[Turbofan] Add concept of FP register aliasing on ARM 32.

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/compilation-info.h"
 #include "src/compiler/code-generator-impl.h"
 #include "src/compiler/gap-resolver.h"
@@ skipping 118 matching lines @@
   MemOperand ToMemOperand(InstructionOperand* op) const {
     DCHECK_NOT_NULL(op);
     DCHECK(op->IsStackSlot() || op->IsFPStackSlot());
     return SlotToMemOperand(AllocatedOperand::cast(op)->index());
   }
 
   MemOperand SlotToMemOperand(int slot) const {
     FrameOffset offset = frame_access_state()->GetFrameOffset(slot);
     return MemOperand(offset.from_stack_pointer() ? sp : fp, offset.offset());
   }
-
-  FloatRegister InputFloat32Register(size_t index) {
-    return ToFloat32Register(instr_->InputAt(index));
-  }
-
-  FloatRegister OutputFloat32Register() {
-    return ToFloat32Register(instr_->Output());
-  }
-
-  FloatRegister ToFloat32Register(InstructionOperand* op) {
-    return LowDwVfpRegister::from_code(ToDoubleRegister(op).code()).low();
-  }
 };
 
 namespace {
 
-class OutOfLineLoadFloat32 final : public OutOfLineCode {
+class OutOfLineLoadFloat final : public OutOfLineCode {
  public:
-  OutOfLineLoadFloat32(CodeGenerator* gen, SwVfpRegister result)
+  OutOfLineLoadFloat(CodeGenerator* gen, SwVfpRegister result)
       : OutOfLineCode(gen), result_(result) {}
 
   void Generate() final {
     // Compute sqrtf(-1.0f), which results in a quiet single-precision NaN.
     __ vmov(result_, -1.0f);
     __ vsqrt(result_, result_);
   }
 
