ARM: Make use of d16-d31 when available.

src/arm/assembler-arm.cc

 // Copyright (c) 1994-2006 Sun Microsystems Inc.
 // All Rights Reserved.
 //
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions
 // are met:
 //
 // - Redistributions of source code must retain the above copyright notice,
 // this list of conditions and the following disclaimer.
 //
@@ skipping 33 matching lines @@
 namespace v8 {
 namespace internal {
 
 #ifdef DEBUG
 bool CpuFeatures::initialized_ = false;
 #endif
 unsigned CpuFeatures::supported_ = 0;
 unsigned CpuFeatures::found_by_runtime_probing_ = 0;
 
 
+ExternalReference ExternalReference::cpu_features() {
+  ASSERT(CpuFeatures::initialized_);
+  return ExternalReference(&CpuFeatures::supported_);
+}
+
 // Get the CPU features enabled by the build. For cross compilation the
 // preprocessor symbols CAN_USE_ARMV7_INSTRUCTIONS and CAN_USE_VFP3_INSTRUCTIONS
 // can be defined to enable ARMv7 and VFPv3 instructions when building the
 // snapshot.
 static unsigned CpuFeaturesImpliedByCompiler() {
   unsigned answer = 0;
 #ifdef CAN_USE_ARMV7_INSTRUCTIONS
   answer |= 1u << ARMv7;
 #endif  // CAN_USE_ARMV7_INSTRUCTIONS
 #ifdef CAN_USE_VFP3_INSTRUCTIONS
   answer |= 1u << VFP3 | 1u << VFP2 | 1u << ARMv7;
 #endif  // CAN_USE_VFP3_INSTRUCTIONS
 #ifdef CAN_USE_VFP2_INSTRUCTIONS
   answer |= 1u << VFP2;
 #endif  // CAN_USE_VFP2_INSTRUCTIONS
+#ifdef CAN_USE_VFP32DREGS
+  answer |= 1u << VFP32DREGS;
+#endif  // CAN_USE_VFP32DREGS
 
 #ifdef __arm__
   // If the compiler is allowed to use VFP then we can use VFP too in our code
   // generation even when generating snapshots. ARMv7 and hardware floating
   // point support implies VFPv3, see ARM DDI 0406B, page A1-6.
 #if defined(CAN_USE_ARMV7_INSTRUCTIONS) && defined(__VFP_FP__) \
     && !defined(__SOFTFP__)
   answer |= 1u << VFP3 | 1u << ARMv7 | 1u << VFP2;
 #endif  // defined(CAN_USE_ARMV7_INSTRUCTIONS) && defined(__VFP_FP__)
         // && !defined(__SOFTFP__)
 #endif  // _arm__
   if (answer & (1u << ARMv7)) {
     answer |= 1u << UNALIGNED_ACCESSES;
   }
 
   return answer;
 }
 
 
 const char* DwVfpRegister::AllocationIndexToString(int index) {
   if (CpuFeatures::IsSupported(VFP2)) {
-    ASSERT(index >= 0 && index < kMaxNumAllocatableRegisters);
+    ASSERT(index >= 0 && index < NumAllocatableRegisters());
+    ASSERT(kScratchDoubleReg.code() - kDoubleRegZero.code() ==
+           kNumReservedRegisters - 1);
+    if (index >= kDoubleRegZero.code())
+      index += kNumReservedRegisters;
+
+    // TODO(hans): Maybe this could just use VFPRegisters::Name()?
     const char* const names[] = {
       "d0",
       "d1",
       "d2",
       "d3",
       "d4",
       "d5",
       "d6",
       "d7",
       "d8",
       "d9",
       "d10",
       "d11",
       "d12",
       "d13"
+      "d14",
+      "d15",
+      "d16",
+      "d17",
+      "d18",
+      "d19",
+      "d20",
+      "d21",
+      "d22",
+      "d23",
+      "d24",
+      "d25",
+      "d26",
+      "d27",
+      "d28",
+      "d29",
+      "d30",
+      "d31"
     };
     return names[index];
   } else {
     ASSERT(index == 0);
     return "sfpd0";
   }
 }
 
 
 void CpuFeatures::Probe() {
@@ skipping 25 matching lines @@
     supported_ |= 1u << ARMv7;
   }
 
