[wasm] Implement I32x4ReplaceLane, I32x4Add, I32x4Sub.

R=bbudge@chromium.org, titzer@chromium.org

Review-Url: https://codereview.chromium.org/2385393002
Cr-Commit-Position: refs/heads/master@{#41505}

src/compiler/common-operator.cc

@@ -1131,6 +1131,43 @@ const Operator* CommonOperatorBuilder::ResizeMergeOrPhi(const Operator* op,
   }
 }
 
+const Operator* CommonOperatorBuilder::Int32x4ExtractLane(int32_t lane_number) {
+  DCHECK(0 <= lane_number && lane_number < 4);
+  return new (zone()) Operator1<int32_t>(              // --
+      IrOpcode::kInt32x4ExtractLane, Operator::kPure,  // opcode
+      "Int32x4ExtractLane",                            // name
+      1, 0, 0, 1, 0, 0,                                // counts
+      lane_number);                                    // parameter
+}
+
+const Operator* CommonOperatorBuilder::Int32x4ReplaceLane(int32_t lane_number) {
+  DCHECK(0 <= lane_number && lane_number < 4);
+  return new (zone()) Operator1<int32_t>(              // --
+      IrOpcode::kInt32x4ReplaceLane, Operator::kPure,  // opcode
+      "Int32x4ReplaceLane",                            // name
+      2, 0, 0, 1, 0, 0,                                // counts
+      lane_number);                                    // parameter
+}
+
+const Operator* CommonOperatorBuilder::Float32x4ExtractLane(
+    int32_t lane_number) {
+  DCHECK(0 <= lane_number && lane_number < 4);
+  return new (zone()) Operator1<int32_t>(                // --
+      IrOpcode::kFloat32x4ExtractLane, Operator::kPure,  // opcode
+      "Float32x4ExtractLane",                            // name
+      1, 0, 0, 1, 0, 0,                                  // counts
+      lane_number);                                      // parameter
+}
+
+const Operator* CommonOperatorBuilder::Float32x4ReplaceLane(
+    int32_t lane_number) {
+  DCHECK(0 <= lane_number && lane_number < 4);
+  return new (zone()) Operator1<int32_t>(                // --
+      IrOpcode::kFloat32x4ReplaceLane, Operator::kPure,  // opcode
+      "Float32x4ReplaceLane",                            // name
+      2, 0, 0, 1, 0, 0,                                  // counts
+      lane_number);                                      // parameter
+}
 
 const FrameStateFunctionInfo*
 CommonOperatorBuilder::CreateFrameStateFunctionInfo(

src/compiler/common-operator.h

@@ -260,6 +260,12 @@ class V8_EXPORT_PRIVATE CommonOperatorBuilder final
   // with {size} inputs.
   const Operator* ResizeMergeOrPhi(const Operator* op, int size);
 
+  // Simd Operators
+  const Operator* Int32x4ExtractLane(int32_t);
+  const Operator* Int32x4ReplaceLane(int32_t);
+  const Operator* Float32x4ExtractLane(int32_t);
+  const Operator* Float32x4ReplaceLane(int32_t);
+
   // Constructs function info for frame state construction.
   const FrameStateFunctionInfo* CreateFrameStateFunctionInfo(
       FrameStateType type, int parameter_count, int local_count,

