    }                                                                        \
-  }
-SIMD_UNOP_PREFIX_LIST(VISIT_SIMD_UNOP_PREFIX)
-#undef VISIT_SIMD_UNOP_PREFIX
-#undef SIMD_UNOP_PREFIX_LIST
-
 // The implementation of AnyTrue is the same for all shapes.
 #define VISIT_SIMD_ANYTRUE(Opcode)                                  \
   void InstructionSelector::Visit##Opcode(Node* node) {             \