//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the "Instituto Nokia de Tecnologia" and
-// is distributed under the University of Illinois Open Source
+// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// Registers are identified with 4-bit ID numbers.
-class ARMReg<bits<4> num, string n, list<Register> aliases = []> : Register<n> {
+class ARMReg<bits<4> num, string n, list<Register> subregs = []> : Register<n> {
field bits<4> Num;
let Namespace = "ARM";
- let Aliases = aliases;
+ let SubRegs = subregs;
}
-class ARMFReg<bits<5> num, string n> : Register<n> {
- field bits<5> Num;
+class ARMFReg<bits<6> num, string n> : Register<n> {
+ field bits<6> Num;
let Namespace = "ARM";
}
+// Subregister indices.
+let Namespace = "ARM" in {
+// Note: Code depends on these having consecutive numbers.
+def ssub_0 : SubRegIndex { let NumberHack = 1; }
+def ssub_1 : SubRegIndex { let NumberHack = 2; }
+def ssub_2 : SubRegIndex { let NumberHack = 3; }
+def ssub_3 : SubRegIndex { let NumberHack = 4; }
+
+def dsub_0 : SubRegIndex { let NumberHack = 5; }
+def dsub_1 : SubRegIndex { let NumberHack = 6; }
+def dsub_2 : SubRegIndex { let NumberHack = 7; }
+def dsub_3 : SubRegIndex { let NumberHack = 8; }
+def dsub_4 : SubRegIndex { let NumberHack = 9; }
+def dsub_5 : SubRegIndex { let NumberHack = 10; }
+def dsub_6 : SubRegIndex { let NumberHack = 11; }
+def dsub_7 : SubRegIndex { let NumberHack = 12; }
+
+def qsub_0 : SubRegIndex { let NumberHack = 13; }
+def qsub_1 : SubRegIndex { let NumberHack = 14; }
+def qsub_2 : SubRegIndex { let NumberHack = 15; }
+def qsub_3 : SubRegIndex { let NumberHack = 16; }
+
+def qqsub_0 : SubRegIndex { let NumberHack = 17; }
+def qqsub_1 : SubRegIndex { let NumberHack = 18; }
+}
+
// Integer registers
-def R0 : ARMReg< 0, "r0">, DwarfRegNum<0>;
-def R1 : ARMReg< 1, "r1">, DwarfRegNum<1>;
-def R2 : ARMReg< 2, "r2">, DwarfRegNum<2>;
-def R3 : ARMReg< 3, "r3">, DwarfRegNum<3>;
-def R4 : ARMReg< 4, "r4">, DwarfRegNum<4>;
-def R5 : ARMReg< 5, "r5">, DwarfRegNum<5>;
-def R6 : ARMReg< 6, "r6">, DwarfRegNum<6>;
-def R7 : ARMReg< 7, "r7">, DwarfRegNum<7>;
-def R8 : ARMReg< 8, "r8">, DwarfRegNum<8>;
-def R9 : ARMReg< 9, "r9">, DwarfRegNum<9>;
-def R10 : ARMReg<10, "r10">, DwarfRegNum<10>;
-def R11 : ARMReg<11, "r11">, DwarfRegNum<11>;
-def R12 : ARMReg<12, "r12">, DwarfRegNum<12>;
-def SP : ARMReg<13, "sp">, DwarfRegNum<13>;
-def LR : ARMReg<14, "lr">, DwarfRegNum<14>;
-def PC : ARMReg<15, "pc">, DwarfRegNum<15>;
+def R0 : ARMReg< 0, "r0">, DwarfRegNum<[0]>;
+def R1 : ARMReg< 1, "r1">, DwarfRegNum<[1]>;
+def R2 : ARMReg< 2, "r2">, DwarfRegNum<[2]>;
+def R3 : ARMReg< 3, "r3">, DwarfRegNum<[3]>;