   if (FLAG_enable_sudiv) {
     supported_ |= 1u << SUDIV;
   }
 
   if (FLAG_enable_movw_movt) {
     supported_ |= 1u << MOVW_MOVT_IMMEDIATE_LOADS;
   }
+
+  if (FLAG_enable_32dregs) {
+    supported_ |= 1u << VFP32DREGS;
+  }
+
 #else  // __arm__
   // Probe for additional features not already known to be available.
   if (!IsSupported(VFP3) && OS::ArmCpuHasFeature(VFP3)) {
     // This implementation also sets the VFP flags if runtime
     // detection of VFP returns true. VFPv3 implies ARMv7 and VFP2, see ARM DDI
     // 0406B, page A1-6.
     found_by_runtime_probing_ |= 1u << VFP3 | 1u << ARMv7 | 1u << VFP2;
   } else if (!IsSupported(VFP2) && OS::ArmCpuHasFeature(VFP2)) {
     found_by_runtime_probing_ |= 1u << VFP2;
   }
 
   if (!IsSupported(ARMv7) && OS::ArmCpuHasFeature(ARMv7)) {
     found_by_runtime_probing_ |= 1u << ARMv7;
   }
 
   if (!IsSupported(SUDIV) && OS::ArmCpuHasFeature(SUDIV)) {
     found_by_runtime_probing_ |= 1u << SUDIV;
   }
 
   if (!IsSupported(UNALIGNED_ACCESSES) && OS::ArmCpuHasFeature(ARMv7)) {
     found_by_runtime_probing_ |= 1u << UNALIGNED_ACCESSES;
   }
 
   if (OS::GetCpuImplementer() == QUALCOMM_IMPLEMENTER &&
       OS::ArmCpuHasFeature(ARMv7)) {
     found_by_runtime_probing_ |= 1u << MOVW_MOVT_IMMEDIATE_LOADS;
   }
 
+  if (!IsSupported(VFP32DREGS) && OS::ArmCpuHasFeature(VFP32DREGS)) {
+    found_by_runtime_probing_ |= 1u << VFP32DREGS;
+  }
+
   supported_ |= found_by_runtime_probing_;
 #endif
 
   // Assert that VFP3 implies VFP2 and ARMv7.
   ASSERT(!IsSupported(VFP3) || (IsSupported(VFP2) && IsSupported(ARMv7)));
 }
 
 
 // -----------------------------------------------------------------------------
 // Implementation of RelocInfo
@@ skipping 1550 matching lines @@
 }
 
 
 // Support for VFP.
 
 void Assembler::vldr(const DwVfpRegister dst,
                      const Register base,
                      int offset,
                      const Condition cond) {
   // Ddst = MEM(Rbase + offset).
-  // Instruction details available in ARM DDI 0406A, A8-628.
-  // cond(31-28) | 1101(27-24)| U001(23-20) | Rbase(19-16) |
-  // Vdst(15-12) | 1011(11-8) | offset
+  // Instruction details available in ARM DDI 0406C.b, A8-924.
+  // cond(31-28) | 1101(27-24)| U(23) | D(22) | 01(21-20) | Rbase(19-16) |
+  // Vd(15-12) | 1011(11-8) | offset
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   int u = 1;
   if (offset < 0) {
     offset = -offset;
     u = 0;
   }
+  int vd, d;
+  dst.split_code(&vd, &d);
 
   ASSERT(offset >= 0);
   if ((offset % 4) == 0 && (offset / 4) < 256) {
-    emit(cond | u*B23 | 0xD1*B20 | base.code()*B16 | dst.code()*B12 |
+    emit(cond | 0xD*B24 | u*B23 | d*B22 | B20 | base.code()*B16 | vd*B12 |
          0xB*B8 | ((offset / 4) & 255));
   } else {
     // Larger offsets must be handled by computing the correct address
     // in the ip register.
     ASSERT(!base.is(ip));
     if (u == 1) {
       add(ip, base, Operand(offset));
     } else {
       sub(ip, base, Operand(offset));
     }
-    emit(cond | 0xD1*B20 | ip.code()*B16 | dst.code()*B12 | 0xB*B8);
+    emit(cond | 0xD*B24 | d*B22 | B20 | ip.code()*B16 | vd*B12 | 0xB*B8);
   }
 }
 