  private:
   SwVfpRegister const result_;
@@ skipping 938 matching lines @@
       if (instr->InputAt(2)->IsImmediate()) {
         __ AsrPair(i.OutputRegister(0), i.OutputRegister(1), i.InputRegister(0),
                    i.InputRegister(1), i.InputInt32(2));
       } else {
         __ AsrPair(i.OutputRegister(0), i.OutputRegister(1), i.InputRegister(0),
                    i.InputRegister(1), kScratchReg, i.InputRegister(2));
       }
       break;
     case kArmVcmpF32:
       if (instr->InputAt(1)->IsFPRegister()) {
-        __ VFPCompareAndSetFlags(i.InputFloat32Register(0),
-                                 i.InputFloat32Register(1));
+        __ VFPCompareAndSetFlags(i.InputFloatRegister(0),
+                                 i.InputFloatRegister(1));
       } else {
         DCHECK(instr->InputAt(1)->IsImmediate());
         // 0.0 is the only immediate supported by vcmp instructions.
         DCHECK(i.InputFloat32(1) == 0.0f);
-        __ VFPCompareAndSetFlags(i.InputFloat32Register(0), i.InputFloat32(1));
+        __ VFPCompareAndSetFlags(i.InputFloatRegister(0), i.InputFloat32(1));
       }
       DCHECK_EQ(SetCC, i.OutputSBit());
       break;
     case kArmVaddF32:
-      __ vadd(i.OutputFloat32Register(), i.InputFloat32Register(0),
-              i.InputFloat32Register(1));
+      __ vadd(i.OutputFloatRegister(), i.InputFloatRegister(0),
+              i.InputFloatRegister(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVsubF32:
-      __ vsub(i.OutputFloat32Register(), i.InputFloat32Register(0),
-              i.InputFloat32Register(1));
+      __ vsub(i.OutputFloatRegister(), i.InputFloatRegister(0),
+              i.InputFloatRegister(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmulF32:
-      __ vmul(i.OutputFloat32Register(), i.InputFloat32Register(0),
-              i.InputFloat32Register(1));
+      __ vmul(i.OutputFloatRegister(), i.InputFloatRegister(0),
+              i.InputFloatRegister(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmlaF32:
-      __ vmla(i.OutputFloat32Register(), i.InputFloat32Register(1),
-              i.InputFloat32Register(2));
+      __ vmla(i.OutputFloatRegister(), i.InputFloatRegister(1),
+              i.InputFloatRegister(2));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmlsF32:
-      __ vmls(i.OutputFloat32Register(), i.InputFloat32Register(1),
-              i.InputFloat32Register(2));
+      __ vmls(i.OutputFloatRegister(), i.InputFloatRegister(1),
+              i.InputFloatRegister(2));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVdivF32:
-      __ vdiv(i.OutputFloat32Register(), i.InputFloat32Register(0),
-              i.InputFloat32Register(1));
+      __ vdiv(i.OutputFloatRegister(), i.InputFloatRegister(0),
+              i.InputFloatRegister(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVsqrtF32:
-      __ vsqrt(i.OutputFloat32Register(), i.InputFloat32Register(0));
+      __ vsqrt(i.OutputFloatRegister(), i.InputFloatRegister(0));
       break;
     case kArmVabsF32:
-      __ vabs(i.OutputFloat32Register(), i.InputFloat32Register(0));
+      __ vabs(i.OutputFloatRegister(), i.InputFloatRegister(0));
       break;
     case kArmVnegF32:
-      __ vneg(i.OutputFloat32Register(), i.InputFloat32Register(0));
+      __ vneg(i.OutputFloatRegister(), i.InputFloatRegister(0));
       break;
     case kArmVcmpF64:
       if (instr->InputAt(1)->IsFPRegister()) {
         __ VFPCompareAndSetFlags(i.InputDoubleRegister(0),
                                  i.InputDoubleRegister(1));
       } else {
         DCHECK(instr->InputAt(1)->IsImmediate());
         // 0.0 is the only immediate supported by vcmp instructions.
         DCHECK(i.InputDouble(1) == 0.0);
         __ VFPCompareAndSetFlags(i.InputDoubleRegister(0), i.InputDouble(1));
@@ skipping 48 matching lines @@
       __ vsqrt(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       break;
     case kArmVabsF64:
       __ vabs(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       break;
     case kArmVnegF64:
       __ vneg(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       break;
     case kArmVrintmF32: {
       CpuFeatureScope scope(masm(), ARMv8);
-      __ vrintm(i.OutputFloat32Register(), i.InputFloat32Register(0));
+      __ vrintm(i.OutputFloatRegister(), i.InputFloatRegister(0));
       break;
     }
     case kArmVrintmF64: {
       CpuFeatureScope scope(masm(), ARMv8);
       __ vrintm(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       break;
     }
     case kArmVrintpF32: {
       CpuFeatureScope scope(masm(), ARMv8);
-      __ vrintp(i.OutputFloat32Register(), i.InputFloat32Register(0));