+def R4 : ARMReg< 4, "r4">, DwarfRegNum<[4]>;
+def R5 : ARMReg< 5, "r5">, DwarfRegNum<[5]>;
+def R6 : ARMReg< 6, "r6">, DwarfRegNum<[6]>;
+def R7 : ARMReg< 7, "r7">, DwarfRegNum<[7]>;
+def R8 : ARMReg< 8, "r8">, DwarfRegNum<[8]>;
+def R9 : ARMReg< 9, "r9">, DwarfRegNum<[9]>;
+def R10 : ARMReg<10, "r10">, DwarfRegNum<[10]>;
+def R11 : ARMReg<11, "r11">, DwarfRegNum<[11]>;
+def R12 : ARMReg<12, "r12">, DwarfRegNum<[12]>;
+def SP : ARMReg<13, "sp">, DwarfRegNum<[13]>;
+def LR : ARMReg<14, "lr">, DwarfRegNum<[14]>;
+def PC : ARMReg<15, "pc">, DwarfRegNum<[15]>;
// Float registers
def S0 : ARMFReg< 0, "s0">; def S1 : ARMFReg< 1, "s1">;
// Aliases of the F* registers used to hold 64-bit fp values (doubles)
def D0 : ARMReg< 0, "d0", [S0, S1]>;
-def D1 : ARMReg< 1, "d1", [S2, S3]>;
+def D1 : ARMReg< 1, "d1", [S2, S3]>;
def D2 : ARMReg< 2, "d2", [S4, S5]>;
def D3 : ARMReg< 3, "d3", [S6, S7]>;
def D4 : ARMReg< 4, "d4", [S8, S9]>;
def D14 : ARMReg<14, "d14", [S28, S29]>;
def D15 : ARMReg<15, "d15", [S30, S31]>;
+// VFP3 defines 16 additional double registers
+def D16 : ARMFReg<16, "d16">; def D17 : ARMFReg<17, "d17">;
+def D18 : ARMFReg<18, "d18">; def D19 : ARMFReg<19, "d19">;
+def D20 : ARMFReg<20, "d20">; def D21 : ARMFReg<21, "d21">;
+def D22 : ARMFReg<22, "d22">; def D23 : ARMFReg<23, "d23">;
+def D24 : ARMFReg<24, "d24">; def D25 : ARMFReg<25, "d25">;
+def D26 : ARMFReg<26, "d26">; def D27 : ARMFReg<27, "d27">;
+def D28 : ARMFReg<28, "d28">; def D29 : ARMFReg<29, "d29">;
+def D30 : ARMFReg<30, "d30">; def D31 : ARMFReg<31, "d31">;
+
+// Advanced SIMD (NEON) defines 16 quad-word aliases
+def Q0 : ARMReg< 0, "q0", [D0, D1]>;
+def Q1 : ARMReg< 1, "q1", [D2, D3]>;
+def Q2 : ARMReg< 2, "q2", [D4, D5]>;
+def Q3 : ARMReg< 3, "q3", [D6, D7]>;
+def Q4 : ARMReg< 4, "q4", [D8, D9]>;
+def Q5 : ARMReg< 5, "q5", [D10, D11]>;
+def Q6 : ARMReg< 6, "q6", [D12, D13]>;
+def Q7 : ARMReg< 7, "q7", [D14, D15]>;
+def Q8 : ARMReg< 8, "q8", [D16, D17]>;
+def Q9 : ARMReg< 9, "q9", [D18, D19]>;
+def Q10 : ARMReg<10, "q10", [D20, D21]>;
+def Q11 : ARMReg<11, "q11", [D22, D23]>;
+def Q12 : ARMReg<12, "q12", [D24, D25]>;
+def Q13 : ARMReg<13, "q13", [D26, D27]>;
+def Q14 : ARMReg<14, "q14", [D28, D29]>;
+def Q15 : ARMReg<15, "q15", [D30, D31]>;
+
+// Pseudo 256-bit registers to represent pairs of Q registers. These should
+// never be present in the emitted code.
+// These are used for NEON load / store instructions, e.g. vld4, vst3.
+// NOTE: It's possible to define more QQ registers since technical the
+// starting D register number doesn't have to be multiple of 4. e.g.
+// D1, D2, D3, D4 would be a legal quad. But that would make the sub-register
+// stuffs very messy.
