[wasm] implement simd lowering for replaceLane, load, store and test for phi

BUG=v8:4124
TEST:test-run-wasm-simd-lowering
R=bradnelson@chromium.org,titzer@chromium.org,mtrofin@chromium.org

Review-Url: https://codereview.chromium.org/2498283002
Cr-Commit-Position: refs/heads/master@{#41443}

src/compiler/simd-scalar-lowering.h

@@ -31,6 +31,7 @@ class SimdScalarLowering {
   enum class SimdType : uint8_t { kInt32, kFloat32 };
 
   static const int kMaxLanes = 4;
+  static const int kLaneWidth = 16 / kMaxLanes;
 
   struct Replacement {
     Node* node[kMaxLanes];
@@ -53,6 +54,12 @@ class SimdScalarLowering {
   SimdType ReplacementType(Node* node);
   void PreparePhiReplacement(Node* phi);
   void SetLoweredType(Node* node, Node* output);
+  void GetIndexNodes(Node* index, Node** new_indices);
+  void LowerLoadOp(MachineRepresentation rep, Node* node,
+                   const Operator* load_op);
+  void LowerStoreOp(MachineRepresentation rep, Node* node,
+                    const Operator* store_op, SimdType rep_type);
+  void LowerBinaryOp(Node* node, SimdType rep_type, const Operator* op);
 
   struct NodeState {
     Node* node;

src/compiler/wasm-compiler.cc

@@ -3124,9 +3124,6 @@ Node* WasmGraphBuilder::SimdOp(wasm::WasmOpcode opcode,
     case wasm::kExprI32x4Add:
       return graph()->NewNode(jsgraph()->machine()->Int32x4Add(), inputs[0],
                               inputs[1]);
-    case wasm::kExprF32x4ExtractLane:
-      return graph()->NewNode(jsgraph()->machine()->Float32x4ExtractLane(),
-                              inputs[0], inputs[1]);
     case wasm::kExprF32x4Splat:
       return graph()->NewNode(jsgraph()->machine()->CreateFloat32x4(),
                               inputs[0], inputs[0], inputs[0], inputs[0]);
@@ -3152,6 +3149,20 @@ Node* WasmGraphBuilder::SimdExtractLane(wasm::WasmOpcode opcode, uint8_t lane,
   }
 }
 
+Node* WasmGraphBuilder::SimdReplaceLane(wasm::WasmOpcode opcode, uint8_t lane,
+                                        Node* input, Node* replacement) {
+  switch (opcode) {
+    case wasm::kExprI32x4ReplaceLane:
+      return graph()->NewNode(jsgraph()->machine()->Int32x4ReplaceLane(), input,
+                              Int32Constant(lane), replacement);
+    case wasm::kExprF32x4ReplaceLane:
+      return graph()->NewNode(jsgraph()->machine()->Float32x4ReplaceLane(),
+                              input, Int32Constant(lane), replacement);
+    default:
+      return graph()->NewNode(UnsupportedOpcode(opcode), nullptr);
+  }
+}
+
 static void RecordFunctionCompilation(CodeEventListener::LogEventsAndTags tag,
                                       Isolate* isolate, Handle<Code> code,
                                       const char* message, uint32_t index,

src/compiler/wasm-compiler.h

@@ -218,6 +218,8 @@ class WasmGraphBuilder {
 
   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);
 
  private:
   static const int kDefaultBufferSize = 16;

src/wasm/ast-decoder.cc

@@ -1338,6 +1338,18 @@ class WasmFullDecoder : public WasmDecoder {
     return operand.length;
   }
 
+  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);
+      Push(LocalType::kSimd128, node);
+    }
+    return operand.length;
+  }
+
   unsigned DecodeSimdOpcode(WasmOpcode opcode) {
     unsigned len = 0;
     switch (opcode) {
@@ -1349,6 +1361,14 @@ class WasmFullDecoder : public WasmDecoder {
         len = ExtractLane(opcode, LocalType::kFloat32);
         break;
       }
+      case kExprI32x4ReplaceLane: {
+        len = ReplaceLane(opcode, LocalType::kWord32);
+        break;
+      }
+      case kExprF32x4ReplaceLane: {
+        len = ReplaceLane(opcode, LocalType::kFloat32);
+        break;
+      }
       default: {
         FunctionSig* sig = WasmOpcodes::Signature(opcode);
         if (sig != nullptr) {

src/wasm/wasm-macro-gen.h

@@ -629,6 +629,10 @@ class LocalDeclEncoder {
 #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)
 
