[wasm-relaxed-simd][x64] Prototype i8x16 relaxed swizzle

Bug: v8:12284
Change-Id: I3519dfa3302c5851585c7430a9930ccd81507090
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3206502
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Deepti Gandluri <gdeepti@chromium.org>
Cr-Commit-Position: refs/heads/main@{#77327}

src/compiler/backend/x64/instruction-selector-x64.cc

@@ -3687,13 +3687,18 @@ void InstructionSelector::VisitI8x16Shuffle(Node* node) { UNREACHABLE(); }
 void InstructionSelector::VisitI8x16Swizzle(Node* node) {
   InstructionCode op = kX64I8x16Swizzle;
 
-  auto m = V128ConstMatcher(node->InputAt(1));
-  if (m.HasResolvedValue()) {
-    // If the indices vector is a const, check if they are in range, or if the
-    // top bit is set, then we can avoid the paddusb in the codegen and simply
-    // emit a pshufb
-    auto imms = m.ResolvedValue().immediate();
-    op |= MiscField::encode(wasm::SimdSwizzle::AllInRangeOrTopBitSet(imms));
+  bool relaxed = OpParameter<bool>(node->op());
+  if (relaxed) {
+    op |= MiscField::encode(true);
+  } else {
+    auto m = V128ConstMatcher(node->InputAt(1));
+    if (m.HasResolvedValue()) {
+      // If the indices vector is a const, check if they are in range, or if the
+      // top bit is set, then we can avoid the paddusb in the codegen and simply
+      // emit a pshufb.
+      auto imms = m.ResolvedValue().immediate();
+      op |= MiscField::encode(wasm::SimdSwizzle::AllInRangeOrTopBitSet(imms));
+    }
   }
 
   X64OperandGenerator g(this);

src/compiler/machine-operator.cc

@@ -597,8 +597,7 @@ std::ostream& operator<<(std::ostream& os, TruncateKind kind) {
   V(I64x2AllTrue, Operator::kNoProperties, 1, 0, 1)                        \
   V(I32x4AllTrue, Operator::kNoProperties, 1, 0, 1)                        \
   V(I16x8AllTrue, Operator::kNoProperties, 1, 0, 1)                        \
-  V(I8x16AllTrue, Operator::kNoProperties, 1, 0, 1)                        \
-  V(I8x16Swizzle, Operator::kNoProperties, 2, 0, 1)
+  V(I8x16AllTrue, Operator::kNoProperties, 1, 0, 1)
 
 // The format is:
 // V(Name, properties, value_input_count, control_input_count, output_count)
@@ -1282,6 +1281,19 @@ struct MachineOperatorGlobalCache {
   STACK_POINTER_GREATER_THAN(CodeStubAssembler)
   STACK_POINTER_GREATER_THAN(Wasm)
 #undef STACK_POINTER_GREATER_THAN
+
+  struct I8x16SwizzleOperator final : public Operator1<bool> {
+    I8x16SwizzleOperator()
+        : Operator1<bool>(IrOpcode::kI8x16Swizzle, Operator::kPure,
+                          "I8x16Swizzle", 2, 0, 0, 1, 0, 0, false) {}
+  };
+  I8x16SwizzleOperator kI8x16Swizzle;
+  struct I8x16RelaxedSwizzleOperator final : public Operator1<bool> {
+    I8x16RelaxedSwizzleOperator()
+        : Operator1<bool>(IrOpcode::kI8x16Swizzle, Operator::kPure,
+                          "I8x16RelaxedSwizzle", 2, 0, 0, 1, 0, 0, true) {}
+  };
+  I8x16RelaxedSwizzleOperator kI8x16RelaxedSwizzle;
 };
 
 struct CommentOperator : public Operator1<const char*> {
@@ -2003,6 +2015,14 @@ const Operator* MachineOperatorBuilder::I8x16Shuffle(
       S128ImmediateParameter(shuffle));
 }
 
+const Operator* MachineOperatorBuilder::I8x16Swizzle(bool relaxed) {
+  if (relaxed) {
+    return &cache_.kI8x16RelaxedSwizzle;
+  } else {
+    return &cache_.kI8x16Swizzle;
+  }
+}
+
 StackCheckKind StackCheckKindOf(Operator const* op) {
   DCHECK_EQ(IrOpcode::kStackPointerGreaterThan, op->opcode());
   return OpParameter<StackCheckKind>(op);

src/compiler/machine-operator.h

@@ -905,7 +905,7 @@ class V8_EXPORT_PRIVATE MachineOperatorBuilder final
   const Operator* S128Select();
   const Operator* S128AndNot();
 
