1 //===- ARMRegisterInfo.td - ARM Register defs -------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 //===----------------------------------------------------------------------===//
11 // Declarations that describe the ARM register file
12 //===----------------------------------------------------------------------===//
14 // Registers are identified with 4-bit ID numbers.
15 class ARMReg<bits<4> num, string n, list<Register> subregs = []> : Register<n> {
17 let Namespace = "ARM";
18 let SubRegs = subregs;
21 class ARMFReg<bits<6> num, string n> : Register<n> {
23 let Namespace = "ARM";
26 // Subregister indices.
27 let Namespace = "ARM" in {
28 // Note: Code depends on these having consecutive numbers.
29 def ssub_0 : SubRegIndex;
30 def ssub_1 : SubRegIndex;
31 def ssub_2 : SubRegIndex;
32 def ssub_3 : SubRegIndex;
34 def dsub_0 : SubRegIndex;
35 def dsub_1 : SubRegIndex;
36 def dsub_2 : SubRegIndex;
37 def dsub_3 : SubRegIndex;
38 def dsub_4 : SubRegIndex;
39 def dsub_5 : SubRegIndex;
40 def dsub_6 : SubRegIndex;
41 def dsub_7 : SubRegIndex;
43 def qsub_0 : SubRegIndex;
44 def qsub_1 : SubRegIndex;
45 def qsub_2 : SubRegIndex;
46 def qsub_3 : SubRegIndex;
48 def qqsub_0 : SubRegIndex;
49 def qqsub_1 : SubRegIndex;
53 def R0 : ARMReg< 0, "r0">, DwarfRegNum<[0]>;
54 def R1 : ARMReg< 1, "r1">, DwarfRegNum<[1]>;
55 def R2 : ARMReg< 2, "r2">, DwarfRegNum<[2]>;
56 def R3 : ARMReg< 3, "r3">, DwarfRegNum<[3]>;
57 def R4 : ARMReg< 4, "r4">, DwarfRegNum<[4]>;
58 def R5 : ARMReg< 5, "r5">, DwarfRegNum<[5]>;
59 def R6 : ARMReg< 6, "r6">, DwarfRegNum<[6]>;
60 def R7 : ARMReg< 7, "r7">, DwarfRegNum<[7]>;
61 def R8 : ARMReg< 8, "r8">, DwarfRegNum<[8]>;
62 def R9 : ARMReg< 9, "r9">, DwarfRegNum<[9]>;
63 def R10 : ARMReg<10, "r10">, DwarfRegNum<[10]>;
64 def R11 : ARMReg<11, "r11">, DwarfRegNum<[11]>;
65 def R12 : ARMReg<12, "r12">, DwarfRegNum<[12]>;
66 def SP : ARMReg<13, "sp">, DwarfRegNum<[13]>;
67 def LR : ARMReg<14, "lr">, DwarfRegNum<[14]>;
68 def PC : ARMReg<15, "pc">, DwarfRegNum<[15]>;
71 def S0 : ARMFReg< 0, "s0">; def S1 : ARMFReg< 1, "s1">;
72 def S2 : ARMFReg< 2, "s2">; def S3 : ARMFReg< 3, "s3">;
73 def S4 : ARMFReg< 4, "s4">; def S5 : ARMFReg< 5, "s5">;
74 def S6 : ARMFReg< 6, "s6">; def S7 : ARMFReg< 7, "s7">;
75 def S8 : ARMFReg< 8, "s8">; def S9 : ARMFReg< 9, "s9">;
76 def S10 : ARMFReg<10, "s10">; def S11 : ARMFReg<11, "s11">;
77 def S12 : ARMFReg<12, "s12">; def S13 : ARMFReg<13, "s13">;
78 def S14 : ARMFReg<14, "s14">; def S15 : ARMFReg<15, "s15">;
79 def S16 : ARMFReg<16, "s16">; def S17 : ARMFReg<17, "s17">;
80 def S18 : ARMFReg<18, "s18">; def S19 : ARMFReg<19, "s19">;
81 def S20 : ARMFReg<20, "s20">; def S21 : ARMFReg<21, "s21">;
82 def S22 : ARMFReg<22, "s22">; def S23 : ARMFReg<23, "s23">;
83 def S24 : ARMFReg<24, "s24">; def S25 : ARMFReg<25, "s25">;
84 def S26 : ARMFReg<26, "s26">; def S27 : ARMFReg<27, "s27">;
85 def S28 : ARMFReg<28, "s28">; def S29 : ARMFReg<29, "s29">;
86 def S30 : ARMFReg<30, "s30">; def S31 : ARMFReg<31, "s31">;
88 // Aliases of the F* registers used to hold 64-bit fp values (doubles)
89 let SubRegIndices = [ssub_0, ssub_1] in {
