[ARM] Improve VFP register moves.

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/assembler-inl.h"
 #include "src/compilation-info.h"
 #include "src/compiler/code-generator-impl.h"
@@ skipping 1594 matching lines @@
       }
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmPoke: {
       int const slot = MiscField::decode(instr->opcode());
       __ str(i.InputRegister(0), MemOperand(sp, slot * kPointerSize));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmF32x4Splat: {
-      __ vdup(i.OutputSimd128Register(), i.InputFloatRegister(0));
+      SwVfpRegister src = i.InputFloatRegister(0);
+      DwVfpRegister src_d = DwVfpRegister::from_code(src.code() / 2);
+      int src_index = src.code() & 0x1;
+      __ vdup(Neon32, i.OutputSimd128Register(), src_d, src_index);
       break;
     }
     case kArmF32x4ExtractLane: {
       __ ExtractLane(i.OutputFloatRegister(), i.InputSimd128Register(0),
-                     kScratchReg, i.InputInt8(1));
+                     i.InputInt8(1));
       break;
     }
     case kArmF32x4ReplaceLane: {
       __ ReplaceLane(i.OutputSimd128Register(), i.InputSimd128Register(0),
-                     i.InputFloatRegister(2), kScratchReg, i.InputInt8(1));
+                     i.InputFloatRegister(2), i.InputInt8(1));
       break;
     }
     case kArmF32x4SConvertI32x4: {
       __ vcvt_f32_s32(i.OutputSimd128Register(), i.InputSimd128Register(0));
       break;
     }
     case kArmF32x4UConvertI32x4: {
       __ vcvt_f32_u32(i.OutputSimd128Register(), i.InputSimd128Register(0));
       break;
     }
@@ skipping 576 matching lines @@
       int src0_code = src0.code() * 4;
       int src1_code = src1.code() * 4;
       int32_t shuffle = i.InputInt32(2);
       for (int i = 0; i < 4; i++) {
         int lane = shuffle & 0x7;
         int src_code = src0_code;
         if (lane >= 4) {
           src_code = src1_code;
           lane &= 0x3;
         }
-        __ VmovExtended(dst_code + i, src_code + lane, kScratchReg);
+        __ VmovExtended(dst_code + i, src_code + lane);
         shuffle >>= 8;
       }
       break;
     }
     case kArmS32x4TransposeRight: {
       Simd128Register dst = i.OutputSimd128Register(),
                       src1 = i.InputSimd128Register(1);
       DCHECK(dst.is(i.InputSimd128Register(0)));
       // src0 = [4, 5, 6, 7], src1 = [0, 1, 2, 3] (flipped from TransposeLeft).
       __ vmov(kScratchQuadReg, src1);
@@ skipping 798 matching lines @@
       } else {
         DCHECK(destination->IsDoubleStackSlot());
         __ vstr(src, g.ToMemOperand(destination));
       }
     } else if (rep == MachineRepresentation::kFloat32) {
       // 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);
+        __ VmovExtended(dst_code, src_code);
       } else {
         DCHECK(destination->IsFloatStackSlot());
-        __ VmovExtended(g.ToMemOperand(destination), src_code, kScratchReg);
+        __ VmovExtended(g.ToMemOperand(destination), src_code);
       }
     } else {
       DCHECK_EQ(MachineRepresentation::kSimd128, rep);
       QwNeonRegister src = g.ToSimd128Register(source);
       if (destination->IsSimd128Register()) {
         QwNeonRegister dst = g.ToSimd128Register(destination);
         __ Move(dst, src);
       } else {
         DCHECK(destination->IsSimd128StackSlot());
         MemOperand dst = g.ToMemOperand(destination);
         __ add(kScratchReg, dst.rn(), Operand(dst.offset()));
         __ vst1(Neon8, NeonListOperand(src.low(), 2),
                 NeonMemOperand(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 if (rep == MachineRepresentation::kFloat32) {
         // 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);
+        __ VmovExtended(dst_code, src);
       } else {
         DCHECK_EQ(MachineRepresentation::kSimd128, rep);
         QwNeonRegister dst = g.ToSimd128Register(destination);
         __ add(kScratchReg, src.rn(), Operand(src.offset()));
         __ vld1(Neon8, NeonListOperand(dst.low(), 2),
                 NeonMemOperand(kScratchReg));
       }
     } else if (rep == MachineRepresentation::kFloat64) {
       DCHECK(destination->IsFPStackSlot());
       if (rep == MachineRepresentation::kFloat64) {
@@ skipping 63 matching lines @@
         DCHECK(destination->IsFPStackSlot());
         MemOperand dst = g.ToMemOperand(destination);
         __ Move(temp, src);
         __ vldr(src, dst);
         __ vstr(temp, dst);
       }
     } else if (rep == MachineRepresentation::kFloat32) {
       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);
+        __ VmovExtended(temp.low().code(), src_code);
+        __ VmovExtended(src_code, dst_code);
+        __ VmovExtended(dst_code, temp.low().code());
       } else {
         DCHECK(destination->IsFPStackSlot());
         MemOperand dst = g.ToMemOperand(destination);
-        __ VmovExtended(temp.low().code(), src_code, kScratchReg);
-        __ VmovExtended(src_code, dst, kScratchReg);
+        __ VmovExtended(temp.low().code(), src_code);
+        __ VmovExtended(src_code, dst);
         __ vstr(temp.low(), dst);
       }
     } else {
       DCHECK_EQ(MachineRepresentation::kSimd128, rep);
       QwNeonRegister src = g.ToSimd128Register(source);
       if (destination->IsFPRegister()) {
         QwNeonRegister dst = g.ToSimd128Register(destination);
         __ Swap(src, dst);
       } else {
         DCHECK(destination->IsFPStackSlot());
@@ skipping 71 matching lines @@
       padding_size -= v8::internal::Assembler::kInstrSize;
     }
   }
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8