[wasm-simd] Unify sse and avx impl for i16x8 shifts

The implementation is pretty much the same, and we instead delegate to a
macro assembler to decide if we want the sse or avx instruction.

This unification will simplify optimization of constant shifts later on.

Bug: v8:10115
Change-Id: I68e60cb3fd51156438989812be189f71e6e47ba7
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2026470Reviewed-by: Deepti Gandluri <gdeepti@chromium.org>
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#66121}

src/codegen/ia32/macro-assembler-ia32.h

@@ -335,10 +335,13 @@ class V8_EXPORT_PRIVATE TurboAssembler : public TurboAssemblerBase {
   AVX_PACKED_OP3(Minpd, minpd)
   AVX_PACKED_OP3(Maxpd, maxpd)
   AVX_PACKED_OP3(Cmpunordpd, cmpunordpd)
+  AVX_PACKED_OP3(Psllw, psllw)
   AVX_PACKED_OP3(Pslld, pslld)
   AVX_PACKED_OP3(Psllq, psllq)
+  AVX_PACKED_OP3(Psrlw, psrlw)
   AVX_PACKED_OP3(Psrld, psrld)
   AVX_PACKED_OP3(Psrlq, psrlq)
+  AVX_PACKED_OP3(Psraw, psraw)
   AVX_PACKED_OP3(Psrad, psrad)
   AVX_PACKED_OP3(Paddq, paddq)
   AVX_PACKED_OP3(Psubq, psubq)

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

@@ -2818,44 +2818,22 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       }
       break;
     }
-    case kSSEI16x8Shl: {
-      DCHECK_EQ(i.OutputSimd128Register(), i.InputSimd128Register(0));
-      XMMRegister tmp = i.TempSimd128Register(0);
-      Register shift = i.InputRegister(1);
-      // Take shift value modulo 16.
-      __ and_(shift, 15);
-      __ movd(tmp, shift);
-      __ psllw(i.OutputSimd128Register(), tmp);
-      break;
-    }
-    case kAVXI16x8Shl: {
-      CpuFeatureScope avx_scope(tasm(), AVX);
-      XMMRegister tmp = i.TempSimd128Register(0);
-      Register shift = i.InputRegister(1);
-      // Take shift value modulo 16.
-      __ and_(shift, 15);
-      __ movd(tmp, shift);
-      __ vpsllw(i.OutputSimd128Register(), i.InputSimd128Register(0), tmp);
-      break;
-    }
-    case kSSEI16x8ShrS: {
-      DCHECK_EQ(i.OutputSimd128Register(), i.InputSimd128Register(0));
+    case kIA32I16x8Shl: {
       XMMRegister tmp = i.TempSimd128Register(0);
       Register shift = i.InputRegister(1);
       // Take shift value modulo 16.
       __ and_(shift, 15);
-      __ movd(tmp, shift);
-      __ psraw(i.OutputSimd128Register(), tmp);
+      __ Movd(tmp, shift);
+      __ Psllw(i.OutputSimd128Register(), i.InputSimd128Register(0), tmp);
       break;
     }
-    case kAVXI16x8ShrS: {
-      CpuFeatureScope avx_scope(tasm(), AVX);
+    case kIA32I16x8ShrS: {
       XMMRegister tmp = i.TempSimd128Register(0);
       Register shift = i.InputRegister(1);
       // Take shift value modulo 16.
       __ and_(shift, 15);
-      __ movd(tmp, shift);
-      __ vpsraw(i.OutputSimd128Register(), i.InputSimd128Register(0), tmp);
+      __ Movd(tmp, shift);
+      __ Psraw(i.OutputSimd128Register(), i.InputSimd128Register(0), tmp);
       break;
     }
     case kSSEI16x8SConvertI32x4: {
@@ -3022,24 +3000,13 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       __ Pmovzxbw(dst, dst);
       break;
     }
-    case kSSEI16x8ShrU: {
-      DCHECK_EQ(i.OutputSimd128Register(), i.InputSimd128Register(0));
-      XMMRegister tmp = i.TempSimd128Register(0);
-      Register shift = i.InputRegister(1);
-      // Take shift value modulo 16.
-      __ and_(shift, 15);
-      __ movd(tmp, shift);
-      __ psrlw(i.OutputSimd128Register(), tmp);
-      break;
-    }
-    case kAVXI16x8ShrU: {
-      CpuFeatureScope avx_scope(tasm(), AVX);
+    case kIA32I16x8ShrU: {
       XMMRegister tmp = i.TempSimd128Register(0);
       Register shift = i.InputRegister(1);
       // Take shift value modulo 16.
       __ and_(shift, 15);
-      __ movd(tmp, shift);
-      __ vpsrlw(i.OutputSimd128Register(), i.InputSimd128Register(0), tmp);
+      __ Movd(tmp, shift);
+      __ Psrlw(i.OutputSimd128Register(), i.InputSimd128Register(0), tmp);
       break;
     }
     case kSSEI16x8UConvertI32x4: {