+      __ vrintp(i.OutputFloatRegister(), i.InputFloatRegister(0));
       break;
     }
     case kArmVrintpF64: {
       CpuFeatureScope scope(masm(), ARMv8);
       __ vrintp(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       break;
     }
     case kArmVrintzF32: {
       CpuFeatureScope scope(masm(), ARMv8);
-      __ vrintz(i.OutputFloat32Register(), i.InputFloat32Register(0));
+      __ vrintz(i.OutputFloatRegister(), i.InputFloatRegister(0));
       break;
     }
     case kArmVrintzF64: {
       CpuFeatureScope scope(masm(), ARMv8);
       __ vrintz(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       break;
     }
     case kArmVrintaF64: {
       CpuFeatureScope scope(masm(), ARMv8);
       __ vrinta(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       break;
     }
     case kArmVrintnF32: {
       CpuFeatureScope scope(masm(), ARMv8);
-      __ vrintn(i.OutputFloat32Register(), i.InputFloat32Register(0));
+      __ vrintn(i.OutputFloatRegister(), i.InputFloatRegister(0));
       break;
     }
     case kArmVrintnF64: {
       CpuFeatureScope scope(masm(), ARMv8);
       __ vrintn(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
       break;
     }
     case kArmVcvtF32F64: {
-      __ vcvt_f32_f64(i.OutputFloat32Register(), i.InputDoubleRegister(0));
+      __ vcvt_f32_f64(i.OutputFloatRegister(), i.InputDoubleRegister(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtF64F32: {
-      __ vcvt_f64_f32(i.OutputDoubleRegister(), i.InputFloat32Register(0));
+      __ vcvt_f64_f32(i.OutputDoubleRegister(), i.InputFloatRegister(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtF32S32: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
       __ vmov(scratch, i.InputRegister(0));
-      __ vcvt_f32_s32(i.OutputFloat32Register(), scratch);
+      __ vcvt_f32_s32(i.OutputFloatRegister(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtF32U32: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
       __ vmov(scratch, i.InputRegister(0));
-      __ vcvt_f32_u32(i.OutputFloat32Register(), scratch);
+      __ vcvt_f32_u32(i.OutputFloatRegister(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtF64S32: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
       __ vmov(scratch, i.InputRegister(0));
       __ vcvt_f64_s32(i.OutputDoubleRegister(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtF64U32: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
       __ vmov(scratch, i.InputRegister(0));
       __ vcvt_f64_u32(i.OutputDoubleRegister(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtS32F32: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
-      __ vcvt_s32_f32(scratch, i.InputFloat32Register(0));
+      __ vcvt_s32_f32(scratch, i.InputFloatRegister(0));
       __ vmov(i.OutputRegister(), scratch);
       // Avoid INT32_MAX as an overflow indicator and use INT32_MIN instead,
       // because INT32_MIN allows easier out-of-bounds detection.
       __ cmn(i.OutputRegister(), Operand(1));
       __ mov(i.OutputRegister(), Operand(INT32_MIN), SBit::LeaveCC, vs);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtU32F32: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
-      __ vcvt_u32_f32(scratch, i.InputFloat32Register(0));
+      __ vcvt_u32_f32(scratch, i.InputFloatRegister(0));
       __ vmov(i.OutputRegister(), scratch);
       // Avoid UINT32_MAX as an overflow indicator and use 0 instead,
       // because 0 allows easier out-of-bounds detection.
       __ cmn(i.OutputRegister(), Operand(1));
       __ adc(i.OutputRegister(), i.OutputRegister(), Operand::Zero());
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtS32F64: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
       __ vcvt_s32_f64(scratch, i.InputDoubleRegister(0));
       __ vmov(i.OutputRegister(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVcvtU32F64: {
       SwVfpRegister scratch = kScratchDoubleReg.low();
       __ vcvt_u32_f64(scratch, i.InputDoubleRegister(0));
       __ vmov(i.OutputRegister(), scratch);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVmovU32F32:
-      __ vmov(i.OutputRegister(), i.InputFloat32Register(0));
+      __ vmov(i.OutputRegister(), i.InputFloatRegister(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmovF32U32:
-      __ vmov(i.OutputFloat32Register(), i.InputRegister(0));
+      __ vmov(i.OutputFloatRegister(), i.InputRegister(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmovLowU32F64:
       __ VmovLow(i.OutputRegister(), i.InputDoubleRegister(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVmovLowF64U32:
       __ VmovLow(i.OutputDoubleRegister(), i.InputRegister(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
