Reland "[wasm-simd][ia32] Prototype sign select"

This is a reland of 716dae3a

Original change's description:
> [wasm-simd][ia32] Prototype sign select
>
> The implementation is the same as on x64.
>
> Bug: v8:10983
> Change-Id: I2654ce4a627ca5cc6c759051ab9034c528d9f25a
> Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2567194
> Reviewed-by: Bill Budge <bbudge@chromium.org>
> Commit-Queue: Zhi An Ng <zhin@chromium.org>
> Cr-Commit-Position: refs/heads/master@{#71606}

Bug: v8:10983
Change-Id: I05af92ec2d3531dd2e0d27353cc665967fb5c387
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2574001
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Bill Budge <bbudge@chromium.org>
Cr-Commit-Position: refs/heads/master@{#71627}

src/codegen/ia32/assembler-ia32.cc

@@ -2973,6 +2973,24 @@ void Assembler::vpshufd(XMMRegister dst, Operand src, uint8_t shuffle) {
   EMIT(shuffle);
 }
 
+void Assembler::vblendvps(XMMRegister dst, XMMRegister src1, XMMRegister src2,
+                          XMMRegister mask) {
+  vinstr(0x4A, dst, src1, src2, k66, k0F3A, kW0);
+  EMIT(mask.code() << 4);
+}
+
+void Assembler::vblendvpd(XMMRegister dst, XMMRegister src1, XMMRegister src2,
+                          XMMRegister mask) {
+  vinstr(0x4B, dst, src1, src2, k66, k0F3A, kW0);
+  EMIT(mask.code() << 4);
+}
+
+void Assembler::vpblendvb(XMMRegister dst, XMMRegister src1, XMMRegister src2,
+                          XMMRegister mask) {
+  vinstr(0x4C, dst, src1, src2, k66, k0F3A, kW0);
+  EMIT(mask.code() << 4);
+}
+
 void Assembler::vpblendw(XMMRegister dst, XMMRegister src1, Operand src2,
                          uint8_t mask) {
   vinstr(0x0E, dst, src1, src2, k66, k0F3A, kWIG);

src/codegen/ia32/assembler-ia32.h

@@ -1406,6 +1406,13 @@ class V8_EXPORT_PRIVATE Assembler : public AssemblerBase {
   }
   void vpshufd(XMMRegister dst, Operand src, uint8_t shuffle);
 
+  void vblendvps(XMMRegister dst, XMMRegister src1, XMMRegister src2,
+                 XMMRegister mask);
+  void vblendvpd(XMMRegister dst, XMMRegister src1, XMMRegister src2,
+                 XMMRegister mask);
+  void vpblendvb(XMMRegister dst, XMMRegister src1, XMMRegister src2,
+                 XMMRegister mask);
+
   void vpblendw(XMMRegister dst, XMMRegister src1, XMMRegister src2,
                 uint8_t mask) {
     vpblendw(dst, src1, Operand(src2), mask);
@@ -1692,6 +1699,9 @@ class V8_EXPORT_PRIVATE Assembler : public AssemblerBase {
 
   SSE4_INSTRUCTION_LIST(DECLARE_SSE4_INSTRUCTION)
   SSE4_RM_INSTRUCTION_LIST(DECLARE_SSE4_INSTRUCTION)
+  DECLARE_SSE4_INSTRUCTION(blendvps, 66, 0F, 38, 14)
+  DECLARE_SSE4_INSTRUCTION(blendvpd, 66, 0F, 38, 15)
+  DECLARE_SSE4_INSTRUCTION(pblendvb, 66, 0F, 38, 10)
 #undef DECLARE_SSE4_INSTRUCTION
 
 #define DECLARE_SSE34_AVX_INSTRUCTION(instruction, prefix, escape1, escape2,  \

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

@@ -505,6 +505,7 @@ class V8_EXPORT_PRIVATE TurboAssembler : public TurboAssemblerBase {
     Pshufb(dst, src, Operand(mask));
   }
   void Pshufb(XMMRegister dst, XMMRegister src, Operand mask);
+
   void Pblendw(XMMRegister dst, XMMRegister src, uint8_t imm8) {
     Pblendw(dst, Operand(src), imm8);
   }

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

@@ -535,6 +535,32 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
     }                                                        \
   } while (false)
 
