[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 223 matching lines @@
     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;
   }
 
   unsigned OpcodeLength(const byte* pc) {
-    switch (static_cast<WasmOpcode>(*pc)) {
+    WasmOpcode opcode = ExtractWasmOpcode(pc);

[titzer, 2016/10/12 19:06:55]

Could you do the trick in the main decoder loop of making the SIMD prefix one of
the cases here, so we save a branch in the common case?

[aseemgarg, 2016/10/17 19:01:44]

Done. Actually seems like we don't need this at all.

[titzer, 2016/10/19 18:31:49]

No, you do, since we still want to be able to iterate over SIMD bytecodes, which
are not a single byte.

[aseemgarg, 2016/10/19 23:37:33]

Done.

+    switch (opcode) {
 #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 46 matching lines @@
       case kExprI8Const:
         return 2;
       case kExprF32Const:
         return 5;
       case kExprF64Const:
         return 9;
       default:
         return 1;
     }
   }
+
+ private:
+  WasmOpcode ExtractWasmOpcode(const byte* pc) {
+    WasmOpcode opcode = static_cast<WasmOpcode>(*pc);
+    if (opcode == kSimdPrefix) {
+      opcode = static_cast<WasmOpcode>((opcode << 8) | *(pc + 1));
+    }
+    return opcode;
+  }
 };
 
 static const int32_t kNullCatch = -1;
 
 // The full WASM decoder for bytecode. Both verifies bytecode and generates
 // a TurboFan IR graph.
 class WasmFullDecoder : public WasmDecoder {
  public:
   WasmFullDecoder(Zone* zone, TFBuilder* builder, const FunctionBody& body)
       : WasmDecoder(body.module, body.sig, body.start, body.end),
@@ skipping 920 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) {

[titzer, 2016/10/12 19:06:55]

Please make a LaneOperand class similar to the others in wasm-opcodes.h

I think you can keep it in the .cc file for now, though.

[aseemgarg, 2016/10/17 19:01:44]

Done.

+    unsigned len = 0;
+    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(type, node);
+    len++;
+    return len;
+  }
+
   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 663 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