src/compiler/instruction-selector.cc

@@ -1396,6 +1396,12 @@ void InstructionSelector::VisitNode(Node* node) {
       return MarkAsSimd128(node), VisitCreateInt32x4(node);
     case IrOpcode::kInt32x4ExtractLane:
       return MarkAsWord32(node), VisitInt32x4ExtractLane(node);
+    case IrOpcode::kInt32x4ReplaceLane:
+      return MarkAsSimd128(node), VisitInt32x4ReplaceLane(node);
+    case IrOpcode::kInt32x4Add:
+      return MarkAsSimd128(node), VisitInt32x4Add(node);
+    case IrOpcode::kInt32x4Sub:
+      return MarkAsSimd128(node), VisitInt32x4Sub(node);
     default:
       V8_Fatal(__FILE__, __LINE__, "Unexpected operator #%d:%s @ node #%d",
                node->opcode(), node->op()->mnemonic(), node->id());
@@ -1729,6 +1735,14 @@ void InstructionSelector::VisitCreateInt32x4(Node* node) { UNIMPLEMENTED(); }
 void InstructionSelector::VisitInt32x4ExtractLane(Node* node) {
   UNIMPLEMENTED();
 }
+
+void InstructionSelector::VisitInt32x4ReplaceLane(Node* node) {
+  UNIMPLEMENTED();
+}
+
+void InstructionSelector::VisitInt32x4Add(Node* node) { UNIMPLEMENTED(); }
+
+void InstructionSelector::VisitInt32x4Sub(Node* node) { UNIMPLEMENTED(); }
 #endif  // !V8_TARGET_ARCH_X64
 
 void InstructionSelector::VisitFinishRegion(Node* node) { EmitIdentity(node); }

src/compiler/simd-scalar-lowering.cc

@@ -399,9 +399,7 @@ void SimdScalarLowering::LowerNode(Node* node) {
     }
     case IrOpcode::kInt32x4ExtractLane:
     case IrOpcode::kFloat32x4ExtractLane: {
-      Node* laneNode = node->InputAt(1);
-      DCHECK_EQ(laneNode->opcode(), IrOpcode::kInt32Constant);
-      int32_t lane = OpParameter<int32_t>(laneNode);
+      int32_t lane = OpParameter<int32_t>(node);
       Node* rep_node[kMaxLanes] = {
           GetReplacementsWithType(node->InputAt(0), rep_type)[lane], nullptr,
           nullptr, nullptr};
@@ -410,11 +408,9 @@ void SimdScalarLowering::LowerNode(Node* node) {
     }
     case IrOpcode::kInt32x4ReplaceLane:
     case IrOpcode::kFloat32x4ReplaceLane: {
-      DCHECK_EQ(3, node->InputCount());
-      Node* laneNode = node->InputAt(1);
-      Node* repNode = node->InputAt(2);
-      DCHECK_EQ(laneNode->opcode(), IrOpcode::kInt32Constant);
-      int32_t lane = OpParameter<int32_t>(laneNode);
+      DCHECK_EQ(2, node->InputCount());
+      Node* repNode = node->InputAt(1);
+      int32_t lane = OpParameter<int32_t>(node);
       DCHECK(lane >= 0 && lane <= 3);
       Node** rep_node = GetReplacementsWithType(node->InputAt(0), rep_type);
       if (HasReplacement(0, repNode)) {

src/compiler/wasm-compiler.cc

@@ -3124,6 +3124,9 @@ Node* WasmGraphBuilder::SimdOp(wasm::WasmOpcode opcode,
     case wasm::kExprI32x4Add:
       return graph()->NewNode(jsgraph()->machine()->Int32x4Add(), inputs[0],
                               inputs[1]);
+    case wasm::kExprI32x4Sub:
+      return graph()->NewNode(jsgraph()->machine()->Int32x4Sub(), inputs[0],
+                              inputs[1]);
     case wasm::kExprF32x4Splat:
       return graph()->NewNode(jsgraph()->machine()->CreateFloat32x4(),
                               inputs[0], inputs[0], inputs[0], inputs[0]);
@@ -3135,29 +3138,21 @@ Node* WasmGraphBuilder::SimdOp(wasm::WasmOpcode opcode,
   }
 }
 