+// Helper macro to help define signselect opcodes. This should not be used for
+// i16x8.signselect, because there is no native word-sized blend instruction.
+// We choose a helper macro here instead of a macro-assembler function because
+// the blend instruction requires xmm0 as an implicit argument, and the codegen
+// relies on xmm0 being the scratch register, so we can freely overwrite it as
+// required.
+#define ASSEMBLE_SIMD_SIGN_SELECT(BLEND_OP)       \
+  do {                                            \
+    XMMRegister dst = i.OutputSimd128Register();  \
+    XMMRegister src1 = i.InputSimd128Register(0); \
+    XMMRegister src2 = i.InputSimd128Register(1); \
+    XMMRegister mask = i.InputSimd128Register(2); \
+    if (CpuFeatures::IsSupported(AVX)) {          \
+      CpuFeatureScope avx_scope(tasm(), AVX);     \
+      __ v##BLEND_OP(dst, src1, src2, mask);      \
+    } else {                                      \
+      CpuFeatureScope scope(tasm(), SSE4_1);      \
+      DCHECK_EQ(dst, src1);                       \
+      DCHECK_EQ(kScratchDoubleReg, xmm0);         \
+      if (mask != xmm0) {                         \
+        __ movaps(xmm0, mask);                    \
+      }                                           \
+      __ BLEND_OP(dst, src2);                     \
+    }                                             \
+  } while (false)
+
 void CodeGenerator::AssembleDeconstructFrame() {
   __ mov(esp, ebp);
   __ pop(ebp);
@@ -2158,6 +2184,37 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
                  i.InputSimd128Register(1));
       break;
     }
+    case kIA32I8x16SignSelect: {
+      ASSEMBLE_SIMD_SIGN_SELECT(pblendvb);
+      break;
+    }
+    case kIA32I16x8SignSelect: {
+      XMMRegister dst = i.OutputSimd128Register();
+      XMMRegister src1 = i.InputSimd128Register(0);
+      XMMRegister src2 = i.InputSimd128Register(1);
+      XMMRegister mask = i.InputSimd128Register(2);
+      if (CpuFeatures::IsSupported(AVX)) {
+        CpuFeatureScope avx_scope(tasm(), AVX);
+        __ vpsraw(kScratchDoubleReg, mask, 15);
+        __ vpblendvb(dst, src1, src2, kScratchDoubleReg);
+      } else {
+        CpuFeatureScope sse_scope(tasm(), SSE4_1);
+        DCHECK_EQ(dst, src1);
+        DCHECK_EQ(kScratchDoubleReg, xmm0);
+        __ pxor(kScratchDoubleReg, kScratchDoubleReg);
+        __ pcmpgtw(kScratchDoubleReg, mask);
+        __ pblendvb(dst, src2);
+      }
+      break;
+    }
+    case kIA32I32x4SignSelect: {
+      ASSEMBLE_SIMD_SIGN_SELECT(blendvps);
+      break;
+    }
+    case kIA32I64x2SignSelect: {
+      ASSEMBLE_SIMD_SIGN_SELECT(blendvpd);
+      break;
+    }
     case kSSEF32x4Splat: {
       DCHECK_EQ(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       XMMRegister dst = i.OutputSimd128Register();

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

@@ -150,6 +150,7 @@ namespace compiler {
   V(IA32I64x2ShrU)                 \
   V(IA32I64x2BitMask)              \
   V(IA32I64x2Eq)                   \
+  V(IA32I64x2SignSelect)           \
   V(SSEF32x4Splat)                 \
   V(AVXF32x4Splat)                 \
   V(SSEF32x4ExtractLane)           \
@@ -236,6 +237,7 @@ namespace compiler {
   V(IA32I32x4Abs)                  \
   V(IA32I32x4BitMask)              \
   V(IA32I32x4DotI16x8S)            \
+  V(IA32I32x4SignSelect)           \
   V(IA32I16x8Splat)                \
   V(IA32I16x8ExtractLaneS)         \
   V(IA32I16x8SConvertI8x16Low)     \
@@ -289,6 +291,7 @@ namespace compiler {
   V(IA32I16x8RoundingAverageU)     \
   V(IA32I16x8Abs)                  \
   V(IA32I16x8BitMask)              \
+  V(IA32I16x8SignSelect)           \
   V(IA32I8x16Splat)                \
   V(IA32I8x16ExtractLaneS)         \
   V(IA32Pinsrb)                    \
@@ -342,6 +345,7 @@ namespace compiler {
   V(IA32I8x16RoundingAverageU)     \
   V(IA32I8x16Abs)                  \
   V(IA32I8x16BitMask)              \
+  V(IA32I8x16SignSelect)           \
   V(IA32S128Const)                 \
   V(IA32S128Zero)                  \
   V(IA32S128AllOnes)               \