+def QQ0 : ARMReg<0, "qq0", [Q0, Q1]>;
+def QQ1 : ARMReg<1, "qq1", [Q2, Q3]>;
+def QQ2 : ARMReg<2, "qq2", [Q4, Q5]>;
+def QQ3 : ARMReg<3, "qq3", [Q6, Q7]>;
+def QQ4 : ARMReg<4, "qq4", [Q8, Q9]>;
+def QQ5 : ARMReg<5, "qq5", [Q10, Q11]>;
+def QQ6 : ARMReg<6, "qq6", [Q12, Q13]>;
+def QQ7 : ARMReg<7, "qq7", [Q14, Q15]>;
+
+// Pseudo 512-bit registers to represent four consecutive Q registers.
+def QQQQ0 : ARMReg<0, "qqqq0", [QQ0, QQ1]>;
+def QQQQ1 : ARMReg<1, "qqqq1", [QQ2, QQ3]>;
+def QQQQ2 : ARMReg<2, "qqqq2", [QQ4, QQ5]>;
+def QQQQ3 : ARMReg<3, "qqqq3", [QQ6, QQ7]>;
+
+// Current Program Status Register.
+def CPSR : ARMReg<0, "cpsr">;
+
+def FPSCR : ARMReg<1, "fpscr">;
+
// Register classes.
//
// pc == Program Counter
// sp == Stack Pointer
// r12 == ip (scratch)
// r7 == Frame Pointer (thumb-style backtraces)
+// r9 == May be reserved as Thread Register
// r11 == Frame Pointer (arm-style backtraces)
// r10 == Stack Limit
//
def GPR : RegisterClass<"ARM", [i32], 32, [R0, R1, R2, R3, R4, R5, R6,
- R7, R8, R9, R10, R12, R11,
- LR, SP, PC]> {
+ R7, R8, R9, R10, R11, R12,
+ SP, LR, PC]> {
let MethodProtos = [{
iterator allocation_order_begin(const MachineFunction &MF) const;
iterator allocation_order_end(const MachineFunction &MF) const;
}];
- // FIXME: We are reserving r12 in case the PEI needs to use it to
- // generate large stack offset. Make it available once we have register
- // scavenging.
let MethodBodies = [{
// FP is R11, R9 is available.
static const unsigned ARM_GPR_AO_1[] = {
ARM::R0, ARM::R1, ARM::R2, ARM::R3,
+ ARM::R12,ARM::LR,
ARM::R4, ARM::R5, ARM::R6, ARM::R7,
ARM::R8, ARM::R9, ARM::R10,
- ARM::LR, ARM::R11 };
+ ARM::R11 };
// FP is R11, R9 is not available.
static const unsigned ARM_GPR_AO_2[] = {
ARM::R0, ARM::R1, ARM::R2, ARM::R3,
+ ARM::R12,ARM::LR,
ARM::R4, ARM::R5, ARM::R6, ARM::R7,
ARM::R8, ARM::R10,
- ARM::LR, ARM::R11 };
- // FP is R7, R9 is available.
+ ARM::R11 };
+ // FP is R7, R9 is available as non-callee-saved register.
+ // This is used by Darwin.
static const unsigned ARM_GPR_AO_3[] = {
ARM::R0, ARM::R1, ARM::R2, ARM::R3,
- ARM::R4, ARM::R5, ARM::R6, ARM::R8,
- ARM::R9, ARM::R10,ARM::R11,
- ARM::LR, ARM::R7 };
+ ARM::R9, ARM::R12,ARM::LR,
+ ARM::R4, ARM::R5, ARM::R6,
+ ARM::R8, ARM::R10,ARM::R11,ARM::R7 };
// FP is R7, R9 is not available.
static const unsigned ARM_GPR_AO_4[] = {
ARM::R0, ARM::R1, ARM::R2, ARM::R3,
- ARM::R4, ARM::R5, ARM::R6, ARM::R8,
- ARM::R10,ARM::R11,
- ARM::LR, ARM::R7 };
- // FP is R7, only low registers available.