-Node* WasmGraphBuilder::SimdExtractLane(wasm::WasmOpcode opcode, uint8_t lane,
-                                        Node* input) {
+Node* WasmGraphBuilder::SimdLaneOp(wasm::WasmOpcode opcode, uint8_t lane,
+                                   const NodeVector& inputs) {
   switch (opcode) {
     case wasm::kExprI32x4ExtractLane:
-      return graph()->NewNode(jsgraph()->machine()->Int32x4ExtractLane(), input,
-                              Int32Constant(lane));
-    case wasm::kExprF32x4ExtractLane:
-      return graph()->NewNode(jsgraph()->machine()->Float32x4ExtractLane(),
-                              input, Int32Constant(lane));
-    default:
-      return graph()->NewNode(UnsupportedOpcode(opcode), nullptr);
-  }
-}
-
-Node* WasmGraphBuilder::SimdReplaceLane(wasm::WasmOpcode opcode, uint8_t lane,
-                                        Node* input, Node* replacement) {
-  switch (opcode) {
+      return graph()->NewNode(jsgraph()->common()->Int32x4ExtractLane(lane),
+                              inputs[0]);
     case wasm::kExprI32x4ReplaceLane:
-      return graph()->NewNode(jsgraph()->machine()->Int32x4ReplaceLane(), input,
-                              Int32Constant(lane), replacement);
+      return graph()->NewNode(jsgraph()->common()->Int32x4ReplaceLane(lane),
+                              inputs[0], inputs[1]);
+    case wasm::kExprF32x4ExtractLane:
+      return graph()->NewNode(jsgraph()->common()->Float32x4ExtractLane(lane),
+                              inputs[0]);
     case wasm::kExprF32x4ReplaceLane:
-      return graph()->NewNode(jsgraph()->machine()->Float32x4ReplaceLane(),
-                              input, Int32Constant(lane), replacement);
+      return graph()->NewNode(jsgraph()->common()->Float32x4ReplaceLane(lane),
+                              inputs[0], inputs[1]);
     default:
       return graph()->NewNode(UnsupportedOpcode(opcode), nullptr);
   }

src/compiler/wasm-compiler.h

@@ -217,9 +217,9 @@ class WasmGraphBuilder {
   Node* CreateS128Value(int32_t value);
 
   Node* SimdOp(wasm::WasmOpcode opcode, const NodeVector& inputs);
-  Node* SimdExtractLane(wasm::WasmOpcode opcode, uint8_t lane, Node* input);
-  Node* SimdReplaceLane(wasm::WasmOpcode opcode, uint8_t lane, Node* input,
-                        Node* replacement);
+
+  Node* SimdLaneOp(wasm::WasmOpcode opcode, uint8_t lane,
+                   const NodeVector& inputs);
 
  private:
   static const int kDefaultBufferSize = 16;

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

@@ -2147,6 +2147,26 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       __ Pextrd(i.OutputRegister(), i.InputSimd128Register(0), i.InputInt8(1));
       break;
     }
+    case kX64Int32x4ReplaceLane: {
+      CpuFeatureScope sse_scope(masm(), SSE4_1);
+      if (instr->InputAt(2)->IsRegister()) {
+        __ Pinsrd(i.OutputSimd128Register(), i.InputRegister(2),
+                  i.InputInt8(1));
+      } else {
+        __ Pinsrd(i.OutputSimd128Register(), i.InputOperand(2), i.InputInt8(1));
+      }
+      break;
+    }
+    case kX64Int32x4Add: {
+      CpuFeatureScope sse_scope(masm(), SSE4_1);
+      __ paddd(i.OutputSimd128Register(), i.InputSimd128Register(1));
+      break;
+    }
+    case kX64Int32x4Sub: {
+      CpuFeatureScope sse_scope(masm(), SSE4_1);
+      __ psubd(i.OutputSimd128Register(), i.InputSimd128Register(1));
+      break;
+    }
     case kCheckedLoadInt8:
       ASSEMBLE_CHECKED_LOAD_INTEGER(movsxbl);
       break;

src/compiler/x64/instruction-codes-x64.h

@@ -147,7 +147,10 @@ namespace compiler {
   V(X64Xchgw)                      \
   V(X64Xchgl)                      \
   V(X64Int32x4Create)              \
-  V(X64Int32x4ExtractLane)
+  V(X64Int32x4ExtractLane)         \
+  V(X64Int32x4ReplaceLane)         \
+  V(X64Int32x4Add)                 \
+  V(X64Int32x4Sub)
 