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

@@ -233,10 +233,8 @@ namespace compiler {
   V(IA32I16x8SConvertI8x16Low)     \
   V(IA32I16x8SConvertI8x16High)    \
   V(IA32I16x8Neg)                  \
-  V(SSEI16x8Shl)                   \
-  V(AVXI16x8Shl)                   \
-  V(SSEI16x8ShrS)                  \
-  V(AVXI16x8ShrS)                  \
+  V(IA32I16x8Shl)                  \
+  V(IA32I16x8ShrS)                 \
   V(SSEI16x8SConvertI32x4)         \
   V(AVXI16x8SConvertI32x4)         \
   V(SSEI16x8Add)                   \
@@ -265,8 +263,7 @@ namespace compiler {
   V(AVXI16x8GeS)                   \
   V(IA32I16x8UConvertI8x16Low)     \
   V(IA32I16x8UConvertI8x16High)    \
-  V(SSEI16x8ShrU)                  \
-  V(AVXI16x8ShrU)                  \
+  V(IA32I16x8ShrU)                 \
   V(SSEI16x8UConvertI32x4)         \
   V(AVXI16x8UConvertI32x4)         \
   V(SSEI16x8AddSaturateU)          \

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

@@ -214,10 +214,8 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kIA32I16x8SConvertI8x16Low:
     case kIA32I16x8SConvertI8x16High:
     case kIA32I16x8Neg:
-    case kSSEI16x8Shl:
-    case kAVXI16x8Shl:
-    case kSSEI16x8ShrS:
-    case kAVXI16x8ShrS:
+    case kIA32I16x8Shl:
+    case kIA32I16x8ShrS:
     case kSSEI16x8SConvertI32x4:
     case kAVXI16x8SConvertI32x4:
     case kSSEI16x8Add:
@@ -246,8 +244,7 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kAVXI16x8GeS:
     case kIA32I16x8UConvertI8x16Low:
     case kIA32I16x8UConvertI8x16High:
-    case kSSEI16x8ShrU:
-    case kAVXI16x8ShrU:
+    case kIA32I16x8ShrU:
     case kSSEI16x8UConvertI32x4:
     case kAVXI16x8UConvertI32x4:
     case kSSEI16x8AddSaturateU:

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

@@ -305,18 +305,13 @@ void VisitRRISimd(InstructionSelector* selector, Node* node,
 }
 
 void VisitRROSimdShift(InstructionSelector* selector, Node* node,
-                       ArchOpcode avx_opcode, ArchOpcode sse_opcode) {
+                       ArchOpcode opcode) {
   IA32OperandGenerator g(selector);
   InstructionOperand operand0 = g.UseUniqueRegister(node->InputAt(0));
   InstructionOperand operand1 = g.UseUniqueRegister(node->InputAt(1));
   InstructionOperand temps[] = {g.TempSimd128Register()};
-  if (selector->IsSupported(AVX)) {
-    selector->Emit(avx_opcode, g.DefineAsRegister(node), operand0, operand1,
-                   arraysize(temps), temps);
-  } else {
-    selector->Emit(sse_opcode, g.DefineSameAsFirst(node), operand0, operand1,
-                   arraysize(temps), temps);
-  }
+  selector->Emit(opcode, g.DefineSameAsFirst(node), operand0, operand1,
+                 arraysize(temps), temps);
 }
 
 void VisitRROI8x16SimdRightShift(InstructionSelector* selector, Node* node,
@@ -2130,17 +2125,15 @@ void InstructionSelector::VisitWord32AtomicPairCompareExchange(Node* node) {
   V(S1x8AllTrue)             \
   V(S1x16AllTrue)
 
-#define SIMD_SHIFT_OPCODES(V) \
-  V(I16x8Shl)                 \
-  V(I16x8ShrS)                \
-  V(I16x8ShrU)
-
 #define SIMD_SHIFT_OPCODES_UNIFED_SSE_AVX(V) \
   V(I64x2Shl)                                \
   V(I64x2ShrU)                               \
   V(I32x4Shl)                                \
   V(I32x4ShrS)                               \
-  V(I32x4ShrU)
+  V(I32x4ShrU)                               \
+  V(I16x8Shl)                                \
+  V(I16x8ShrS)                               \
+  V(I16x8ShrU)
 
 #define SIMD_I8X16_RIGHT_SHIFT_OPCODES(V) \
   V(I8x16ShrS)                            \
@@ -2360,17 +2353,9 @@ VISIT_SIMD_REPLACE_LANE(F32x4)
 VISIT_SIMD_REPLACE_LANE_USE_REG(F64x2)
 #undef VISIT_SIMD_REPLACE_LANE_USE_REG
 
-#define VISIT_SIMD_SHIFT(Opcode)                               \
-  void InstructionSelector::Visit##Opcode(Node* node) {        \
-    VisitRROSimdShift(this, node, kAVX##Opcode, kSSE##Opcode); \
-  }
-SIMD_SHIFT_OPCODES(VISIT_SIMD_SHIFT)
-#undef VISIT_SIMD_SHIFT
-#undef SIMD_SHIFT_OPCODES
-
-#define VISIT_SIMD_SHIFT_UNIFIED_SSE_AVX(Opcode)                 \
-  void InstructionSelector::Visit##Opcode(Node* node) {          \
-    VisitRROSimdShift(this, node, kIA32##Opcode, kIA32##Opcode); \
+#define VISIT_SIMD_SHIFT_UNIFIED_SSE_AVX(Opcode)        \
+  void InstructionSelector::Visit##Opcode(Node* node) { \
+    VisitRROSimdShift(this, node, kIA32##Opcode);       \
   }
 SIMD_SHIFT_OPCODES_UNIFED_SSE_AVX(VISIT_SIMD_SHIFT_UNIFIED_SSE_AVX)
 #undef VISIT_SIMD_SHIFT_UNIFIED_SSE_AVX