[wasm] simd scalar lowering F32x4Add and I32x4Add

src/wasm/ast-decoder.cc

 // Copyright 2015 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/signature.h"
 
 #include "src/bit-vector.h"
 #include "src/flags.h"
 #include "src/handles.h"
 #include "src/zone/zone-containers.h"
@@ skipping 132 matching lines @@
   }
 };
 
 // Macros that build nodes only if there is a graph and the current SSA
 // environment is reachable from start. This avoids problems with malformed
 // TF graphs when decoding inputs that have unreachable code.
 #define BUILD(func, ...) \
   (build() ? CheckForException(builder_->func(__VA_ARGS__)) : nullptr)
 #define BUILD0(func) (build() ? CheckForException(builder_->func()) : nullptr)
 
+struct LaneOperand {
+  uint8_t lane;
+  unsigned length;
+
+  inline LaneOperand(Decoder* decoder, const byte* pc) {
+    lane = decoder->checked_read_u8(pc, 2, "lane");
+    length = 1;
+  }
+};
+
 // Generic Wasm bytecode decoder with utilities for decoding operands,
 // lengths, etc.
 class WasmDecoder : public Decoder {
  public:
   WasmDecoder(ModuleEnv* module, FunctionSig* sig, const byte* start,
               const byte* end)
       : Decoder(start, end),
         module_(module),
         sig_(sig),
         total_locals_(0),
@@ skipping 70 matching lines @@
     error(pc, pc + 1, "invalid break depth: %u", operand.depth);
     return false;
   }
 
   bool Validate(const byte* pc, BranchTableOperand& operand,
                 size_t block_depth) {
     // TODO(titzer): add extra redundant validation for br_table here?
     return true;
   }
 
+  inline bool Validate(const byte* pc, LaneOperand& operand) {
+    if (operand.lane < 0 || operand.lane > 3) {
+      error(pc_, pc_ + 2, "invalid extract lane value");
+      return false;
+    } else {
+      return true;
+    }
+  }
+
   unsigned OpcodeLength(const byte* pc) {
-    switch (static_cast<WasmOpcode>(*pc)) {
+    switch (static_cast<byte>(*pc)) {
 #define DECLARE_OPCODE_CASE(name, opcode, sig) case kExpr##name:
       FOREACH_LOAD_MEM_OPCODE(DECLARE_OPCODE_CASE)
       FOREACH_STORE_MEM_OPCODE(DECLARE_OPCODE_CASE)
 #undef DECLARE_OPCODE_CASE
       {
         MemoryAccessOperand operand(this, pc, UINT32_MAX);
         return 1 + operand.length;
       }
       case kExprBr:
       case kExprBrIf: {
@@ skipping 42 matching lines @@
       case kExprI64Const: {
         ImmI64Operand operand(this, pc);
         return 1 + operand.length;
       }
       case kExprI8Const:
         return 2;
       case kExprF32Const:
         return 5;
       case kExprF64Const:
         return 9;
+      case kSimdPrefix: {
+        byte simd_index = *(pc + 1);
+        WasmOpcode opcode =
+            static_cast<WasmOpcode>(kSimdPrefix << 8 | simd_index);
+        switch (opcode) {
+#define DECLARE_OPCODE_CASE(name, opcode, sig) case kExpr##name:
+          FOREACH_SIMD_0_OPERAND_OPCODE(DECLARE_OPCODE_CASE)
+#undef DECLARE_OPCODE_CASE
+          {
+            return 2;
+          }
+#define DECLARE_OPCODE_CASE(name, opcode, sig) case kExpr##name:
+          FOREACH_SIMD_1_OPERAND_OPCODE(DECLARE_OPCODE_CASE)
+#undef DECLARE_OPCODE_CASE
+          {
+            return 3;
+          }
+          default:
+            UNREACHABLE();
+        }
+      }
       default:
         return 1;
     }
   }
 };
 
 static const int32_t kNullCatch = -1;
 
 // The full WASM decoder for bytecode. Both verifies bytecode and generates
 // a TurboFan IR graph.
@@ skipping 925 matching lines @@
   int DecodeStoreMem(LocalType type, MachineType mem_type) {
     MemoryAccessOperand operand(this, pc_,
                                 ElementSizeLog2Of(mem_type.representation()));
     Value val = Pop(1, type);
     Value index = Pop(0, kAstI32);
     BUILD(StoreMem, mem_type, index.node, operand.offset, operand.alignment,
           val.node, position());
     return 1 + operand.length;
   }
 
+  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);
+      Push(type, node);
+    }
+    return operand.length;
+  }
+
   unsigned DecodeSimdOpcode(WasmOpcode opcode) {
     unsigned len = 0;
     switch (opcode) {
       case kExprI32x4ExtractLane: {
-        uint8_t lane = this->checked_read_u8(pc_, 2, "lane number");
-        if (lane < 0 || lane > 3) {
-          error(pc_, pc_ + 2, "invalid extract lane value");
-        }
-        TFNode* input = Pop(0, LocalType::kSimd128).node;
-        TFNode* node = BUILD(SimdExtractLane, opcode, lane, input);
-        Push(LocalType::kWord32, node);
-        len++;
+        len = ExtractLane(opcode, LocalType::kWord32);
+        break;
+      }
+      case kExprF32x4ExtractLane: {
+        len = ExtractLane(opcode, LocalType::kFloat32);
         break;
       }
       default: {
         FunctionSig* sig = WasmOpcodes::Signature(opcode);
         if (sig != nullptr) {
           compiler::NodeVector inputs(sig->parameter_count(), zone_);
           for (size_t i = sig->parameter_count(); i > 0; i--) {
             Value val = Pop(static_cast<int>(i - 1), sig->GetParam(i - 1));
             inputs[i - 1] = val.node;
           }
@@ skipping 677 matching lines @@
 BitVector* AnalyzeLoopAssignmentForTesting(Zone* zone, size_t num_locals,
                                            const byte* start, const byte* end) {
   FunctionBody body = {nullptr, nullptr, nullptr, start, end};
   WasmFullDecoder decoder(zone, nullptr, body);
   return decoder.AnalyzeLoopAssignmentForTesting(start, num_locals);
 }
 
 }  // namespace wasm
 }  // namespace internal
 }  // namespace v8