 // Addressing modes represent the "shape" of inputs to an instruction.
 // Many instructions support multiple addressing modes. Addressing modes

src/compiler/x64/instruction-scheduler-x64.cc

@@ -125,6 +125,9 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kX64Inc32:
     case kX64Int32x4Create:
     case kX64Int32x4ExtractLane:
+    case kX64Int32x4ReplaceLane:
+    case kX64Int32x4Add:
+    case kX64Int32x4Sub:
       return (instr->addressing_mode() == kMode_None)
           ? kNoOpcodeFlags
           : kIsLoadOperation | kHasSideEffect;

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

@@ -2379,8 +2379,29 @@ void InstructionSelector::VisitCreateInt32x4(Node* node) {
 
 void InstructionSelector::VisitInt32x4ExtractLane(Node* node) {
   X64OperandGenerator g(this);
+  int32_t lane = OpParameter<int32_t>(node);
   Emit(kX64Int32x4ExtractLane, g.DefineAsRegister(node),
-       g.UseRegister(node->InputAt(0)), g.UseImmediate(node->InputAt(1)));
+       g.UseRegister(node->InputAt(0)), g.UseImmediate(lane));
+}
+
+void InstructionSelector::VisitInt32x4ReplaceLane(Node* node) {
+  X64OperandGenerator g(this);
+  int32_t lane = OpParameter<int32_t>(node);
+  Emit(kX64Int32x4ReplaceLane, g.DefineSameAsFirst(node),
+       g.UseRegister(node->InputAt(0)), g.UseImmediate(lane),
+       g.Use(node->InputAt(1)));
+}
+
+void InstructionSelector::VisitInt32x4Add(Node* node) {
+  X64OperandGenerator g(this);
+  Emit(kX64Int32x4Add, g.DefineSameAsFirst(node),
+       g.UseRegister(node->InputAt(0)), g.UseRegister(node->InputAt(1)));
+}
+
+void InstructionSelector::VisitInt32x4Sub(Node* node) {
+  X64OperandGenerator g(this);
+  Emit(kX64Int32x4Sub, g.DefineSameAsFirst(node),
+       g.UseRegister(node->InputAt(0)), g.UseRegister(node->InputAt(1)));
 }
 
 // static

src/wasm/ast-decoder.cc

@@ -1331,8 +1331,9 @@ class WasmFullDecoder : public WasmDecoder {
   unsigned ExtractLane(WasmOpcode opcode, LocalType type) {
     LaneOperand operand(this, pc_);
     if (Validate(pc_, operand)) {
-      TFNode* input = Pop(0, LocalType::kSimd128).node;
-      TFNode* node = BUILD(SimdExtractLane, opcode, operand.lane, input);
+      compiler::NodeVector inputs(1, zone_);
+      inputs[0] = Pop(0, LocalType::kSimd128).node;
+      TFNode* node = BUILD(SimdLaneOp, opcode, operand.lane, inputs);
       Push(type, node);
     }
     return operand.length;
@@ -1341,10 +1342,10 @@ class WasmFullDecoder : public WasmDecoder {
   unsigned ReplaceLane(WasmOpcode opcode, LocalType type) {
     LaneOperand operand(this, pc_);
     if (Validate(pc_, operand)) {
-      TFNode* input = Pop(0, LocalType::kSimd128).node;
-      TFNode* replacement = Pop(1, type).node;
-      TFNode* node =
-          BUILD(SimdReplaceLane, opcode, operand.lane, input, replacement);
+      compiler::NodeVector inputs(2, zone_);
+      inputs[1] = Pop(1, type).node;
+      inputs[0] = Pop(0, LocalType::kSimd128).node;
+      TFNode* node = BUILD(SimdLaneOp, opcode, operand.lane, inputs);
       Push(LocalType::kSimd128, node);
     }
     return operand.length;