+ ARM::R12,ARM::LR,
+ ARM::R4, ARM::R5, ARM::R6,
+ ARM::R8, ARM::R10,ARM::R11,
+ ARM::R7 };
+ // FP is R7, R9 is available as callee-saved register.
+ // This is used by non-Darwin platform in Thumb mode.
+ static const unsigned ARM_GPR_AO_5[] = {
+ ARM::R0, ARM::R1, ARM::R2, ARM::R3,
+ ARM::R12,ARM::LR,
+ ARM::R4, ARM::R5, ARM::R6,
+ ARM::R8, ARM::R9, ARM::R10,ARM::R11,ARM::R7 };
+
+ // For Thumb1 mode, we don't want to allocate hi regs at all, as we
+ // don't know how to spill them. If we make our prologue/epilogue code
+ // smarter at some point, we can go back to using the above allocation
+ // orders for the Thumb1 instructions that know how to use hi regs.
static const unsigned THUMB_GPR_AO[] = {
- ARM::R0, ARM::R1, ARM::R2,
+ ARM::R0, ARM::R1, ARM::R2, ARM::R3,
ARM::R4, ARM::R5, ARM::R6, ARM::R7 };
GPRClass::iterator
GPRClass::allocation_order_begin(const MachineFunction &MF) const {
const TargetMachine &TM = MF.getTarget();
const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
- if (Subtarget.isThumb())
+ if (Subtarget.isThumb1Only())
return THUMB_GPR_AO;
- if (Subtarget.useThumbBacktraces()) {
+ if (Subtarget.isTargetDarwin()) {
if (Subtarget.isR9Reserved())
return ARM_GPR_AO_4;
else
} else {
if (Subtarget.isR9Reserved())
return ARM_GPR_AO_2;
+ else if (Subtarget.isThumb())
+ return ARM_GPR_AO_5;
else
return ARM_GPR_AO_1;
}
GPRClass::iterator
GPRClass::allocation_order_end(const MachineFunction &MF) const {
const TargetMachine &TM = MF.getTarget();
- const MRegisterInfo *RI = TM.getRegisterInfo();
+ const TargetRegisterInfo *RI = TM.getRegisterInfo();
const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
GPRClass::iterator I;
- if (Subtarget.isThumb())
+
+ if (Subtarget.isThumb1Only()) {
I = THUMB_GPR_AO + (sizeof(THUMB_GPR_AO)/sizeof(unsigned));
- else if (Subtarget.useThumbBacktraces()) {
+ // Mac OS X requires FP not to be clobbered for backtracing purpose.
+ return (Subtarget.isTargetDarwin() || RI->hasFP(MF)) ? I-1 : I;
+ }
+
+ if (Subtarget.isTargetDarwin()) {
if (Subtarget.isR9Reserved())
I = ARM_GPR_AO_4 + (sizeof(ARM_GPR_AO_4)/sizeof(unsigned));
else
} else {
if (Subtarget.isR9Reserved())
I = ARM_GPR_AO_2 + (sizeof(ARM_GPR_AO_2)/sizeof(unsigned));
+ else if (Subtarget.isThumb())
+ I = ARM_GPR_AO_5 + (sizeof(ARM_GPR_AO_5)/sizeof(unsigned));
else
I = ARM_GPR_AO_1 + (sizeof(ARM_GPR_AO_1)/sizeof(unsigned));
}
}];
}
+// Thumb registers are R0-R7 normally. Some instructions can still use
+// the general GPR register class above (MOV, e.g.)
+def tGPR : RegisterClass<"ARM", [i32], 32, [R0, R1, R2, R3, R4, R5, R6, R7]> {
+ let MethodProtos = [{
+ iterator allocation_order_begin(const MachineFunction &MF) const;
+ iterator allocation_order_end(const MachineFunction &MF) const;
+ }];
+ let MethodBodies = [{
+ static const unsigned THUMB_tGPR_AO[] = {
+ ARM::R0, ARM::R1, ARM::R2, ARM::R3,
+ ARM::R4, ARM::R5, ARM::R6, ARM::R7 };
+
+ // FP is R7, only low registers available.