 
 void Assembler::vldr(const DwVfpRegister dst,
                      const MemOperand& operand,
                      const Condition cond) {
   ASSERT(!operand.rm().is_valid());
   ASSERT(operand.am_ == Offset);
   vldr(dst, operand.rn(), operand.offset(), cond);
@@ skipping 42 matching lines @@
   ASSERT(operand.am_ == Offset);
   vldr(dst, operand.rn(), operand.offset(), cond);
 }
 
 
 void Assembler::vstr(const DwVfpRegister src,
                      const Register base,
                      int offset,
                      const Condition cond) {
   // MEM(Rbase + offset) = Dsrc.
-  // Instruction details available in ARM DDI 0406A, A8-786.
-  // cond(31-28) | 1101(27-24)| U000(23-20) | | Rbase(19-16) |
-  // Vsrc(15-12) | 1011(11-8) | (offset/4)
+  // Instruction details available in ARM DDI 0406C.b, A8-1082.
+  // cond(31-28) | 1101(27-24)| U(23) | D(22) | 00(21-20) | Rbase(19-16) |
+  // Vd(15-12) | 1011(11-8) | (offset/4)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   int u = 1;
   if (offset < 0) {
     offset = -offset;
     u = 0;
   }
   ASSERT(offset >= 0);
+  int vd, d;
+  src.split_code(&vd, &d);
+
   if ((offset % 4) == 0 && (offset / 4) < 256) {
-    emit(cond | u*B23 | 0xD0*B20 | base.code()*B16 | src.code()*B12 |
-         0xB*B8 | ((offset / 4) & 255));
+    emit(cond | 0xD*B24 | u*B23 | d*B22 | base.code()*B16 | vd*B12 | 0xB*B8 |
+         ((offset / 4) & 255));
   } else {
     // Larger offsets must be handled by computing the correct address
     // in the ip register.
     ASSERT(!base.is(ip));
     if (u == 1) {
       add(ip, base, Operand(offset));
     } else {
       sub(ip, base, Operand(offset));
     }
-    emit(cond | 0xD0*B20 | ip.code()*B16 | src.code()*B12 | 0xB*B8);
+    emit(cond | 0xD*B24 | d*B22 | ip.code()*B16 | vd*B12 | 0xB*B8);
   }
 }
 
 
 void Assembler::vstr(const DwVfpRegister src,
                      const MemOperand& operand,
                      const Condition cond) {
   ASSERT(!operand.rm().is_valid());
   ASSERT(operand.am_ == Offset);
   vstr(src, operand.rn(), operand.offset(), cond);
@@ skipping 41 matching lines @@
   ASSERT(operand.am_ == Offset);
   vstr(src, operand.rn(), operand.offset(), cond);
 }
 
 
 void  Assembler::vldm(BlockAddrMode am,
                       Register base,
                       DwVfpRegister first,
                       DwVfpRegister last,
                       Condition cond) {
-  // Instruction details available in ARM DDI 0406A, A8-626.
+  // Instruction details available in ARM DDI 0406C.b, A8-922.
   // cond(31-28) | 110(27-25)| PUDW1(24-20) | Rbase(19-16) |
-  // first(15-12) | 1010(11-8) | (count * 2)
+  // first(15-12) | 1011(11-8) | (count * 2)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   ASSERT_LE(first.code(), last.code());
   ASSERT(am == ia || am == ia_w || am == db_w);
   ASSERT(!base.is(pc));
 
   int sd, d;
   first.split_code(&sd, &d);
   int count = last.code() - first.code() + 1;
   ASSERT(count <= 16);
   emit(cond | B27 | B26 | am | d*B22 | B20 | base.code()*B16 | sd*B12 |
        0xB*B8 | count*2);
 }
 
 
 void  Assembler::vstm(BlockAddrMode am,
                       Register base,
                       DwVfpRegister first,
                       DwVfpRegister last,
                       Condition cond) {
-  // Instruction details available in ARM DDI 0406A, A8-784.
+  // Instruction details available in ARM DDI 0406C.b, A8-1080.
   // cond(31-28) | 110(27-25)| PUDW0(24-20) | Rbase(19-16) |
   // first(15-12) | 1011(11-8) | (count * 2)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   ASSERT_LE(first.code(), last.code());
   ASSERT(am == ia || am == ia_w || am == db_w);
   ASSERT(!base.is(pc));
 
   int sd, d;
   first.split_code(&sd, &d);
   int count = last.code() - first.code() + 1;
@@ skipping 99 matching lines @@
   *encoding |= (hi >> 12) & 0x80000;  // Top bit of the high nybble.
 