src/wasm/wasm-macro-gen.h

@@ -624,15 +624,16 @@ class LocalDeclEncoder {
 #define WASM_SIMD_I32x4_SPLAT(x) x, kSimdPrefix, kExprI32x4Splat & 0xff
 #define WASM_SIMD_I32x4_EXTRACT_LANE(lane, x) \
   x, kSimdPrefix, kExprI32x4ExtractLane & 0xff, static_cast<byte>(lane)
+#define WASM_SIMD_I32x4_REPLACE_LANE(lane, x, y) \
+  x, y, kSimdPrefix, kExprI32x4ReplaceLane & 0xff, static_cast<byte>(lane)
 #define WASM_SIMD_I32x4_ADD(x, y) x, y, kSimdPrefix, kExprI32x4Add & 0xff
+#define WASM_SIMD_I32x4_SUB(x, y) x, y, kSimdPrefix, kExprI32x4Sub & 0xff
 #define WASM_SIMD_F32x4_SPLAT(x) x, kSimdPrefix, kExprF32x4Splat & 0xff
 #define WASM_SIMD_F32x4_EXTRACT_LANE(lane, x) \
   x, kSimdPrefix, kExprF32x4ExtractLane & 0xff, static_cast<byte>(lane)
 #define WASM_SIMD_F32x4_ADD(x, y) x, y, kSimdPrefix, kExprF32x4Add & 0xff
-#define WASM_SIMD_I32x4_REPLACE_LANE(lane, x, y) \
-  y, x, kSimdPrefix, kExprI32x4ReplaceLane & 0xff, static_cast<byte>(lane)
 #define WASM_SIMD_F32x4_REPLACE_LANE(lane, x, y) \
-  y, x, kSimdPrefix, kExprF32x4ReplaceLane & 0xff, static_cast<byte>(lane)
+  x, y, kSimdPrefix, kExprF32x4ReplaceLane & 0xff, static_cast<byte>(lane)
 
 #define SIG_ENTRY_v_v kWasmFunctionTypeForm, 0, 0
 #define SIZEOF_SIG_ENTRY_v_v 3

src/wasm/wasm-opcodes.h

@@ -392,11 +392,11 @@ const WasmCodePosition kNoCodePosition = -1;
 
 #define FOREACH_SIMD_1_OPERAND_OPCODE(V) \
   V(F32x4ExtractLane, 0xe501, _)         \
+  V(F32x4ReplaceLane, 0xe502, _)         \
   V(I32x4ExtractLane, 0xe51c, _)         \
+  V(I32x4ReplaceLane, 0xe51d, _)         \
   V(I16x8ExtractLane, 0xe539, _)         \
   V(I8x16ExtractLane, 0xe558, _)         \
-  V(F32x4ReplaceLane, 0xe502, _)         \
-  V(I32x4ReplaceLane, 0xe51d, _)         \
   V(I16x8ReplaceLane, 0xe53a, _)         \
   V(I8x16ReplaceLane, 0xe559, _)
 
@@ -484,11 +484,9 @@ const WasmCodePosition kNoCodePosition = -1;
 #define FOREACH_SIMD_SIGNATURE(V)                  \
   V(s_s, kAstS128, kAstS128)                       \
   V(s_f, kAstS128, kAstF32)                        \
-  V(s_sif, kAstS128, kAstS128, kAstI32, kAstF32)   \
   V(s_ss, kAstS128, kAstS128, kAstS128)            \
   V(s_sss, kAstS128, kAstS128, kAstS128, kAstS128) \
   V(s_i, kAstS128, kAstI32)                        \
-  V(s_sii, kAstS128, kAstS128, kAstI32, kAstI32)   \
   V(s_si, kAstS128, kAstS128, kAstI32)
 
 #define FOREACH_PREFIX(V) \

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

@@ -13,7 +13,26 @@ using namespace v8::internal;
 using namespace v8::internal::compiler;
 using namespace v8::internal::wasm;
 
