[wasm-simd] Share i8x16.splat implementation

The optimal implementation is in TurboFan x64 codegen, move it into
shared-macro-assembler, and have TurboFan ia32 and Liftoff use it. The
optimal implementation accounts for AVX2 support.

We add a couple of AVX2 instruction to ia32 in sse-instr.h, not all of
them are used, but follow-up patches will use them, so we add support
(including diassembly and test) in this change.

Drive-by clean up to test-disasm-x64.cc to merge 2 AVX2 test sections.

Bug: v8:11589
Change-Id: I1c8d7deb0f8bb70b29e7a680e5dbcfb09ca5505b
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3092555Reviewed-by: Clemens Backes <clemensb@chromium.org>
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/main@{#76352}

src/codegen/ia32/assembler-ia32.h

@@ -1790,6 +1790,19 @@ class V8_EXPORT_PRIVATE Assembler : public AssemblerBase {
   SSE4_RM_INSTRUCTION_LIST(DECLARE_SSE4_AVX_RM_INSTRUCTION)
 #undef DECLARE_SSE4_AVX_RM_INSTRUCTION
 
+  // AVX2 instructions
+#define AVX2_INSTRUCTION(instr, prefix, escape1, escape2, opcode)           \
+  void instr(XMMRegister dst, XMMRegister src) {                            \
+    vinstr(0x##opcode, dst, xmm0, src, k##prefix, k##escape1##escape2, kW0, \
+           AVX2);                                                           \
+  }                                                                         \
+  void instr(XMMRegister dst, Operand src) {                                \
+    vinstr(0x##opcode, dst, xmm0, src, k##prefix, k##escape1##escape2, kW0, \
+           AVX2);                                                           \
+  }
+  AVX2_BROADCAST_LIST(AVX2_INSTRUCTION)
+#undef AVX2_INSTRUCTION
+
   // Prefetch src position into cache level.
   // Level 1, 2 or 3 specifies CPU cache level. Level 0 specifies a
   // non-temporal

src/codegen/ia32/sse-instr.h

@@ -102,4 +102,10 @@
   V(pmovzxdq, 66, 0F, 38, 35)       \
   V(ptest, 66, 0F, 38, 17)
 
+// These require AVX2, and we only define the VEX-128 versions.
+#define AVX2_BROADCAST_LIST(V)    \
+  V(vpbroadcastd, 66, 0F, 38, 58) \
+  V(vpbroadcastb, 66, 0F, 38, 78) \
+  V(vpbroadcastw, 66, 0F, 38, 79)
+
 #endif  // V8_CODEGEN_IA32_SSE_INSTR_H_

src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.cc

@@ -254,6 +254,42 @@ void SharedTurboAssembler::S128Store32Lane(Operand dst, XMMRegister src,
   }
 }
 
+template <typename Op>
+void SharedTurboAssembler::I8x16SplatPreAvx2(XMMRegister dst, Op src,
+                                             XMMRegister scratch) {
+  DCHECK(!CpuFeatures::IsSupported(AVX2));
+  CpuFeatureScope ssse3_scope(this, SSSE3);
+  Movd(dst, src);
+  Xorps(scratch, scratch);
+  Pshufb(dst, scratch);
+}
+
+void SharedTurboAssembler::I8x16Splat(XMMRegister dst, Register src,
+                                      XMMRegister scratch) {
+  if (CpuFeatures::IsSupported(AVX2)) {
+    CpuFeatureScope avx2_scope(this, AVX2);
+    Movd(scratch, src);
+    vpbroadcastb(dst, scratch);
+  } else {
+    I8x16SplatPreAvx2(dst, src, scratch);
+  }
+}
+
+void SharedTurboAssembler::I8x16Splat(XMMRegister dst, Operand src,
+                                      XMMRegister scratch) {
+#if V8_TARGET_ARCH_IA32
+  // Operand on IA32 can be a wrapper for a single register, in which case they
+  // should call I8x16Splat |src| being Register.
+  DCHECK(!src.is_reg_only());
+#endif
+  if (CpuFeatures::IsSupported(AVX2)) {
+    CpuFeatureScope avx2_scope(this, AVX2);
+    vpbroadcastb(dst, src);
+  } else {
+    I8x16SplatPreAvx2(dst, src, scratch);
+  }
+}
+
 void SharedTurboAssembler::I8x16Shl(XMMRegister dst, XMMRegister src1,
                                     uint8_t src2, Register tmp1,
                                     XMMRegister tmp2) {