+ tGPRClass::iterator
+ tGPRClass::allocation_order_begin(const MachineFunction &MF) const {
+ return THUMB_tGPR_AO;
+ }
+
+ tGPRClass::iterator
+ tGPRClass::allocation_order_end(const MachineFunction &MF) const {
+ const TargetMachine &TM = MF.getTarget();
+ const TargetRegisterInfo *RI = TM.getRegisterInfo();
+ const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
+ tGPRClass::iterator I =
+ THUMB_tGPR_AO + (sizeof(THUMB_tGPR_AO)/sizeof(unsigned));
+ // Mac OS X requires FP not to be clobbered for backtracing purpose.
+ return (Subtarget.isTargetDarwin() || RI->hasFP(MF)) ? I-1 : I;
+ }
+ }];
+}
+
+// Scalar single precision floating point register class..
def SPR : RegisterClass<"ARM", [f32], 32, [S0, S1, S2, S3, S4, S5, S6, S7, S8,
S9, S10, S11, S12, S13, S14, S15, S16, S17, S18, S19, S20, S21, S22,
S23, S24, S25, S26, S27, S28, S29, S30, S31]>;
+// Subset of SPR which can be used as a source of NEON scalars for 16-bit
+// operations
+def SPR_8 : RegisterClass<"ARM", [f32], 32,
+ [S0, S1, S2, S3, S4, S5, S6, S7,
+ S8, S9, S10, S11, S12, S13, S14, S15]>;
+
+// Scalar double precision floating point / generic 64-bit vector register
+// class.
// ARM requires only word alignment for double. It's more performant if it
// is double-word alignment though.
-def DPR : RegisterClass<"ARM", [f64], 64, [D0, D1, D2, D3, D4, D5, D6, D7, D8,
- D9, D10, D11, D12, D13, D14, D15]>;
+def DPR : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
+ [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]> {
+ let MethodProtos = [{
+ iterator allocation_order_begin(const MachineFunction &MF) const;
+ iterator allocation_order_end(const MachineFunction &MF) const;
+ }];
+ let MethodBodies = [{
+ // VFP2
+ static const unsigned ARM_DPR_VFP2[] = {
+ ARM::D0, ARM::D1, ARM::D2, ARM::D3,
+ ARM::D4, ARM::D5, ARM::D6, ARM::D7,
+ ARM::D8, ARM::D9, ARM::D10, ARM::D11,
+ ARM::D12, ARM::D13, ARM::D14, ARM::D15 };
+ // VFP3
+ static const unsigned ARM_DPR_VFP3[] = {
+ ARM::D0, ARM::D1, ARM::D2, ARM::D3,
+ ARM::D4, ARM::D5, ARM::D6, ARM::D7,
+ ARM::D8, ARM::D9, ARM::D10, ARM::D11,
+ ARM::D12, ARM::D13, ARM::D14, ARM::D15,
+ ARM::D16, ARM::D17, ARM::D18, ARM::D19,
+ ARM::D20, ARM::D21, ARM::D22, ARM::D23,
+ ARM::D24, ARM::D25, ARM::D26, ARM::D27,
+ ARM::D28, ARM::D29, ARM::D30, ARM::D31 };
+ DPRClass::iterator
+ DPRClass::allocation_order_begin(const MachineFunction &MF) const {
+ const TargetMachine &TM = MF.getTarget();
+ const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
+ if (Subtarget.hasVFP3())
+ return ARM_DPR_VFP3;
+ return ARM_DPR_VFP2;
+ }
+
+ DPRClass::iterator
+ DPRClass::allocation_order_end(const MachineFunction &MF) const {
+ const TargetMachine &TM = MF.getTarget();
+ const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
+ if (Subtarget.hasVFP3())
+ return ARM_DPR_VFP3 + (sizeof(ARM_DPR_VFP3)/sizeof(unsigned));
+ else
+ return ARM_DPR_VFP2 + (sizeof(ARM_DPR_VFP2)/sizeof(unsigned));
+ }
+ }];
+}
+
+// Subset of DPR that are accessible with VFP2 (and so that also have
+// 32-bit SPR subregs).