-  const Operator* I8x16Swizzle();
+  const Operator* I8x16Swizzle(bool relaxed = false);
   const Operator* I8x16Shuffle(const uint8_t shuffle[16]);
 
   const Operator* V128AnyTrue();

src/compiler/wasm-compiler.cc

@@ -5111,8 +5111,11 @@ Node* WasmGraphBuilder::SimdOp(wasm::WasmOpcode opcode, Node* const* inputs) {
     case wasm::kExprI8x16AllTrue:
       return graph()->NewNode(mcgraph()->machine()->I8x16AllTrue(), inputs[0]);
     case wasm::kExprI8x16Swizzle:
-      return graph()->NewNode(mcgraph()->machine()->I8x16Swizzle(), inputs[0],
-                              inputs[1]);
+      return graph()->NewNode(mcgraph()->machine()->I8x16Swizzle(false),
+                              inputs[0], inputs[1]);
+    case wasm::kExprI8x16RelaxedSwizzle:
+      return graph()->NewNode(mcgraph()->machine()->I8x16Swizzle(true),
+                              inputs[0], inputs[1]);
     default:
       FATAL_UNSUPPORTED_OPCODE(opcode);
   }

src/wasm/wasm-opcodes-inl.h

@@ -259,8 +259,6 @@ constexpr const char* WasmOpcodes::OpcodeName(WasmOpcode opcode) {
     CASE_SIMDF_OP(Gt, "gt")
     CASE_SIMDF_OP(Ge, "ge")
     CASE_SIMDF_OP(Abs, "abs")
-    CASE_F32x4_OP(RecipApprox, "recip_approx")
-    CASE_F32x4_OP(RecipSqrtApprox, "recip_sqrt_approx")
     CASE_SIMDF_OP(Min, "min")
     CASE_SIMDF_OP(Max, "max")
     CASE_CONVERT_OP(Convert, F32x4, I32x4, "i32x4", "convert")
@@ -306,8 +304,6 @@ constexpr const char* WasmOpcodes::OpcodeName(WasmOpcode opcode) {
     CASE_I8x16_OP(Shuffle, "shuffle")
     CASE_V128_OP(AnyTrue, "any_true")
     CASE_SIMDI_OP(AllTrue, "all_true")
-    CASE_SIMDF_OP(Qfma, "qfma")
-    CASE_SIMDF_OP(Qfms, "qfms")
 
     CASE_S128_OP(Load32Zero, "load32_zero")
     CASE_S128_OP(Load64Zero, "load64_zero")
@@ -366,6 +362,13 @@ constexpr const char* WasmOpcodes::OpcodeName(WasmOpcode opcode) {
     CASE_F32x4_OP(DemoteF64x2Zero, "demote_f64x2_zero")
     CASE_F64x2_OP(PromoteLowF32x4, "promote_low_f32x4")
 
+    // Relaxed SIMD opcodes.
+    CASE_F32x4_OP(RecipApprox, "recip_approx")
+    CASE_F32x4_OP(RecipSqrtApprox, "recip_sqrt_approx")
+    CASE_SIMDF_OP(Qfma, "qfma")
+    CASE_SIMDF_OP(Qfms, "qfms")
+    CASE_I8x16_OP(RelaxedSwizzle, "relaxed_swizzle");
+
     // Atomic operations.
     CASE_OP(AtomicNotify, "atomic.notify")
     CASE_INT_OP(AtomicWait, "atomic.wait")

src/wasm/wasm-opcodes.h

@@ -515,6 +515,7 @@ bool V8_EXPORT_PRIVATE IsJSCompatibleSignature(const FunctionSig* sig,
   V(F64x2ConvertLowI32x4U, 0xfdff, s_s)
 
 #define FOREACH_RELAXED_SIMD_OPCODE(V) \
+  V(I8x16RelaxedSwizzle, 0xfda2, s_ss) \
   V(F32x4Qfma, 0xfdaf, s_sss)          \
   V(F32x4Qfms, 0xfdb0, s_sss)          \
   V(F64x2Qfma, 0xfdcf, s_sss)          \

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

@@ -230,6 +230,30 @@ WASM_RELAXED_SIMD_TEST(F32x4RecipSqrtApprox) {
                    false /* !exact */);
 }
 
+#if V8_TARGET_ARCH_X64
+WASM_RELAXED_SIMD_TEST(I8x16RelaxedSwizzle) {
+  // Output is only defined for indices in the range [0,15].
+  WasmRunner<int32_t> r(execution_tier);
+  static const int kElems = kSimd128Size / sizeof(uint8_t);
+  uint8_t* dst = r.builder().AddGlobal<uint8_t>(kWasmS128);
+  uint8_t* src = r.builder().AddGlobal<uint8_t>(kWasmS128);
+  uint8_t* indices = r.builder().AddGlobal<uint8_t>(kWasmS128);
+  BUILD(r,
+        WASM_GLOBAL_SET(
+            0, WASM_SIMD_BINOP(kExprI8x16RelaxedSwizzle, WASM_GLOBAL_GET(1),
+                               WASM_GLOBAL_GET(2))),
+        WASM_ONE);
+  for (int i = 0; i < kElems; i++) {
+    LANE(src, i) = kElems - i - 1;
+    LANE(indices, i) = kElems - i - 1;
+  }
+  CHECK_EQ(1, r.Call());
+  for (int i = 0; i < kElems; i++) {
+    CHECK_EQ(LANE(dst, i), i);
+  }
+}
+#endif  // V8_TARGET_ARCH_X64
+
 #undef WASM_RELAXED_SIMD_TEST
 }  // namespace test_run_wasm_relaxed_simd
 }  // namespace wasm

test/common/wasm/wasm-interpreter.cc

@@ -2761,6 +2761,7 @@ class WasmInterpreterInternals {
         *len += 16;
         return true;
       }
+      case kExprI8x16RelaxedSwizzle:
       case kExprI8x16Swizzle: {
         int16 v2 = Pop().to_s128().to_i8x16();
         int16 v1 = Pop().to_s128().to_i8x16();