   return true;
 }
 
 
 void Assembler::vmov(const DwVfpRegister dst,
                      double imm,
                      const Register scratch,
                      const Condition cond) {
-  // Dd = immediate
-  // Instruction details available in ARM DDI 0406B, A8-640.
   ASSERT(CpuFeatures::IsEnabled(VFP2));
 
   uint32_t enc;
   if (CpuFeatures::IsSupported(VFP3) && FitsVMOVDoubleImmediate(imm, &enc)) {
     // The double can be encoded in the instruction.
-    emit(cond | 0xE*B24 | 0xB*B20 | dst.code()*B12 | 0xB*B8 | enc);
+    //
+    // Dd = immediate
+    // Instruction details available in ARM DDI 0406C.b, A8-936.
+    // cond(31-28) | 11101(27-23) | D(22) | 11(21-20) | imm4H(19-16) |
+    // Vd(15-12) | 101(11-9) | sz=1(8) | imm4L(3-0)
+    int vd, d;
+    dst.split_code(&vd, &d);
+    emit(cond | 0x1D*B23 | d*B22 | 0x3*B20 | vd*B12 | 0x5*B9 | B8 | enc);
   } else if (FLAG_enable_vldr_imm) {
     // TODO(jfb) Temporarily turned off until we have constant blinding or
     //           some equivalent mitigation: an attacker can otherwise control
     //           generated data which also happens to be executable, a Very Bad
     //           Thing indeed.
     //           Blinding gets tricky because we don't have xor, we probably
     //           need to add/subtract without losing precision, which requires a
     //           cookie value that Lithium is probably better positioned to
     //           choose.
     //           We could also add a few peepholes here like detecting 0.0 and
     //           -0.0 and doing a vmov from the sequestered d14, forcing denorms
     //           to zero (we set flush-to-zero), and normalizing NaN values.
     //           We could also detect redundant values.
     //           The code could also randomize the order of values, though
     //           that's tricky because vldr has a limited reach. Furthermore
     //           it breaks load locality.
     RecordRelocInfo(imm);
     vldr(dst, MemOperand(pc, 0), cond);
   } else {
     // Synthesise the double from ARM immediates.
     uint32_t lo, hi;
     DoubleAsTwoUInt32(imm, &lo, &hi);
     mov(ip, Operand(lo));
 
     if (scratch.is(no_reg)) {
-      // Move the low part of the double into the lower of the corresponsing S
-      // registers of D register dst.
-      vmov(dst.low(), ip, cond);
-
-      // Move the high part of the double into the higher of the corresponsing S
-      // registers of D register dst.
+      // Move the low and high part separately using vmov.32.
+      vmov(dst, 0, ip, cond);
       mov(ip, Operand(hi));
-      vmov(dst.high(), ip, cond);
+      vmov(dst, 1, ip, cond);
     } else {
       // Move the low and high parts of the double to a D register in one
       // instruction.
       mov(scratch, Operand(hi));
       vmov(dst, ip, scratch, cond);
     }
   }
 }
 
 
 void Assembler::vmov(const SwVfpRegister dst,
                      const SwVfpRegister src,
                      const Condition cond) {
   // Sd = Sm
   // Instruction details available in ARM DDI 0406B, A8-642.
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   int sd, d, sm, m;
   dst.split_code(&sd, &d);
   src.split_code(&sm, &m);
   emit(cond | 0xE*B24 | d*B22 | 0xB*B20 | sd*B12 | 0xA*B8 | B6 | m*B5 | sm);
 }
 
 
 void Assembler::vmov(const DwVfpRegister dst,
                      const DwVfpRegister src,
                      const Condition cond) {
   // Dd = Dm
-  // Instruction details available in ARM DDI 0406B, A8-642.
+  // Instruction details available in ARM DDI 0406C.b, A8-938.
+  // cond(31-28) | 11101(27-23) | D(22) | 11(21-20) | 0000(19-16) | Vd(15-12) |
+  // 101(11-9) | sz=1(8) | 0(7) | 1(6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 | 0xB*B20 |
-       dst.code()*B12 | 0x5*B9 | B8 | B6 | src.code());
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vm, m;
+  src.split_code(&vm, &m);
+  emit(cond | 0x1D*B23 | d*B22 | 0x3*B20 | vd*B12 | 0x5*B9 | B8 | B6 | m*B5 |
+       vm);
 }
 