src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h

@@ -223,6 +223,7 @@ class V8_EXPORT_PRIVATE SharedTurboAssembler : public TurboAssemblerBase {
   AVX_OP(Pmullw, pmullw)
   AVX_OP(Pmuludq, pmuludq)
   AVX_OP(Por, por)
+  AVX_OP(Pshufb, pshufb)
   AVX_OP(Pshufd, pshufd)
   AVX_OP(Pshufhw, pshufhw)
   AVX_OP(Pshuflw, pshuflw)
@@ -300,6 +301,8 @@ class V8_EXPORT_PRIVATE SharedTurboAssembler : public TurboAssemblerBase {
   void F32x4Splat(XMMRegister dst, DoubleRegister src);
   void F32x4ExtractLane(FloatRegister dst, XMMRegister src, uint8_t lane);
   void S128Store32Lane(Operand dst, XMMRegister src, uint8_t laneidx);
+  void I8x16Splat(XMMRegister dst, Register src, XMMRegister scratch);
+  void I8x16Splat(XMMRegister dst, Operand src, XMMRegister scratch);
   void I8x16Shl(XMMRegister dst, XMMRegister src1, uint8_t src2, Register tmp1,
                 XMMRegister tmp2);
   void I8x16Shl(XMMRegister dst, XMMRegister src1, Register src2, Register tmp1,
@@ -350,6 +353,10 @@ class V8_EXPORT_PRIVATE SharedTurboAssembler : public TurboAssemblerBase {
   // Requires dst == mask when AVX is not supported.
   void S128Select(XMMRegister dst, XMMRegister mask, XMMRegister src1,
                   XMMRegister src2, XMMRegister scratch);
+
+ private:
+  template <typename Op>
+  void I8x16SplatPreAvx2(XMMRegister dst, Op src, XMMRegister scratch);
 };
 }  // namespace internal
 }  // namespace v8

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

@@ -3009,10 +3009,13 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       break;
     }
     case kIA32I8x16Splat: {
-      XMMRegister dst = i.OutputSimd128Register();
-      __ Movd(dst, i.InputOperand(0));
-      __ Pxor(kScratchDoubleReg, kScratchDoubleReg);
-      __ Pshufb(dst, kScratchDoubleReg);
+      if (instr->InputAt(0)->IsRegister()) {
+        __ I8x16Splat(i.OutputSimd128Register(), i.InputRegister(0),
+                      kScratchDoubleReg);
+      } else {
+        __ I8x16Splat(i.OutputSimd128Register(), i.InputOperand(0),
+                      kScratchDoubleReg);
+      }
       break;
     }
     case kIA32I8x16ExtractLaneS: {

src/compiler/backend/x64/code-generator-x64.cc

@@ -3405,25 +3405,11 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
     }
     case kX64I8x16Splat: {
       XMMRegister dst = i.OutputSimd128Register();
-      if (CpuFeatures::IsSupported(AVX2)) {
-        CpuFeatureScope avx_scope(tasm(), AVX);
-        CpuFeatureScope avx2_scope(tasm(), AVX2);
-        if (HasRegisterInput(instr, 0)) {
-          __ vmovd(kScratchDoubleReg, i.InputRegister(0));
-          __ vpbroadcastb(dst, kScratchDoubleReg);
-        } else {
-          __ vpbroadcastb(dst, i.InputOperand(0));
-        }
+      if (HasRegisterInput(instr, 0)) {
+        __ I8x16Splat(dst, i.InputRegister(0), kScratchDoubleReg);
       } else {
-        if (HasRegisterInput(instr, 0)) {
-          __ Movd(dst, i.InputRegister(0));
-        } else {
-          __ Movd(dst, i.InputOperand(0));
-        }
-        __ Xorps(kScratchDoubleReg, kScratchDoubleReg);
-        __ Pshufb(dst, kScratchDoubleReg);
+        __ I8x16Splat(dst, i.InputOperand(0), kScratchDoubleReg);
       }
-
       break;
     }
     case kX64Pextrb: {

