[turbofan] Add backend support for float32.

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

 // Copyright 2014 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/compiler/code-generator.h"
 
 #include "src/arm/macro-assembler-arm.h"
 #include "src/compiler/code-generator-impl.h"
 #include "src/compiler/gap-resolver.h"
 #include "src/compiler/node-matchers.h"
 #include "src/compiler/node-properties-inl.h"
 #include "src/scopes.h"
 
 namespace v8 {
 namespace internal {
 namespace compiler {
 
 #define __ masm()->
 
 
 #define kScratchReg r9
 
 
 // Adds Arm-specific methods to convert InstructionOperands.
-class ArmOperandConverter : public InstructionOperandConverter {
+class ArmOperandConverter FINAL : public InstructionOperandConverter {
  public:
   ArmOperandConverter(CodeGenerator* gen, Instruction* instr)
       : InstructionOperandConverter(gen, instr) {}
 
+  SwVfpRegister OutputFloat32Register(int index = 0) {
+    return ToFloat32Register(instr_->OutputAt(index));
+  }
+
+  SwVfpRegister InputFloat32Register(int index) {
+    return ToFloat32Register(instr_->InputAt(index));
+  }
+
+  SwVfpRegister ToFloat32Register(InstructionOperand* op) {
+    return ToFloat64Register(op).low();
+  }
+
+  LowDwVfpRegister OutputFloat64Register(int index = 0) {
+    return ToFloat64Register(instr_->OutputAt(index));
+  }
+
+  LowDwVfpRegister InputFloat64Register(int index) {
+    return ToFloat64Register(instr_->InputAt(index));
+  }
+
+  LowDwVfpRegister ToFloat64Register(InstructionOperand* op) {
+    return LowDwVfpRegister::from_code(ToDoubleRegister(op).code());
+  }
+
   SBit OutputSBit() const {
     switch (instr_->flags_mode()) {
       case kFlags_branch:
       case kFlags_set:
         return SetCC;
       case kFlags_none:
         return LeaveCC;
     }
     UNREACHABLE();
     return LeaveCC;
@@ skipping 131 matching lines @@
       break;
     case kArchNop:
       // don't emit code for nops.
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArchRet:
       AssembleReturn();
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArchTruncateDoubleToI:
-      __ TruncateDoubleToI(i.OutputRegister(), i.InputDoubleRegister(0));
+      __ TruncateDoubleToI(i.OutputRegister(), i.InputFloat64Register(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmAdd:
       __ add(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
              i.OutputSBit());
       break;
     case kArmAnd:
       __ and_(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
               i.OutputSBit());
       break;
@@ skipping 73 matching lines @@
       break;
     case kArmTst:
       __ tst(i.InputRegister(0), i.InputOperand2(1));
       DCHECK_EQ(SetCC, i.OutputSBit());
       break;
     case kArmTeq:
       __ teq(i.InputRegister(0), i.InputOperand2(1));
       DCHECK_EQ(SetCC, i.OutputSBit());
       break;
     case kArmVcmpF64:
-      __ VFPCompareAndSetFlags(i.InputDoubleRegister(0),
-                               i.InputDoubleRegister(1));
+      __ VFPCompareAndSetFlags(i.InputFloat64Register(0),
+                               i.InputFloat64Register(1));
       DCHECK_EQ(SetCC, i.OutputSBit());
       break;
     case kArmVaddF64:
-      __ vadd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-              i.InputDoubleRegister(1));
+      __ vadd(i.OutputFloat64Register(), i.InputFloat64Register(0),
+              i.InputFloat64Register(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVsubF64:
-      __ vsub(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-              i.InputDoubleRegister(1));
+      __ vsub(i.OutputFloat64Register(), i.InputFloat64Register(0),
+              i.InputFloat64Register(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmulF64:
-      __ vmul(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-              i.InputDoubleRegister(1));
+      __ vmul(i.OutputFloat64Register(), i.InputFloat64Register(0),
+              i.InputFloat64Register(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmlaF64:
-      __ vmla(i.OutputDoubleRegister(), i.InputDoubleRegister(1),
-              i.InputDoubleRegister(2));
+      __ vmla(i.OutputFloat64Register(), i.InputFloat64Register(1),
+              i.InputFloat64Register(2));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmlsF64:
-      __ vmls(i.OutputDoubleRegister(), i.InputDoubleRegister(1),
-              i.InputDoubleRegister(2));
+      __ vmls(i.OutputFloat64Register(), i.InputFloat64Register(1),
+              i.InputFloat64Register(2));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVdivF64:
-      __ vdiv(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
-              i.InputDoubleRegister(1));
+      __ vdiv(i.OutputFloat64Register(), i.InputFloat64Register(0),
