Turbofan: Modify WASM linkage to store floats using only 4 bytes.

Adds instructions for ARM to push floats.

LOG=N
BUG=v8:4124

Review-Url: https://codereview.chromium.org/2024443002
Cr-Commit-Position: refs/heads/master@{#36654}

src/arm/assembler-arm.h

@@ -1320,6 +1320,10 @@ class Assembler : public AssemblerBase {
     vstm(db_w, sp, src, src, cond);
   }
 
+  void vpush(SwVfpRegister src, Condition cond = al) {
+    vstm(db_w, sp, src, src, cond);
+  }
+
   void vpop(DwVfpRegister dst, Condition cond = al) {
     vldm(ia_w, sp, dst, dst, cond);
   }

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

@@ -1204,8 +1204,15 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
     }
     case kArmPush:
       if (instr->InputAt(0)->IsFPRegister()) {
-        __ vpush(i.InputDoubleRegister(0));
-        frame_access_state()->IncreaseSPDelta(kDoubleSize / kPointerSize);
+        LocationOperand* op = LocationOperand::cast(instr->InputAt(0));
+        if (op->representation() == MachineRepresentation::kFloat64) {
+          __ vpush(i.InputFloat64Register(0));
+          frame_access_state()->IncreaseSPDelta(kDoubleSize / kPointerSize);
+        } else {
+          DCHECK_EQ(MachineRepresentation::kFloat32, op->representation());
+          __ vpush(i.InputFloat32Register(0));
+          frame_access_state()->IncreaseSPDelta(1);
+        }
       } else {
         __ push(i.InputRegister(0));
         frame_access_state()->IncreaseSPDelta(1);

src/compiler/ia32/code-generator-ia32.cc

@@ -1230,9 +1230,9 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
     }
     case kIA32PushFloat32:
       if (instr->InputAt(0)->IsFPRegister()) {
-        __ sub(esp, Immediate(kDoubleSize));
+        __ sub(esp, Immediate(kFloatSize));
         __ movss(Operand(esp, 0), i.InputDoubleRegister(0));
-        frame_access_state()->IncreaseSPDelta(kDoubleSize / kPointerSize);
+        frame_access_state()->IncreaseSPDelta(kFloatSize / kPointerSize);
       } else if (HasImmediateInput(instr, 0)) {
         __ Move(kScratchDoubleReg, i.InputDouble(0));
         __ sub(esp, Immediate(kDoubleSize));
@@ -1264,9 +1264,9 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       break;
     case kIA32Push:
       if (instr->InputAt(0)->IsFPRegister()) {
-        __ sub(esp, Immediate(kDoubleSize));
+        __ sub(esp, Immediate(kFloatSize));
         __ movsd(Operand(esp, 0), i.InputDoubleRegister(0));
-        frame_access_state()->IncreaseSPDelta(kDoubleSize / kPointerSize);
+        frame_access_state()->IncreaseSPDelta(kFloatSize / kPointerSize);
       } else if (HasImmediateInput(instr, 0)) {
         __ push(i.InputImmediate(0));
         frame_access_state()->IncreaseSPDelta(1);

src/compiler/wasm-linkage.cc

@@ -199,11 +199,7 @@ struct Allocator {
     return type == kAstF32 || type == kAstF64;
   }
   int Words(LocalType type) {
-    // The code generation for pushing parameters on the stack does not
-    // distinguish between float32 and float64. Therefore also float32 needs
-    // two words.
-    if (kPointerSize < 8 &&
-        (type == kAstI64 || type == kAstF64 || type == kAstF32)) {
+    if (kPointerSize < 8 && (type == kAstI64 || type == kAstF64)) {
       return 2;
     }
     return 1;

test/cctest/compiler/test-run-native-calls.cc

@@ -154,10 +154,7 @@ struct Allocator {
     }
   }
   int StackWords(MachineType type) {
-    // TODO(titzer): hack. float32 occupies 8 bytes on stack.
-    int size = IsFloatingPoint(type.representation())
-                   ? kDoubleSize
-                   : (1 << ElementSizeLog2Of(type.representation()));
+    int size = 1 << ElementSizeLog2Of(type.representation());
     return size <= kPointerSize ? 1 : size / kPointerSize;
   }
   void Reset() {