-WASM_EXEC_TEST(Splat) {
+// TODO(gdeepti): These are tests using sample values to verify functional
+// correctness of opcodes, add more tests for a range of values and macroize
+// tests.
+
+#define WASM_SIMD_CHECK_LANE(TYPE, value, LANE_TYPE, lane_value, lane_index) \
+  WASM_IF(WASM_##LANE_TYPE##_NE(WASM_GET_LOCAL(lane_value),                  \
+                                WASM_SIMD_##TYPE##_EXTRACT_LANE(             \
+                                    lane_index, WASM_GET_LOCAL(value))),     \
+          WASM_RETURN1(WASM_ZERO))
+
+#define WASM_SIMD_CHECK4(TYPE, value, LANE_TYPE, lv0, lv1, lv2, lv3) \
+  WASM_SIMD_CHECK_LANE(TYPE, value, LANE_TYPE, lv0, 0)               \
+  , WASM_SIMD_CHECK_LANE(TYPE, value, LANE_TYPE, lv1, 1),            \
+      WASM_SIMD_CHECK_LANE(TYPE, value, LANE_TYPE, lv2, 2),          \
+      WASM_SIMD_CHECK_LANE(TYPE, value, LANE_TYPE, lv3, 3)
+
+#define WASM_SIMD_CHECK_SPLAT4(TYPE, value, LANE_TYPE, lv) \
+  WASM_SIMD_CHECK4(TYPE, value, LANE_TYPE, lv, lv, lv, lv)
+
+WASM_EXEC_TEST(I32x4Splat) {
   FLAG_wasm_simd_prototype = true;
 