+              i.InputFloat64Register(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmodF64: {
       // TODO(bmeurer): We should really get rid of this special instruction,
       // and generate a CallAddress instruction instead.
       FrameScope scope(masm(), StackFrame::MANUAL);
       __ PrepareCallCFunction(0, 2, kScratchReg);
-      __ MovToFloatParameters(i.InputDoubleRegister(0),
-                              i.InputDoubleRegister(1));
+      __ MovToFloatParameters(i.InputFloat64Register(0),
+                              i.InputFloat64Register(1));
       __ CallCFunction(ExternalReference::mod_two_doubles_operation(isolate()),
                        0, 2);
       // Move the result in the double result register.
-      __ MovFromFloatResult(i.OutputDoubleRegister());
+      __ MovFromFloatResult(i.OutputFloat64Register());
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
+    case kArmVsqrtF64:
+      __ vsqrt(i.OutputFloat64Register(), i.InputFloat64Register(0));
+      break;
     case kArmVnegF64:
-      __ vneg(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
+      __ vneg(i.OutputFloat64Register(), i.InputFloat64Register(0));
       break;
-    case kArmVsqrtF64:
-      __ vsqrt(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
+    case kArmVcvtF32F64: {
+      __ vcvt_f32_f64(i.OutputFloat32Register(), i.InputFloat64Register(0));
+      DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
+    }
+    case kArmVcvtF64F32: {
+      __ vcvt_f64_f32(i.OutputFloat64Register(), i.InputFloat32Register(0));
+      DCHECK_EQ(LeaveCC, i.OutputSBit());
+      break;
+    }
     case kArmVcvtF64S32: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
       __ vmov(scratch, i.InputRegister(0));
-      __ vcvt_f64_s32(i.OutputDoubleRegister(), scratch);
+      __ vcvt_f64_s32(i.OutputFloat64Register(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtF64U32: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
       __ vmov(scratch, i.InputRegister(0));
-      __ vcvt_f64_u32(i.OutputDoubleRegister(), scratch);
+      __ vcvt_f64_u32(i.OutputFloat64Register(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtS32F64: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
-      __ vcvt_s32_f64(scratch, i.InputDoubleRegister(0));
+      __ vcvt_s32_f64(scratch, i.InputFloat64Register(0));
       __ vmov(i.OutputRegister(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtU32F64: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
-      __ vcvt_u32_f64(scratch, i.InputDoubleRegister(0));
+      __ vcvt_u32_f64(scratch, i.InputFloat64Register(0));
       __ vmov(i.OutputRegister(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmLdrb:
       __ ldrb(i.OutputRegister(), i.InputOffset());
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmLdrsb:
       __ ldrsb(i.OutputRegister(), i.InputOffset());
@@ skipping 22 matching lines @@
     case kArmLdr:
       __ ldr(i.OutputRegister(), i.InputOffset());
       break;
     case kArmStr: {
       int index = 0;
       MemOperand operand = i.InputOffset(&index);
       __ str(i.InputRegister(index), operand);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
-    case kArmVldr32: {
-      SwVfpRegister scratch = kScratchDoubleReg.low();
-      __ vldr(scratch, i.InputOffset());
-      __ vcvt_f64_f32(i.OutputDoubleRegister(), scratch);
+    case kArmVldrF32: {
+      __ vldr(i.OutputFloat32Register(), i.InputOffset());
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
-    case kArmVstr32: {
+    case kArmVstrF32: {
       int index = 0;
-      SwVfpRegister scratch = kScratchDoubleReg.low();
       MemOperand operand = i.InputOffset(&index);
-      __ vcvt_f32_f64(scratch, i.InputDoubleRegister(index));
-      __ vstr(scratch, operand);
+      __ vstr(i.InputFloat32Register(index), operand);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
-    case kArmVldr64:
-      __ vldr(i.OutputDoubleRegister(), i.InputOffset());
+    case kArmVldrF64:
+      __ vldr(i.OutputFloat64Register(), i.InputOffset());
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
-    case kArmVstr64: {
+    case kArmVstrF64: {
       int index = 0;
       MemOperand operand = i.InputOffset(&index);
-      __ vstr(i.InputDoubleRegister(index), operand);
+      __ vstr(i.InputFloat64Register(index), operand);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmPush:
       __ Push(i.InputRegister(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmStoreWriteBarrier: {
       Register object = i.InputRegister(0);
       Register index = i.InputRegister(1);
@@ skipping 438 matching lines @@
     }
   }
   MarkLazyDeoptSite();
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8