@@ skipping 37 matching lines @@
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmLdr:
       __ ldr(i.OutputRegister(), i.InputOffset());
       break;
     case kArmStr:
       __ str(i.InputRegister(0), i.InputOffset(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVldrF32: {
-      __ vldr(i.OutputFloat32Register(), i.InputOffset());
+      __ vldr(i.OutputFloatRegister(), i.InputOffset());
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmVstrF32:
-      __ vstr(i.InputFloat32Register(0), i.InputOffset(1));
+      __ vstr(i.InputFloatRegister(0), i.InputOffset(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVldrF64:
       __ vldr(i.OutputDoubleRegister(), i.InputOffset());
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVstrF64:
       __ vstr(i.InputDoubleRegister(0), i.InputOffset(1));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmFloat32Max: {
-      SwVfpRegister result = i.OutputFloat32Register();
-      SwVfpRegister left = i.InputFloat32Register(0);
-      SwVfpRegister right = i.InputFloat32Register(1);
+      SwVfpRegister result = i.OutputFloatRegister();
+      SwVfpRegister left = i.InputFloatRegister(0);
+      SwVfpRegister right = i.InputFloatRegister(1);
       if (left.is(right)) {
         __ Move(result, left);
       } else {
         auto ool = new (zone()) OutOfLineFloat32Max(this, result, left, right);
         __ FloatMax(result, left, right, ool->entry());
         __ bind(ool->exit());
       }
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmFloat64Max: {
       DwVfpRegister result = i.OutputDoubleRegister();
       DwVfpRegister left = i.InputDoubleRegister(0);
       DwVfpRegister right = i.InputDoubleRegister(1);
       if (left.is(right)) {
         __ Move(result, left);
       } else {
         auto ool = new (zone()) OutOfLineFloat64Max(this, result, left, right);
         __ FloatMax(result, left, right, ool->entry());
         __ bind(ool->exit());
       }
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmFloat32Min: {
-      SwVfpRegister result = i.OutputFloat32Register();
-      SwVfpRegister left = i.InputFloat32Register(0);
-      SwVfpRegister right = i.InputFloat32Register(1);
+      SwVfpRegister result = i.OutputFloatRegister();
+      SwVfpRegister left = i.InputFloatRegister(0);
+      SwVfpRegister right = i.InputFloatRegister(1);
       if (left.is(right)) {
         __ Move(result, left);
       } else {
         auto ool = new (zone()) OutOfLineFloat32Min(this, result, left, right);
         __ FloatMin(result, left, right, ool->entry());
         __ bind(ool->exit());
       }
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
@@ skipping 18 matching lines @@
       break;
     }
     case kArmPush:
       if (instr->InputAt(0)->IsFPRegister()) {
         LocationOperand* op = LocationOperand::cast(instr->InputAt(0));
         if (op->representation() == MachineRepresentation::kFloat64) {
           __ vpush(i.InputDoubleRegister(0));
           frame_access_state()->IncreaseSPDelta(kDoubleSize / kPointerSize);
         } else {
           DCHECK_EQ(MachineRepresentation::kFloat32, op->representation());
-          __ vpush(i.InputFloat32Register(0));
+          __ vpush(i.InputFloatRegister(0));
           frame_access_state()->IncreaseSPDelta(1);
         }
       } else {
         __ push(i.InputRegister(0));
         frame_access_state()->IncreaseSPDelta(1);
       }
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmPoke: {
       int const slot = MiscField::decode(instr->opcode());
@@ skipping 10 matching lines @@
     case kCheckedLoadInt16:
       ASSEMBLE_CHECKED_LOAD_INTEGER(ldrsh);
       break;
     case kCheckedLoadUint16:
       ASSEMBLE_CHECKED_LOAD_INTEGER(ldrh);
       break;
     case kCheckedLoadWord32:
       ASSEMBLE_CHECKED_LOAD_INTEGER(ldr);
       break;
     case kCheckedLoadFloat32:
-      ASSEMBLE_CHECKED_LOAD_FP(Float32);
+      ASSEMBLE_CHECKED_LOAD_FP(Float);
       break;
     case kCheckedLoadFloat64:
       ASSEMBLE_CHECKED_LOAD_FP(Double);
       break;
     case kCheckedStoreWord8:
       ASSEMBLE_CHECKED_STORE_INTEGER(strb);
       break;
     case kCheckedStoreWord16:
       ASSEMBLE_CHECKED_STORE_INTEGER(strh);
       break;
     case kCheckedStoreWord32:
       ASSEMBLE_CHECKED_STORE_INTEGER(str);
       break;
     case kCheckedStoreFloat32:
-      ASSEMBLE_CHECKED_STORE_FP(Float32);
+      ASSEMBLE_CHECKED_STORE_FP(Float);
       break;
     case kCheckedStoreFloat64:
       ASSEMBLE_CHECKED_STORE_FP(Double);
       break;
     case kCheckedLoadWord64:
     case kCheckedStoreWord64:
       UNREACHABLE();  // currently unsupported checked int64 load/store.
       break;
 