   // Store SIMD value in a local variable, use extract lane to check lane values
@@ -27,23 +46,92 @@ WASM_EXEC_TEST(Splat) {
   //
   // return 1
   WasmRunner<int32_t> r(kExecuteCompiled, MachineType::Int32());
-  r.AllocateLocal(kAstS128);
+  byte lane_val = 0;
+  byte simd = r.AllocateLocal(kAstS128);
+  BUILD(r, WASM_BLOCK(WASM_SET_LOCAL(simd, WASM_SIMD_I32x4_SPLAT(
+                                               WASM_GET_LOCAL(lane_val))),
+                      WASM_SIMD_CHECK_SPLAT4(I32x4, simd, I32, lane_val),
+                      WASM_RETURN1(WASM_ONE)));
+
+  FOR_INT32_INPUTS(i) { CHECK_EQ(1, r.Call(*i)); }
+}
+
+WASM_EXEC_TEST(I32x4ReplaceLane) {
+  FLAG_wasm_simd_prototype = true;
+  WasmRunner<int32_t> r(kExecuteCompiled, MachineType::Int32(),
+                        MachineType::Int32());
+  byte old_val = 0;
+  byte new_val = 1;
+  byte simd = r.AllocateLocal(kAstS128);
+  BUILD(r, WASM_BLOCK(
+               WASM_SET_LOCAL(simd,
+                              WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(old_val))),
+               WASM_SET_LOCAL(
+                   simd, WASM_SIMD_I32x4_REPLACE_LANE(0, WASM_GET_LOCAL(simd),
+                                                      WASM_GET_LOCAL(new_val))),
+               WASM_SIMD_CHECK4(I32x4, simd, I32, new_val, old_val, old_val,
+                                old_val),
+               WASM_SET_LOCAL(
+                   simd, WASM_SIMD_I32x4_REPLACE_LANE(1, WASM_GET_LOCAL(simd),
+                                                      WASM_GET_LOCAL(new_val))),
+               WASM_SIMD_CHECK4(I32x4, simd, I32, new_val, new_val, old_val,
+                                old_val),
+               WASM_SET_LOCAL(
+                   simd, WASM_SIMD_I32x4_REPLACE_LANE(2, WASM_GET_LOCAL(simd),
+                                                      WASM_GET_LOCAL(new_val))),
+               WASM_SIMD_CHECK4(I32x4, simd, I32, new_val, new_val, new_val,
+                                old_val),
+               WASM_SET_LOCAL(
+                   simd, WASM_SIMD_I32x4_REPLACE_LANE(3, WASM_GET_LOCAL(simd),
+                                                      WASM_GET_LOCAL(new_val))),
+               WASM_SIMD_CHECK_SPLAT4(I32x4, simd, I32, new_val),
+               WASM_RETURN1(WASM_ONE)));
+
+  CHECK_EQ(1, r.Call(1, 2));
+}
+
+WASM_EXEC_TEST(I32x4Add) {
+  FLAG_wasm_simd_prototype = true;
+  WasmRunner<int32_t> r(kExecuteCompiled, MachineType::Int32(),
+                        MachineType::Int32(), MachineType::Int32());
+  byte a = 0;
+  byte b = 1;
+  byte expected = 2;
+  byte simd0 = r.AllocateLocal(kAstS128);
+  byte simd1 = r.AllocateLocal(kAstS128);
   BUILD(r,
         WASM_BLOCK(
-            WASM_SET_LOCAL(1, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(0))),
-            WASM_IF(WASM_I32_NE(WASM_GET_LOCAL(0), WASM_SIMD_I32x4_EXTRACT_LANE(
-                                                       0, WASM_GET_LOCAL(1))),
-                    WASM_RETURN1(WASM_ZERO)),
-            WASM_IF(WASM_I32_NE(WASM_GET_LOCAL(0), WASM_SIMD_I32x4_EXTRACT_LANE(
-                                                       1, WASM_GET_LOCAL(1))),
-                    WASM_RETURN1(WASM_ZERO)),
-            WASM_IF(WASM_I32_NE(WASM_GET_LOCAL(0), WASM_SIMD_I32x4_EXTRACT_LANE(
-                                                       2, WASM_GET_LOCAL(1))),
-                    WASM_RETURN1(WASM_ZERO)),
-            WASM_IF(WASM_I32_NE(WASM_GET_LOCAL(0), WASM_SIMD_I32x4_EXTRACT_LANE(
-                                                       3, WASM_GET_LOCAL(1))),
-                    WASM_RETURN1(WASM_ZERO)),
+            WASM_SET_LOCAL(simd0, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(a))),
+            WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(b))),
+            WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_ADD(WASM_GET_LOCAL(simd0),
+                                                      WASM_GET_LOCAL(simd1))),
+            WASM_SIMD_CHECK_SPLAT4(I32x4, simd1, I32, expected),
             WASM_RETURN1(WASM_ONE)));
 
-  FOR_INT32_INPUTS(i) { CHECK_EQ(1, r.Call(*i)); }
+  FOR_INT32_INPUTS(i) {
+    FOR_INT32_INPUTS(j) { CHECK_EQ(1, r.Call(*i, *j, *i + *j)); }
+  }
+}
+
+WASM_EXEC_TEST(I32x4Sub) {
+  FLAG_wasm_simd_prototype = true;
+  WasmRunner<int32_t> r(kExecuteCompiled, MachineType::Int32(),
+                        MachineType::Int32(), MachineType::Int32());
+  byte a = 0;
+  byte b = 1;
+  byte expected = 2;
+  byte simd0 = r.AllocateLocal(kAstS128);
+  byte simd1 = r.AllocateLocal(kAstS128);
+  BUILD(r,
+        WASM_BLOCK(
+            WASM_SET_LOCAL(simd0, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(a))),
+            WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(b))),
+            WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_SUB(WASM_GET_LOCAL(simd0),
+                                                      WASM_GET_LOCAL(simd1))),
+            WASM_SIMD_CHECK_SPLAT4(I32x4, simd1, I32, expected),
+            WASM_RETURN1(WASM_ONE)));
+
+  FOR_INT32_INPUTS(i) {
+    FOR_INT32_INPUTS(j) { CHECK_EQ(1, r.Call(*i, *j, *i - *j)); }
+  }
 }