 
+void Assembler::vmov(const DwVfpRegister dst,
+                     int index,
+                     const Register src,
+                     const Condition cond) {
+  // Dd[index] = Rt
+  // Instruction details available in ARM DDI 0406C.b, A8-940.
+  // cond(31-28) | 1110(27-24) | 0(23) | opc1=0index(22-21) | 0(20) |
+  // Vd(19-16) | Rt(15-12) | 1011(11-8) | D(7) | opc2=00(6-5) | 1(4) | 0000(3-0)
+  ASSERT(CpuFeatures::IsEnabled(VFP2));
+  ASSERT(index == 0 || index == 1);
+  int vd, d;
+  dst.split_code(&vd, &d);
+  emit(cond | 0xE*B24 | index*B21 | vd*B16 | src.code()*B12 | 0xB*B8 | d*B7 |
+       B4);
+}
+
+
 void Assembler::vmov(const DwVfpRegister dst,
                      const Register src1,
                      const Register src2,
                      const Condition cond) {
   // Dm = <Rt,Rt2>.
-  // Instruction details available in ARM DDI 0406A, A8-646.
+  // Instruction details available in ARM DDI 0406C.b, A8-948.
   // cond(31-28) | 1100(27-24)| 010(23-21) | op=0(20) | Rt2(19-16) |
   // Rt(15-12) | 1011(11-8) | 00(7-6) | M(5) | 1(4) | Vm
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   ASSERT(!src1.is(pc) && !src2.is(pc));
+  int vm, m;
+  dst.split_code(&vm, &m);
   emit(cond | 0xC*B24 | B22 | src2.code()*B16 |
-       src1.code()*B12 | 0xB*B8 | B4 | dst.code());
+       src1.code()*B12 | 0xB*B8 | m*B5 | B4 | vm);
 }
 
 
 void Assembler::vmov(const Register dst1,
                      const Register dst2,
                      const DwVfpRegister src,
                      const Condition cond) {
   // <Rt,Rt2> = Dm.
-  // Instruction details available in ARM DDI 0406A, A8-646.
+  // Instruction details available in ARM DDI 0406C.b, A8-948.
   // cond(31-28) | 1100(27-24)| 010(23-21) | op=1(20) | Rt2(19-16) |
   // Rt(15-12) | 1011(11-8) | 00(7-6) | M(5) | 1(4) | Vm
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   ASSERT(!dst1.is(pc) && !dst2.is(pc));
+  int vm, m;
+  src.split_code(&vm, &m);
   emit(cond | 0xC*B24 | B22 | B20 | dst2.code()*B16 |
-       dst1.code()*B12 | 0xB*B8 | B4 | src.code());
+       dst1.code()*B12 | 0xB*B8 | m*B5 | B4 | vm);
 }
 
 
 void Assembler::vmov(const SwVfpRegister dst,
                      const Register src,
                      const Condition cond) {
   // Sn = Rt.
   // Instruction details available in ARM DDI 0406A, A8-642.
   // cond(31-28) | 1110(27-24)| 000(23-21) | op=0(20) | Vn(19-16) |
   // Rt(15-12) | 1010(11-8) | N(7)=0 | 00(6-5) | 1(4) | 0000(3-0)
@@ skipping 192 matching lines @@
                              VFPConversionMode mode,
                              const Condition cond) {
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   emit(EncodeVCVT(F32, dst.code(), F64, src.code(), mode, cond));
 }
 
 
 void Assembler::vneg(const DwVfpRegister dst,
                      const DwVfpRegister src,
                      const Condition cond) {
+  // Instruction details available in ARM DDI 0406C.b, A8-968.
+  // cond(31-28) | 11101(27-23) | D(22) | 11(21-20) | 0001(19-16) | Vd(15-12) |
+  // 101(11-9) | sz=1(8) | 0(7) | 1(6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 | 0xB*B20 | B16 | dst.code()*B12 |
-       0x5*B9 | B8 | B6 | src.code());
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vm, m;
+  src.split_code(&vm, &m);
+
+  emit(cond | 0x1D*B23 | d*B22 | 0x3*B20 | B16 | vd*B12 | 0x5*B9 | B8 | B6 |
+       m*B5 | vm);
 }
 