     case kAtomicLoadInt8:
@@ skipping 221 matching lines @@
       __ b(&return_label_);
       return;
     } else {
       __ bind(&return_label_);
       AssembleDeconstructFrame();
     }
   }
   __ Ret(pop_count);
 }
 
-
 void CodeGenerator::AssembleMove(InstructionOperand* source,
                                  InstructionOperand* destination) {
   ArmOperandConverter g(this, nullptr);
   // Dispatch on the source and destination operand kinds.  Not all
   // combinations are possible.
   if (source->IsRegister()) {
     DCHECK(destination->IsRegister() || destination->IsStackSlot());
     Register src = g.ToRegister(source);
     if (destination->IsRegister()) {
       __ mov(g.ToRegister(destination), src);
@@ skipping 48 matching lines @@
             __ Move(dst, src_object);
           }
           break;
         }
         case Constant::kRpoNumber:
           UNREACHABLE();  // TODO(dcarney): loading RPO constants on arm.
           break;
       }
       if (destination->IsStackSlot()) __ str(dst, g.ToMemOperand(destination));
     } else if (src.type() == Constant::kFloat32) {
-      if (destination->IsFPStackSlot()) {
+      if (destination->IsFloatStackSlot()) {
         MemOperand dst = g.ToMemOperand(destination);
         __ mov(ip, Operand(bit_cast<int32_t>(src.ToFloat32())));
         __ str(ip, dst);
       } else {
-        SwVfpRegister dst = g.ToFloat32Register(destination);
+        SwVfpRegister dst = g.ToFloatRegister(destination);
         __ vmov(dst, src.ToFloat32());
       }
     } else {
       DCHECK_EQ(Constant::kFloat64, src.type());
       DwVfpRegister dst = destination->IsFPRegister()
                               ? g.ToDoubleRegister(destination)
                               : kScratchDoubleReg;
       __ vmov(dst, src.ToFloat64(), kScratchReg);
-      if (destination->IsFPStackSlot()) {
+      if (destination->IsDoubleStackSlot()) {
         __ vstr(dst, g.ToMemOperand(destination));
       }
     }
   } else if (source->IsFPRegister()) {
-    DwVfpRegister src = g.ToDoubleRegister(source);
-    if (destination->IsFPRegister()) {
-      DwVfpRegister dst = g.ToDoubleRegister(destination);
-      __ Move(dst, src);
+    MachineRepresentation rep = LocationOperand::cast(source)->representation();
+    if (rep == MachineRepresentation::kFloat64) {
+      DwVfpRegister src = g.ToDoubleRegister(source);
+      if (destination->IsDoubleRegister()) {
+        DwVfpRegister dst = g.ToDoubleRegister(destination);
+        __ Move(dst, src);
+      } else {
+        DCHECK(destination->IsDoubleStackSlot());
+        __ vstr(src, g.ToMemOperand(destination));
+      }
     } else {
-      DCHECK(destination->IsFPStackSlot());
-      __ vstr(src, g.ToMemOperand(destination));
+      DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+      // GapResolver may give us reg codes that don't map to actual s-registers.
+      // Generate code to work around those cases.
+      int src_code = LocationOperand::cast(source)->register_code();
+      if (destination->IsFloatRegister()) {
+        int dst_code = LocationOperand::cast(destination)->register_code();
+        __ VmovExtended(dst_code, src_code, kScratchReg);
+      } else {
+        DCHECK(destination->IsFloatStackSlot());
+        __ VmovExtended(g.ToMemOperand(destination), src_code, kScratchReg);
+      }
     }
   } else if (source->IsFPStackSlot()) {
     MemOperand src = g.ToMemOperand(source);
+    MachineRepresentation rep =
+        LocationOperand::cast(destination)->representation();
     if (destination->IsFPRegister()) {
+      if (rep == MachineRepresentation::kFloat64) {
         __ vldr(g.ToDoubleRegister(destination), src);
+      } else {
+        DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+        // GapResolver may give us reg codes that don't map to actual
+        // s-registers. Generate code to work around those cases.
+        int dst_code = LocationOperand::cast(destination)->register_code();
+        __ VmovExtended(dst_code, src, kScratchReg);
+      }
     } else {
       DCHECK(destination->IsFPStackSlot());
+      if (rep == MachineRepresentation::kFloat64) {
         DwVfpRegister temp = kScratchDoubleReg;
         __ vldr(temp, src);
         __ vstr(temp, g.ToMemOperand(destination));
+      } else {
+        DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+        SwVfpRegister temp = kScratchDoubleReg.low();
+        __ vldr(temp, src);
+        __ vstr(temp, g.ToMemOperand(destination));
+      }
     }
   } else {
     UNREACHABLE();
   }
 }
 