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

@@ -129,6 +129,7 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kIA32I64x2ShrU:
     case kIA32I64x2BitMask:
     case kIA32I64x2Eq:
+    case kIA32I64x2SignSelect:
     case kSSEF32x4Splat:
     case kAVXF32x4Splat:
     case kSSEF32x4ExtractLane:
@@ -215,6 +216,7 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kIA32I32x4Abs:
     case kIA32I32x4BitMask:
     case kIA32I32x4DotI16x8S:
+    case kIA32I32x4SignSelect:
     case kIA32I16x8Splat:
     case kIA32I16x8ExtractLaneS:
     case kIA32I16x8SConvertI8x16Low:
@@ -268,6 +270,7 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kIA32I16x8RoundingAverageU:
     case kIA32I16x8Abs:
     case kIA32I16x8BitMask:
+    case kIA32I16x8SignSelect:
     case kIA32I8x16Splat:
     case kIA32I8x16ExtractLaneS:
     case kIA32Pinsrb:
@@ -321,6 +324,7 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kIA32I8x16RoundingAverageU:
     case kIA32I8x16Abs:
     case kIA32I8x16BitMask:
+    case kIA32I8x16SignSelect:
     case kIA32S128Const:
     case kIA32S128Zero:
     case kIA32S128AllOnes:

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

@@ -2993,6 +2993,44 @@ void InstructionSelector::VisitF64x2Pmax(Node* node) {
   VisitPminOrPmax(this, node, kIA32F64x2Pmax);
 }
 
+namespace {
+void VisitSignSelect(InstructionSelector* selector, Node* node,
+                     ArchOpcode opcode) {
+  IA32OperandGenerator g(selector);
+  // signselect(x, y, -1) = x
+  // pblendvb(dst, x, y, -1) = dst <- y, so we need to swap x and y.
+  if (selector->IsSupported(AVX)) {
+    selector->Emit(
+        opcode, g.DefineAsRegister(node), g.UseRegister(node->InputAt(1)),
+        g.UseRegister(node->InputAt(0)), g.UseRegister(node->InputAt(2)));
+  } else {
+    // We would like to fix the mask to be xmm0, since that is what
+    // pblendvb/blendvps/blendvps uses as an implicit operand. However, xmm0 is
+    // also scratch register, so our mask values can be overwritten. Instead, we
+    // manually move the mask to xmm0 inside codegen.
+    selector->Emit(
+        opcode, g.DefineSameAsFirst(node), g.UseRegister(node->InputAt(1)),
+        g.UseRegister(node->InputAt(0)), g.UseRegister(node->InputAt(2)));
+  }
+}
+}  // namespace
+
+void InstructionSelector::VisitI8x16SignSelect(Node* node) {
+  VisitSignSelect(this, node, kIA32I8x16SignSelect);
+}
+
+void InstructionSelector::VisitI16x8SignSelect(Node* node) {
+  VisitSignSelect(this, node, kIA32I16x8SignSelect);
+}
+
+void InstructionSelector::VisitI32x4SignSelect(Node* node) {
+  VisitSignSelect(this, node, kIA32I32x4SignSelect);
+}
+
+void InstructionSelector::VisitI64x2SignSelect(Node* node) {
+  VisitSignSelect(this, node, kIA32I64x2SignSelect);
+}
+
 // static
 MachineOperatorBuilder::Flags
 InstructionSelector::SupportedMachineOperatorFlags() {

src/compiler/backend/instruction-selector.cc

@@ -2826,13 +2826,13 @@ void InstructionSelector::VisitLoadLane(Node* node) { UNIMPLEMENTED(); }
 void InstructionSelector::VisitStoreLane(Node* node) { UNIMPLEMENTED(); }
 #endif  // !V8_TARGET_ARCH_X64 && !V8_TARGET_ARCH_IA32 && !V8_TARGET_ARCH_ARM64
 
-#if !V8_TARGET_ARCH_X64
+#if !V8_TARGET_ARCH_X64 && !V8_TARGET_ARCH_IA32
 // TODO(v8:10983) Prototyping sign select.
 void InstructionSelector::VisitI8x16SignSelect(Node* node) { UNIMPLEMENTED(); }
 void InstructionSelector::VisitI16x8SignSelect(Node* node) { UNIMPLEMENTED(); }
 void InstructionSelector::VisitI32x4SignSelect(Node* node) { UNIMPLEMENTED(); }
 void InstructionSelector::VisitI64x2SignSelect(Node* node) { UNIMPLEMENTED(); }
-#endif  // !V8_TARGET_ARCH_X64
+#endif  // !V8_TARGET_ARCH_X64 && !V8_TARGET_ARCH_IA32
 