90 def D0 : ARMReg< 0, "d0", [S0, S1]>;
91 def D1 : ARMReg< 1, "d1", [S2, S3]>;
92 def D2 : ARMReg< 2, "d2", [S4, S5]>;
93 def D3 : ARMReg< 3, "d3", [S6, S7]>;
94 def D4 : ARMReg< 4, "d4", [S8, S9]>;
95 def D5 : ARMReg< 5, "d5", [S10, S11]>;
96 def D6 : ARMReg< 6, "d6", [S12, S13]>;
97 def D7 : ARMReg< 7, "d7", [S14, S15]>;
98 def D8 : ARMReg< 8, "d8", [S16, S17]>;
99 def D9 : ARMReg< 9, "d9", [S18, S19]>;
100 def D10 : ARMReg<10, "d10", [S20, S21]>;
101 def D11 : ARMReg<11, "d11", [S22, S23]>;
102 def D12 : ARMReg<12, "d12", [S24, S25]>;
103 def D13 : ARMReg<13, "d13", [S26, S27]>;
104 def D14 : ARMReg<14, "d14", [S28, S29]>;
105 def D15 : ARMReg<15, "d15", [S30, S31]>;
108 // VFP3 defines 16 additional double registers
109 def D16 : ARMFReg<16, "d16">; def D17 : ARMFReg<17, "d17">;
110 def D18 : ARMFReg<18, "d18">; def D19 : ARMFReg<19, "d19">;
111 def D20 : ARMFReg<20, "d20">; def D21 : ARMFReg<21, "d21">;
112 def D22 : ARMFReg<22, "d22">; def D23 : ARMFReg<23, "d23">;
113 def D24 : ARMFReg<24, "d24">; def D25 : ARMFReg<25, "d25">;
114 def D26 : ARMFReg<26, "d26">; def D27 : ARMFReg<27, "d27">;
115 def D28 : ARMFReg<28, "d28">; def D29 : ARMFReg<29, "d29">;
116 def D30 : ARMFReg<30, "d30">; def D31 : ARMFReg<31, "d31">;
118 // Advanced SIMD (NEON) defines 16 quad-word aliases
119 let SubRegIndices = [dsub_0, dsub_1],
120 CompositeIndices = [(ssub_2 dsub_1, ssub_0),
121 (ssub_3 dsub_1, ssub_1)] in {
122 def Q0 : ARMReg< 0, "q0", [D0, D1]>;
123 def Q1 : ARMReg< 1, "q1", [D2, D3]>;
124 def Q2 : ARMReg< 2, "q2", [D4, D5]>;
125 def Q3 : ARMReg< 3, "q3", [D6, D7]>;
126 def Q4 : ARMReg< 4, "q4", [D8, D9]>;
127 def Q5 : ARMReg< 5, "q5", [D10, D11]>;
128 def Q6 : ARMReg< 6, "q6", [D12, D13]>;
129 def Q7 : ARMReg< 7, "q7", [D14, D15]>;
131 let SubRegIndices = [dsub_0, dsub_1] in {
132 def Q8 : ARMReg< 8, "q8", [D16, D17]>;
133 def Q9 : ARMReg< 9, "q9", [D18, D19]>;
134 def Q10 : ARMReg<10, "q10", [D20, D21]>;
135 def Q11 : ARMReg<11, "q11", [D22, D23]>;
136 def Q12 : ARMReg<12, "q12", [D24, D25]>;
137 def Q13 : ARMReg<13, "q13", [D26, D27]>;
138 def Q14 : ARMReg<14, "q14", [D28, D29]>;
139 def Q15 : ARMReg<15, "q15", [D30, D31]>;
142 // Pseudo 256-bit registers to represent pairs of Q registers. These should
143 // never be present in the emitted code.
144 // These are used for NEON load / store instructions, e.g. vld4, vst3.
145 // NOTE: It's possible to define more QQ registers since technical the
146 // starting D register number doesn't have to be multiple of 4. e.g.
147 // D1, D2, D3, D4 would be a legal quad. But that would make the sub-register
148 // stuffs very messy.
149 let SubRegIndices = [qsub_0, qsub_1],
150 CompositeIndices = [(dsub_2 qsub_1, dsub_0),
151 (dsub_3 qsub_1, dsub_1)] in {
152 def QQ0 : ARMReg<0, "qq0", [Q0, Q1]>;
153 def QQ1 : ARMReg<1, "qq1", [Q2, Q3]>;
154 def QQ2 : ARMReg<2, "qq2", [Q4, Q5]>;
155 def QQ3 : ARMReg<3, "qq3", [Q6, Q7]>;
156 def QQ4 : ARMReg<4, "qq4", [Q8, Q9]>;
157 def QQ5 : ARMReg<5, "qq5", [Q10, Q11]>;
158 def QQ6 : ARMReg<6, "qq6", [Q12, Q13]>;
159 def QQ7 : ARMReg<7, "qq7", [Q14, Q15]>;
162 // Pseudo 512-bit registers to represent four consecutive Q registers.