 
 void CodeGenerator::AssembleSwap(InstructionOperand* source,
                                  InstructionOperand* destination) {
   ArmOperandConverter g(this, nullptr);
@@ skipping 19 matching lines @@
     DCHECK(destination->IsStackSlot());
     Register temp_0 = kScratchReg;
     SwVfpRegister temp_1 = kScratchDoubleReg.low();
     MemOperand src = g.ToMemOperand(source);
     MemOperand dst = g.ToMemOperand(destination);
     __ ldr(temp_0, src);
     __ vldr(temp_1, dst);
     __ str(temp_0, dst);
     __ vstr(temp_1, src);
   } else if (source->IsFPRegister()) {
+    MachineRepresentation rep = LocationOperand::cast(source)->representation();
     LowDwVfpRegister temp = kScratchDoubleReg;
-    DwVfpRegister src = g.ToDoubleRegister(source);
-    if (destination->IsFPRegister()) {
-      DwVfpRegister dst = g.ToDoubleRegister(destination);
-      __ vswp(src, dst);
+    if (rep == MachineRepresentation::kFloat64) {
+      DwVfpRegister src = g.ToDoubleRegister(source);
+      if (destination->IsFPRegister()) {
+        DwVfpRegister dst = g.ToDoubleRegister(destination);
+        __ vswp(src, dst);
+      } else {
+        DCHECK(destination->IsFPStackSlot());
+        MemOperand dst = g.ToMemOperand(destination);
+        __ Move(temp, src);
+        __ vldr(src, dst);
+        __ vstr(temp, dst);
+      }
     } else {
-      DCHECK(destination->IsFPStackSlot());
-      MemOperand dst = g.ToMemOperand(destination);
-      __ Move(temp, src);
-      __ vldr(src, dst);
-      __ vstr(temp, dst);
+      DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+      int src_code = LocationOperand::cast(source)->register_code();
+      if (destination->IsFPRegister()) {
+        int dst_code = LocationOperand::cast(destination)->register_code();
+        __ VmovExtended(temp.low().code(), src_code, kScratchReg);
+        __ VmovExtended(src_code, dst_code, kScratchReg);
+        __ VmovExtended(dst_code, temp.low().code(), kScratchReg);
+      } else {
+        DCHECK(destination->IsFPStackSlot());
+        MemOperand dst = g.ToMemOperand(destination);
+        __ VmovExtended(temp.low().code(), src_code, kScratchReg);
+        __ VmovExtended(src_code, dst, kScratchReg);
+        __ vstr(temp.low(), dst);
+      }
     }
   } else if (source->IsFPStackSlot()) {
     DCHECK(destination->IsFPStackSlot());
     Register temp_0 = kScratchReg;
     LowDwVfpRegister temp_1 = kScratchDoubleReg;
     MemOperand src0 = g.ToMemOperand(source);
     MemOperand dst0 = g.ToMemOperand(destination);
-    MemOperand src1(src0.rn(), src0.offset() + kPointerSize);
-    MemOperand dst1(dst0.rn(), dst0.offset() + kPointerSize);
-    __ vldr(temp_1, dst0);  // Save destination in temp_1.
-    __ ldr(temp_0, src0);   // Then use temp_0 to copy source to destination.
-    __ str(temp_0, dst0);
-    __ ldr(temp_0, src1);
-    __ str(temp_0, dst1);
-    __ vstr(temp_1, src0);
+    MachineRepresentation rep = LocationOperand::cast(source)->representation();
+    if (rep == MachineRepresentation::kFloat64) {
+      MemOperand src1(src0.rn(), src0.offset() + kPointerSize);
+      MemOperand dst1(dst0.rn(), dst0.offset() + kPointerSize);
+      __ vldr(temp_1, dst0);  // Save destination in temp_1.
+      __ ldr(temp_0, src0);   // Then use temp_0 to copy source to destination.
+      __ str(temp_0, dst0);
+      __ ldr(temp_0, src1);
+      __ str(temp_0, dst1);
+      __ vstr(temp_1, src0);
+    } else {
+      DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+      __ vldr(temp_1.low(), dst0);  // Save destination in temp_1.
+      __ ldr(temp_0, src0);  // Then use temp_0 to copy source to destination.
+      __ str(temp_0, dst0);
+      __ vstr(temp_1.low(), src0);
+    }
   } else {
     // No other combinations are possible.
     UNREACHABLE();
   }
 }
 
-
 void CodeGenerator::AssembleJumpTable(Label** targets, size_t target_count) {
   // On 32-bit ARM we emit the jump tables inline.
   UNREACHABLE();
 }
 
 
 void CodeGenerator::EnsureSpaceForLazyDeopt() {
   if (!info()->ShouldEnsureSpaceForLazyDeopt()) {
     return;
   }
@@ skipping 12 matching lines @@
       padding_size -= v8::internal::Assembler::kInstrSize;
     }
   }
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8