 #if !V8_TARGET_ARCH_X64 && !V8_TARGET_ARCH_ARM64 && !V8_TARGET_ARCH_ARM && \
     !V8_TARGET_ARCH_IA32 && !V8_TARGET_ARCH_MIPS64 && !V8_TARGET_ARCH_MIPS

src/diagnostics/ia32/disasm-ia32.cc

@@ -856,6 +856,24 @@ int DisassemblerIA32::AVXInstruction(byte* data) {
         AppendToBuffer(",%d", Imm8(current));
         current++;
         break;
+      case 0x4A:
+        AppendToBuffer("vblendvps %s,%s,", NameOfXMMRegister(regop),
+                       NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        AppendToBuffer(",%s", NameOfXMMRegister(*current >> 4));
+        break;
+      case 0x4B:
+        AppendToBuffer("vblendvps %s,%s,", NameOfXMMRegister(regop),
+                       NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        AppendToBuffer(",%s", NameOfXMMRegister(*current >> 4));
+        break;
+      case 0x4C:
+        AppendToBuffer("vpblendvb %s,%s,", NameOfXMMRegister(regop),
+                       NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        AppendToBuffer(",%s", NameOfXMMRegister(*current >> 4));
+        break;
       default:
         UnimplementedInstruction();
     }
@@ -2189,6 +2207,21 @@ int DisassemblerIA32::InstructionDecode(v8::internal::Vector<char> out_buffer,
               SSE4_INSTRUCTION_LIST(SSE34_DIS_CASE)
               SSE4_RM_INSTRUCTION_LIST(SSE34_DIS_CASE)
 #undef SSE34_DIS_CASE
+              case 0x10:
+                AppendToBuffer("pblendvb %s,", NameOfXMMRegister(regop));
+                data += PrintRightXMMOperand(data);
+                AppendToBuffer(",xmm0");
+                break;
+              case 0x14:
+                AppendToBuffer("blendvps %s,", NameOfXMMRegister(regop));
+                data += PrintRightXMMOperand(data);
+                AppendToBuffer(",xmm0");
+                break;
+              case 0x15:
+                AppendToBuffer("blendvps %s,", NameOfXMMRegister(regop));
+                data += PrintRightXMMOperand(data);
+                AppendToBuffer(",xmm0");
+                break;
               default:
                 UnimplementedInstruction();
             }

test/cctest/test-disasm-ia32.cc

@@ -627,6 +627,10 @@ TEST(DisasmIa320) {
       __ pinsrd(xmm1, Operand(edx, 4), 0);
       __ extractps(eax, xmm1, 0);
 
+      __ blendvps(xmm3, xmm1);
+      __ blendvpd(xmm3, xmm1);
+      __ pblendvb(xmm3, xmm1);
+
       SSE4_INSTRUCTION_LIST(EMIT_SSE34_INSTR)
       SSE4_RM_INSTRUCTION_LIST(EMIT_SSE34_INSTR)
     }
@@ -786,6 +790,10 @@ TEST(DisasmIa320) {
       __ vpinsrd(xmm0, xmm1, eax, 0);
       __ vpinsrd(xmm0, xmm1, Operand(edx, 4), 0);
 
+      __ vblendvps(xmm3, xmm1, xmm4, xmm6);
+      __ vblendvpd(xmm3, xmm1, xmm4, xmm6);
+      __ vpblendvb(xmm3, xmm1, xmm4, xmm6);
+
       __ vcvtdq2ps(xmm1, xmm0);
       __ vcvtdq2ps(xmm1, Operand(ebx, ecx, times_4, 10000));
       __ vcvttps2dq(xmm1, xmm0);

test/cctest/wasm/test-run-wasm-simd.cc

@@ -765,7 +765,7 @@ WASM_SIMD_TEST(F32x4Le) {
   RunF32x4CompareOpTest(execution_tier, lower_simd, kExprF32x4Le, LessEqual);
 }
 
-#if V8_TARGET_ARCH_X64
+#if V8_TARGET_ARCH_X64 || V8_TARGET_ARCH_IA32
 // TODO(v8:10983) Prototyping sign select.
 template <typename T>
 void RunSignSelect(TestExecutionTier execution_tier, LowerSimd lower_simd,
@@ -822,7 +822,7 @@ WASM_SIMD_TEST_NO_LOWERING(I64x2SignSelect) {
   RunSignSelect<int64_t>(execution_tier, lower_simd, kExprI64x2SignSelect,
                          kExprI64x2Splat, mask);
 }
-#endif  // V8_TARGET_ARCH_X64
+#endif  // V8_TARGET_ARCH_X64 || V8_TARGET_ARCH_IA32
 
 #if V8_TARGET_ARCH_X64 || V8_TARGET_ARCH_ARM64 || V8_TARGET_ARCH_S390X
 WASM_SIMD_TEST_NO_LOWERING(F32x4Qfma) {