src/diagnostics/ia32/disasm-ia32.cc

@@ -786,6 +786,15 @@ int DisassemblerIA32::AVXInstruction(byte* data) {
         SSSE3_UNOP_INSTRUCTION_LIST(DECLARE_SSE_AVX_RM_DIS_CASE)
         SSE4_RM_INSTRUCTION_LIST(DECLARE_SSE_AVX_RM_DIS_CASE)
 #undef DECLARE_SSE_AVX_RM_DIS_CASE
+
+#define DISASSEMBLE_AVX2_BROADCAST(instruction, _1, _2, _3, code)     \
+  case 0x##code:                                                      \
+    AppendToBuffer("" #instruction " %s,", NameOfXMMRegister(regop)); \
+    current += PrintRightXMMOperand(current);                         \
+    break;
+        AVX2_BROADCAST_LIST(DISASSEMBLE_AVX2_BROADCAST)
+#undef DISASSEMBLE_AVX2_BROADCAST
+
       default:
         UnimplementedInstruction();
     }

src/wasm/baseline/ia32/liftoff-assembler-ia32.h

@@ -2917,9 +2917,7 @@ void LiftoffAssembler::emit_i8x16_popcnt(LiftoffRegister dst,
 
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
-  Movd(dst.fp(), src.gp());
-  Pxor(liftoff::kScratchDoubleReg, liftoff::kScratchDoubleReg);
-  Pshufb(dst.fp(), liftoff::kScratchDoubleReg);
+  I8x16Splat(dst.fp(), src.gp(), liftoff::kScratchDoubleReg);
 }
 
 void LiftoffAssembler::emit_i16x8_splat(LiftoffRegister dst,

src/wasm/baseline/x64/liftoff-assembler-x64.h

@@ -2502,9 +2502,7 @@ void LiftoffAssembler::emit_i8x16_popcnt(LiftoffRegister dst,
 
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
-  Movd(dst.fp(), src.gp());
-  Pxor(kScratchDoubleReg, kScratchDoubleReg);
-  Pshufb(dst.fp(), kScratchDoubleReg);
+  I8x16Splat(dst.fp(), src.gp(), kScratchDoubleReg);
 }
 
 void LiftoffAssembler::emit_i16x8_splat(LiftoffRegister dst,

test/cctest/test-disasm-ia32.cc

@@ -865,6 +865,18 @@ TEST(DisasmIa320) {
     }
   }
 
+  // AVX2 instructions.
+  {
+    if (CpuFeatures::IsSupported(AVX2)) {
+      CpuFeatureScope scope(&assm, AVX2);
+#define EMIT_AVX2_BROADCAST(instruction, notUsed1, notUsed2, notUsed3, \
+                            notUsed4)                                  \
+  __ instruction(xmm0, xmm1);                                          \
+  __ instruction(xmm0, Operand(ebx, ecx, times_4, 10000));
+      AVX2_BROADCAST_LIST(EMIT_AVX2_BROADCAST)
+    }
+  }
+
   // FMA3 instruction
   {
     if (CpuFeatures::IsSupported(FMA3)) {

test/cctest/test-disasm-x64.cc

@@ -869,13 +869,6 @@ TEST(DisasmX64) {
     if (CpuFeatures::IsSupported(AVX2)) {
       CpuFeatureScope scope(&assm, AVX2);
       __ vbroadcastss(xmm1, xmm2);
-    }
-  }
-
-  // AVX2 instructions.
-  {
-    if (CpuFeatures::IsSupported(AVX2)) {
-      CpuFeatureScope scope(&assm, AVX2);
 #define EMIT_AVX2_BROADCAST(instruction, notUsed1, notUsed2, notUsed3, \
                             notUsed4)                                  \
   __ instruction(xmm0, xmm1);                                          \