 #define SIG_ENTRY_v_v kWasmFunctionTypeForm, 0, 0
 #define SIZEOF_SIG_ENTRY_v_v 3

src/wasm/wasm-opcodes.h

@@ -276,7 +276,6 @@ const WasmCodePosition kNoCodePosition = -1;
 
 #define FOREACH_SIMD_0_OPERAND_OPCODE(V) \
   V(F32x4Splat, 0xe500, s_f)             \
-  V(F32x4ReplaceLane, 0xe502, s_sif)     \
   V(F32x4Abs, 0xe503, s_s)               \
   V(F32x4Neg, 0xe504, s_s)               \
   V(F32x4Sqrt, 0xe505, s_s)              \
@@ -299,7 +298,6 @@ const WasmCodePosition kNoCodePosition = -1;
   V(F32x4FromInt32x4, 0xe519, s_s)       \
   V(F32x4FromUint32x4, 0xe51a, s_s)      \
   V(I32x4Splat, 0xe51b, s_i)             \
-  V(I32x4ReplaceLane, 0xe51d, s_sii)     \
   V(I32x4Neg, 0xe51e, s_s)               \
   V(I32x4Add, 0xe51f, s_ss)              \
   V(I32x4Sub, 0xe520, s_ss)              \
@@ -327,7 +325,6 @@ const WasmCodePosition kNoCodePosition = -1;
   V(I32x4Ge_u, 0xe536, s_ss)             \
   V(Ui32x4FromFloat32x4, 0xe537, s_s)    \
   V(I16x8Splat, 0xe538, s_i)             \
-  V(I16x8ReplaceLane, 0xe53a, s_sii)     \
   V(I16x8Neg, 0xe53b, s_s)               \
   V(I16x8Add, 0xe53c, s_ss)              \
   V(I16x8AddSaturate_s, 0xe53d, s_ss)    \
@@ -357,7 +354,6 @@ const WasmCodePosition kNoCodePosition = -1;
   V(I16x8Gt_u, 0xe555, s_ss)             \
   V(I16x8Ge_u, 0xe556, s_ss)             \
   V(I8x16Splat, 0xe557, s_i)             \
-  V(I8x16ReplaceLane, 0xe559, s_sii)     \
   V(I8x16Neg, 0xe55a, s_s)               \
   V(I8x16Add, 0xe55b, s_ss)              \
   V(I8x16AddSaturate_s, 0xe55c, s_ss)    \
@@ -398,7 +394,11 @@ const WasmCodePosition kNoCodePosition = -1;
   V(F32x4ExtractLane, 0xe501, _)         \
   V(I32x4ExtractLane, 0xe51c, _)         \
   V(I16x8ExtractLane, 0xe539, _)         \
-  V(I8x16ExtractLane, 0xe558, _)
+  V(I8x16ExtractLane, 0xe558, _)         \
+  V(F32x4ReplaceLane, 0xe502, _)         \
+  V(I32x4ReplaceLane, 0xe51d, _)         \
+  V(I16x8ReplaceLane, 0xe53a, _)         \
+  V(I8x16ReplaceLane, 0xe559, _)
 
 #define FOREACH_ATOMIC_OPCODE(V)               \
   V(I32AtomicAdd8S, 0xe601, i_ii)              \

test/cctest/wasm/wasm-run-utils.h

@@ -791,15 +791,17 @@ class WasmRunner {
 };
 
 // A macro to define tests that run in different engine configurations.
-// Currently only supports compiled tests, but a future
-// RunWasmInterpreted_##name version will allow each test to also run in the
-// interpreter.
 #define WASM_EXEC_TEST(name)                                               \
   void RunWasm_##name(WasmExecutionMode execution_mode);                   \
   TEST(RunWasmCompiled_##name) { RunWasm_##name(kExecuteCompiled); }       \
   TEST(RunWasmInterpreted_##name) { RunWasm_##name(kExecuteInterpreted); } \
   void RunWasm_##name(WasmExecutionMode execution_mode)
 
+#define WASM_EXEC_COMPILED_TEST(name)                                \
+  void RunWasm_##name(WasmExecutionMode execution_mode);             \
+  TEST(RunWasmCompiled_##name) { RunWasm_##name(kExecuteCompiled); } \
+  void RunWasm_##name(WasmExecutionMode execution_mode)
+
 }  // namespace
 
 #endif