 
 void Assembler::vabs(const DwVfpRegister dst,
                      const DwVfpRegister src,
                      const Condition cond) {
+  // Instruction details available in ARM DDI 0406C.b, A8-524.
+  // cond(31-28) | 11101(27-23) | D(22) | 11(21-20) | 0000(19-16) | Vd(15-12) |
+  // 101(11-9) | sz=1(8) | 1(7) | 1(6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 | 0xB*B20 | dst.code()*B12 |
-       0x5*B9 | B8 | 0x3*B6 | src.code());
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vm, m;
+  src.split_code(&vm, &m);
+  emit(cond | 0x1D*B23 | d*B22 | 0x3*B20 | vd*B12 | 0x5*B9 | B8 | B7 | B6 |
+       m*B5 | vm);
 }
 
 
 void Assembler::vadd(const DwVfpRegister dst,
                      const DwVfpRegister src1,
                      const DwVfpRegister src2,
                      const Condition cond) {
   // Dd = vadd(Dn, Dm) double precision floating point addition.
   // Dd = D:Vd; Dm=M:Vm; Dn=N:Vm.
-  // Instruction details available in ARM DDI 0406A, A8-536.
-  // cond(31-28) | 11100(27-23)| D=?(22) | 11(21-20) | Vn(19-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | N(7)=0 | 0(6) | M=?(5) | 0(4) | Vm(3-0)
+  // Instruction details available in ARM DDI 0406C.b, A8-830.
+  // cond(31-28) | 11100(27-23)| D(22) | 11(21-20) | Vn(19-16) |
+  // Vd(15-12) | 101(11-9) | sz=1(8) | N(7) | 0(6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 | 0x3*B20 | src1.code()*B16 |
-       dst.code()*B12 | 0x5*B9 | B8 | src2.code());
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vn, n;
+  src1.split_code(&vn, &n);
+  int vm, m;
+  src2.split_code(&vm, &m);
+  emit(cond | 0x1C*B23 | d*B22 | 0x3*B20 | vn*B16 | vd*B12 | 0x5*B9 | B8 |
+       n*B7 | m*B5 | vm);
 }
 
 
 void Assembler::vsub(const DwVfpRegister dst,
                      const DwVfpRegister src1,
                      const DwVfpRegister src2,
                      const Condition cond) {
   // Dd = vsub(Dn, Dm) double precision floating point subtraction.
   // Dd = D:Vd; Dm=M:Vm; Dn=N:Vm.
-  // Instruction details available in ARM DDI 0406A, A8-784.
-  // cond(31-28) | 11100(27-23)| D=?(22) | 11(21-20) | Vn(19-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | N(7)=0 | 1(6) | M=?(5) | 0(4) | Vm(3-0)
+  // Instruction details available in ARM DDI 0406C.b, A8-1086.
+  // cond(31-28) | 11100(27-23)| D(22) | 11(21-20) | Vn(19-16) |
+  // Vd(15-12) | 101(11-9) | sz=1(8) | N(7) | 1(6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 | 0x3*B20 | src1.code()*B16 |
-       dst.code()*B12 | 0x5*B9 | B8 | B6 | src2.code());
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vn, n;
+  src1.split_code(&vn, &n);
+  int vm, m;
+  src2.split_code(&vm, &m);
+  emit(cond | 0x1C*B23 | d*B22 | 0x3*B20 | vn*B16 | vd*B12 | 0x5*B9 | B8 |
+       n*B7 | B6 | m*B5 | vm);
 }
 
 
 void Assembler::vmul(const DwVfpRegister dst,
                      const DwVfpRegister src1,
                      const DwVfpRegister src2,
                      const Condition cond) {
   // Dd = vmul(Dn, Dm) double precision floating point multiplication.
   // Dd = D:Vd; Dm=M:Vm; Dn=N:Vm.
-  // Instruction details available in ARM DDI 0406A, A8-784.
-  // cond(31-28) | 11100(27-23)| D=?(22) | 10(21-20) | Vn(19-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | N(7)=0 | 0(6) | M=?(5) | 0(4) | Vm(3-0)
+  // Instruction details available in ARM DDI 0406C.b, A8-960.
+  // cond(31-28) | 11100(27-23)| D(22) | 10(21-20) | Vn(19-16) |
+  // Vd(15-12) | 101(11-9) | sz=1(8) | N(7) | 0(6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 | 0x2*B20 | src1.code()*B16 |
-       dst.code()*B12 | 0x5*B9 | B8 | src2.code());
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vn, n;
+  src1.split_code(&vn, &n);
+  int vm, m;
+  src2.split_code(&vm, &m);
+  emit(cond | 0x1C*B23 | d*B22 | 0x2*B20 | vn*B16 | vd*B12 | 0x5*B9 | B8 |
+       n*B7 | m*B5 | vm);
 }
 