+def DPR_VFP2 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
+ [D0, D1, D2, D3, D4, D5, D6, D7,
+ D8, D9, D10, D11, D12, D13, D14, D15]> {
+ let SubRegClasses = [(SPR ssub_0, ssub_1)];
+}
+
+// Subset of DPR which can be used as a source of NEON scalars for 16-bit
+// operations
+def DPR_8 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
+ [D0, D1, D2, D3, D4, D5, D6, D7]> {
+ let SubRegClasses = [(SPR_8 ssub_0, ssub_1)];
+}
+
+// Generic 128-bit vector register class.
+def QPR : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64], 128,
+ [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7,
+ Q8, Q9, Q10, Q11, Q12, Q13, Q14, Q15]> {
+ let SubRegClasses = [(DPR dsub_0, dsub_1)];
+}
+
+// Subset of QPR that have 32-bit SPR subregs.
+def QPR_VFP2 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
+ 128,
+ [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]> {
+ let SubRegClasses = [(SPR ssub_0, ssub_1, ssub_2, ssub_3),
+ (DPR_VFP2 dsub_0, dsub_1)];
+}
+
+// Subset of QPR that have DPR_8 and SPR_8 subregs.
+def QPR_8 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
+ 128,
+ [Q0, Q1, Q2, Q3]> {
+ let SubRegClasses = [(SPR_8 ssub_0, ssub_1, ssub_2, ssub_3),
+ (DPR_8 dsub_0, dsub_1)];
+}
+
+// Pseudo 256-bit vector register class to model pairs of Q registers
+// (4 consecutive D registers).
+def QQPR : RegisterClass<"ARM", [v4i64],
+ 256,
+ [QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7]> {
+ let SubRegClasses = [(DPR dsub_0, dsub_1, dsub_2, dsub_3),
+ (QPR qsub_0, qsub_1)];
+}
+
+// Subset of QQPR that have 32-bit SPR subregs.
+def QQPR_VFP2 : RegisterClass<"ARM", [v4i64],
+ 256,
+ [QQ0, QQ1, QQ2, QQ3]> {
+ let SubRegClasses = [(SPR ssub_0, ssub_1, ssub_2, ssub_3),
+ (DPR_VFP2 dsub_0, dsub_1, dsub_2, dsub_3),
+ (QPR_VFP2 qsub_0, qsub_1)];
+
+}
+
+// Pseudo 512-bit vector register class to model 4 consecutive Q registers
+// (8 consecutive D registers).
+def QQQQPR : RegisterClass<"ARM", [v8i64],
+ 256,
+ [QQQQ0, QQQQ1, QQQQ2, QQQQ3]> {
+ let SubRegClasses = [(DPR dsub_0, dsub_1, dsub_2, dsub_3,
+ dsub_4, dsub_5, dsub_6, dsub_7),
+ (QPR qsub_0, qsub_1, qsub_2, qsub_3)];
+}
+
+// Condition code registers.
+def CCR : RegisterClass<"ARM", [i32], 32, [CPSR]>;
+
+//===----------------------------------------------------------------------===//
+// Subregister Set Definitions... now that we have all of the pieces, define the
+// sub registers for each register.
+//
+
+// S sub-registers of D registers.
+def : SubRegSet<ssub_0, [D0, D1, D2, D3, D4, D5, D6, D7,
+ D8, D9, D10, D11, D12, D13, D14, D15],
+ [S0, S2, S4, S6, S8, S10, S12, S14,
+ S16, S18, S20, S22, S24, S26, S28, S30]>;
+def : SubRegSet<ssub_1, [D0, D1, D2, D3, D4, D5, D6, D7,
+ D8, D9, D10, D11, D12, D13, D14, D15],
+ [S1, S3, S5, S7, S9, S11, S13, S15,
+ S17, S19, S21, S23, S25, S27, S29, S31]>;
+
+// S sub-registers of Q registers.