 
 void Assembler::vmla(const DwVfpRegister dst,
                      const DwVfpRegister src1,
                      const DwVfpRegister src2,
                      const Condition cond) {
-  // Instruction details available in ARM DDI 0406C.b, A8-892.
-  // cond(31-28) | 11100(27-23) | D=?(22) | 00(21-20) | Vn(19-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | N=?(7) | op(6)=0 | M=?(5) | 0(4) |
-  // Vm(3-0)
-  unsigned x = (cond | 0x1C*B23 | src1.code()*B16 |
-      dst.code()*B12 | 0x5*B9 | B8 | src2.code());
-  emit(x);
+  // Instruction details available in ARM DDI 0406C.b, A8-932.
+  // cond(31-28) | 11100(27-23) | D(22) | 00(21-20) | Vn(19-16) |
+  // Vd(15-12) | 101(11-9) | sz=1(8) | N(7) | op=0(6) | M(5) | 0(4) | Vm(3-0)
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vn, n;
+  src1.split_code(&vn, &n);
+  int vm, m;
+  src2.split_code(&vm, &m);
+  emit(cond | 0x1C*B23 | d*B22 | vn*B16 | vd*B12 | 0x5*B9 | B8 | n*B7 | m*B5 |
+       vm);
 }
 
 
 void Assembler::vdiv(const DwVfpRegister dst,
                      const DwVfpRegister src1,
                      const DwVfpRegister src2,
                      const Condition cond) {
   // Dd = vdiv(Dn, Dm) double precision floating point division.
   // Dd = D:Vd; Dm=M:Vm; Dn=N:Vm.
-  // Instruction details available in ARM DDI 0406A, A8-584.
-  // cond(31-28) | 11101(27-23)| D=?(22) | 00(21-20) | Vn(19-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | N(7)=? | 0(6) | M=?(5) | 0(4) | Vm(3-0)
+  // Instruction details available in ARM DDI 0406C.b, A8-882.
+  // cond(31-28) | 11101(27-23)| D(22) | 00(21-20) | Vn(19-16) |
+  // Vd(15-12) | 101(11-9) | sz=1(8) | N(7) | 0(6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 | B23 | src1.code()*B16 |
-       dst.code()*B12 | 0x5*B9 | B8 | src2.code());
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vn, n;
+  src1.split_code(&vn, &n);
+  int vm, m;
+  src2.split_code(&vm, &m);
+  emit(cond | 0x1D*B23 | d*B22 | vn*B16 | vd*B12 | 0x5*B9 | B8 | n*B7 | m*B5 |
+       vm);
 }
 
 
 void Assembler::vcmp(const DwVfpRegister src1,
                      const DwVfpRegister src2,
                      const Condition cond) {
   // vcmp(Dd, Dm) double precision floating point comparison.
-  // Instruction details available in ARM DDI 0406A, A8-570.
-  // cond(31-28) | 11101 (27-23)| D=?(22) | 11 (21-20) | 0100 (19-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | E(7)=0 | 1(6) | M(5)=? | 0(4) | Vm(3-0)
+  // Instruction details available in ARM DDI 0406C.b, A8-864.
+  // cond(31-28) | 11101(27-23)| D(22) | 11(21-20) | 0100(19-16) |
+  // Vd(15-12) | 101(11-9) | sz=1(8) | E=0(7) | 1(6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 |B23 | 0x3*B20 | B18 |
-       src1.code()*B12 | 0x5*B9 | B8 | B6 | src2.code());
+  int vd, d;
+  src1.split_code(&vd, &d);
+  int vm, m;
+  src2.split_code(&vm, &m);
+  emit(cond | 0x1D*B23 | d*B22 | 0x3*B20 | 0x4*B16 | vd*B12 | 0x5*B9 | B8 | B6 |
+       m*B5 | vm);
 }
 