+def : SubRegSet<ssub_0, [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7],
+ [S0, S4, S8, S12, S16, S20, S24, S28]>;
+def : SubRegSet<ssub_1, [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7],
+ [S1, S5, S9, S13, S17, S21, S25, S29]>;
+def : SubRegSet<ssub_2, [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7],
+ [S2, S6, S10, S14, S18, S22, S26, S30]>;
+def : SubRegSet<ssub_3, [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7],
+ [S3, S7, S11, S15, S19, S23, S27, S31]>;
+
+// D sub-registers of Q registers.
+def : SubRegSet<dsub_0, [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7,
+ Q8, Q9, Q10, Q11, Q12, Q13, Q14, Q15],
+ [D0, D2, D4, D6, D8, D10, D12, D14,
+ D16, D18, D20, D22, D24, D26, D28, D30]>;
+def : SubRegSet<dsub_1, [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7,
+ Q8, Q9, Q10, Q11, Q12, Q13, Q14, Q15],
+ [D1, D3, D5, D7, D9, D11, D13, D15,
+ D17, D19, D21, D23, D25, D27, D29, D31]>;
+
+// S sub-registers of QQ registers. Note there are no sub-indices
+// for referencing S4 - S7, S12 - S15, and S20 - S23. It doesn't
+// look like we need them.
+def : SubRegSet<ssub_0, [QQ0, QQ1, QQ2, QQ3],
+ [S0, S8, S16, S24]>;
+def : SubRegSet<ssub_1, [QQ0, QQ1, QQ2, QQ3],
+ [S1, S9, S17, S25]>;
+def : SubRegSet<ssub_2, [QQ0, QQ1, QQ2, QQ3],
+ [S2, S10, S18, S26]>;
+def : SubRegSet<ssub_3, [QQ0, QQ1, QQ2, QQ3],
+ [S3, S11, S19, S27]>;
+
+// D sub-registers of QQ registers.
+def : SubRegSet<dsub_0, [QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7],
+ [D0, D4, D8, D12, D16, D20, D24, D28]>;
+def : SubRegSet<dsub_1, [QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7],
+ [D1, D5, D9, D13, D17, D21, D25, D29]>;
+def : SubRegSet<dsub_2, [QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7],
+ [D2, D6, D10, D14, D18, D22, D26, D30]>;
+def : SubRegSet<dsub_3, [QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7],
+ [D3, D7, D11, D15, D19, D23, D27, D31]>;
+
+// Q sub-registers of QQ registers.
+def : SubRegSet<qsub_0, [QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7],
+ [Q0, Q2, Q4, Q6, Q8, Q10, Q12, Q14]>;
+def : SubRegSet<qsub_1,[QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7],
+ [Q1, Q3, Q5, Q7, Q9, Q11, Q13, Q15]>;
+
+
+// D sub-registers of QQQQ registers.
+def : SubRegSet<dsub_0, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [D0, D8, D16, D24]>;
+def : SubRegSet<dsub_1, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [D1, D9, D17, D25]>;
+def : SubRegSet<dsub_2, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [D2, D10, D18, D26]>;
+def : SubRegSet<dsub_3, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [D3, D11, D19, D27]>;
+
+def : SubRegSet<dsub_4, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [D4, D12, D20, D28]>;
+def : SubRegSet<dsub_5, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [D5, D13, D21, D29]>;
+def : SubRegSet<dsub_6, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [D6, D14, D22, D30]>;
+def : SubRegSet<dsub_7, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [D7, D15, D23, D31]>;
+
+// Q sub-registers of QQQQ registers.
+def : SubRegSet<qsub_0, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [Q0, Q4, Q8, Q12]>;
+def : SubRegSet<qsub_1, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [Q1, Q5, Q9, Q13]>;
+def : SubRegSet<qsub_2, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [Q2, Q6, Q10, Q14]>;
+def : SubRegSet<qsub_3, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [Q3, Q7, Q11, Q15]>;
+
+// QQ sub-registers of QQQQ registers.
+def : SubRegSet<qqsub_0, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [QQ0, QQ2, QQ4, QQ6]>;
+def : SubRegSet<qqsub_1, [QQQQ0, QQQQ1, QQQQ2, QQQQ3],
+ [QQ1, QQ3, QQ5, QQ7]>;
+