 
 void Assembler::vcmp(const DwVfpRegister src1,
                      const double src2,
                      const Condition cond) {
-  // vcmp(Dd, Dm) double precision floating point comparison.
-  // Instruction details available in ARM DDI 0406A, A8-570.
-  // cond(31-28) | 11101 (27-23)| D=?(22) | 11 (21-20) | 0101 (19-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | E(7)=0 | 1(6) | M(5)=? | 0(4) | 0000(3-0)
+  // vcmp(Dd, #0.0) double precision floating point comparison.
+  // Instruction details available in ARM DDI 0406C.b, A8-864.
+  // cond(31-28) | 11101(27-23)| D(22) | 11(21-20) | 0101(19-16) |
+  // Vd(15-12) | 101(11-9) | sz=1(8) | E=0(7) | 1(6) | 0(5) | 0(4) | 0000(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   ASSERT(src2 == 0.0);
-  emit(cond | 0xE*B24 |B23 | 0x3*B20 | B18 | B16 |
-       src1.code()*B12 | 0x5*B9 | B8 | B6);
+  int vd, d;
+  src1.split_code(&vd, &d);
+  emit(cond | 0x1D*B23 | d*B22 | 0x3*B20 | 0x5*B16 | vd*B12 | 0x5*B9 | B8 | B6);
 }
 
 
 void Assembler::vmsr(Register dst, Condition cond) {
   // Instruction details available in ARM DDI 0406A, A8-652.
   // cond(31-28) | 1110 (27-24) | 1110(23-20)| 0001 (19-16) |
   // Rt(15-12) | 1010 (11-8) | 0(7) | 00 (6-5) | 1(4) | 0000(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   emit(cond | 0xE*B24 | 0xE*B20 |  B16 |
        dst.code()*B12 | 0xA*B8 | B4);
 }
 
 
 void Assembler::vmrs(Register dst, Condition cond) {
   // Instruction details available in ARM DDI 0406A, A8-652.
   // cond(31-28) | 1110 (27-24) | 1111(23-20)| 0001 (19-16) |
   // Rt(15-12) | 1010 (11-8) | 0(7) | 00 (6-5) | 1(4) | 0000(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
   emit(cond | 0xE*B24 | 0xF*B20 |  B16 |
        dst.code()*B12 | 0xA*B8 | B4);
 }
 
 
 void Assembler::vsqrt(const DwVfpRegister dst,
                       const DwVfpRegister src,
                       const Condition cond) {
-  // cond(31-28) | 11101 (27-23)| D=?(22) | 11 (21-20) | 0001 (19-16) |
-  // Vd(15-12) | 101(11-9) | sz(8)=1 | 11 (7-6) | M(5)=? | 0(4) | Vm(3-0)
+  // Instruction details available in ARM DDI 0406C.b, A8-1058.
+  // cond(31-28) | 11101(27-23)| D(22) | 11(21-20) | 0001(19-16) |
+  // Vd(15-12) | 101(11-9) | sz=1(8) | 11(7-6) | M(5) | 0(4) | Vm(3-0)
   ASSERT(CpuFeatures::IsEnabled(VFP2));
-  emit(cond | 0xE*B24 | B23 | 0x3*B20 | B16 |
-       dst.code()*B12 | 0x5*B9 | B8 | 3*B6 | src.code());
+  int vd, d;
+  dst.split_code(&vd, &d);
+  int vm, m;
+  src.split_code(&vm, &m);
+  emit(cond | 0x1D*B23 | d*B22 | 0x3*B20 | B16 | vd*B12 | 0x5*B9 | B8 | 0x3*B6 |
+       m*B5 | vm);
 }
 
 
 // Pseudo instructions.
 void Assembler::nop(int type) {
   // ARMv6{K/T2} and v7 have an actual NOP instruction but it serializes
   // some of the CPU's pipeline and has to issue. Older ARM chips simply used
   // MOV Rx, Rx as NOP and it performs better even in newer CPUs.
   // We therefore use MOV Rx, Rx, even on newer CPUs, and use Rx to encode
   // a type.
@@ skipping 390 matching lines @@
 
   // Since a constant pool was just emitted, move the check offset forward by
   // the standard interval.
   next_buffer_check_ = pc_offset() + kCheckPoolInterval;
 }
 
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_ARM