163 let SubRegIndices = [qqsub_0, qqsub_1],
164 CompositeIndices = [(qsub_2 qqsub_1, qsub_0),
165 (qsub_3 qqsub_1, qsub_1),
166 (dsub_4 qqsub_1, dsub_0),
167 (dsub_5 qqsub_1, dsub_1),
168 (dsub_6 qqsub_1, dsub_2),
169 (dsub_7 qqsub_1, dsub_3)] in {
170 def QQQQ0 : ARMReg<0, "qqqq0", [QQ0, QQ1]>;
171 def QQQQ1 : ARMReg<1, "qqqq1", [QQ2, QQ3]>;
172 def QQQQ2 : ARMReg<2, "qqqq2", [QQ4, QQ5]>;
173 def QQQQ3 : ARMReg<3, "qqqq3", [QQ6, QQ7]>;
176 // Current Program Status Register.
177 def CPSR : ARMReg<0, "cpsr">;
179 def FPSCR : ARMReg<1, "fpscr">;
183 // pc == Program Counter
184 // lr == Link Register
185 // sp == Stack Pointer
186 // r12 == ip (scratch)
187 // r7 == Frame Pointer (thumb-style backtraces)
188 // r9 == May be reserved as Thread Register
189 // r11 == Frame Pointer (arm-style backtraces)
190 // r10 == Stack Limit
192 def GPR : RegisterClass<"ARM", [i32], 32, [R0, R1, R2, R3, R4, R5, R6,
193 R7, R8, R9, R10, R11, R12,
195 let MethodProtos = [{
196 iterator allocation_order_begin(const MachineFunction &MF) const;
197 iterator allocation_order_end(const MachineFunction &MF) const;
199 let MethodBodies = [{
200 // FP is R11, R9 is available.
201 static const unsigned ARM_GPR_AO_1[] = {
202 ARM::R0, ARM::R1, ARM::R2, ARM::R3,
204 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
205 ARM::R8, ARM::R9, ARM::R10,
207 // FP is R11, R9 is not available.
208 static const unsigned ARM_GPR_AO_2[] = {
209 ARM::R0, ARM::R1, ARM::R2, ARM::R3,
211 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
214 // FP is R7, R9 is available as non-callee-saved register.
215 // This is used by Darwin.
216 static const unsigned ARM_GPR_AO_3[] = {
217 ARM::R0, ARM::R1, ARM::R2, ARM::R3,
218 ARM::R9, ARM::R12,ARM::LR,
219 ARM::R4, ARM::R5, ARM::R6,
220 ARM::R8, ARM::R10,ARM::R11,ARM::R7 };
221 // FP is R7, R9 is not available.
222 static const unsigned ARM_GPR_AO_4[] = {
223 ARM::R0, ARM::R1, ARM::R2, ARM::R3,
225 ARM::R4, ARM::R5, ARM::R6,
226 ARM::R8, ARM::R10,ARM::R11,
228 // FP is R7, R9 is available as callee-saved register.
229 // This is used by non-Darwin platform in Thumb mode.
230 static const unsigned ARM_GPR_AO_5[] = {
231 ARM::R0, ARM::R1, ARM::R2, ARM::R3,
233 ARM::R4, ARM::R5, ARM::R6,
234 ARM::R8, ARM::R9, ARM::R10,ARM::R11,ARM::R7 };
236 // For Thumb1 mode, we don't want to allocate hi regs at all, as we
237 // don't know how to spill them. If we make our prologue/epilogue code
238 // smarter at some point, we can go back to using the above allocation
239 // orders for the Thumb1 instructions that know how to use hi regs.
240 static const unsigned THUMB_GPR_AO[] = {
241 ARM::R0, ARM::R1, ARM::R2, ARM::R3,
242 ARM::R4, ARM::R5, ARM::R6, ARM::R7 };
245 GPRClass::allocation_order_begin(const MachineFunction &MF) const {
246 const TargetMachine &TM = MF.getTarget();
247 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
248 if (Subtarget.isThumb1Only())
250 if (Subtarget.isTargetDarwin()) {
251 if (Subtarget.isR9Reserved())
256 if (Subtarget.isR9Reserved())
258 else if (Subtarget.isThumb())
266 GPRClass::allocation_order_end(const MachineFunction &MF) const {
267 const TargetMachine &TM = MF.getTarget();
268 const TargetRegisterInfo *RI = TM.getRegisterInfo();
269 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
270 GPRClass::iterator I;
272 if (Subtarget.isThumb1Only()) {
273 I = THUMB_GPR_AO + (sizeof(THUMB_GPR_AO)/sizeof(unsigned));
274 // Mac OS X requires FP not to be clobbered for backtracing purpose.
275 return (Subtarget.isTargetDarwin() || RI->hasFP(MF)) ? I-1 : I;
278 if (Subtarget.isTargetDarwin()) {
279 if (Subtarget.isR9Reserved())
280 I = ARM_GPR_AO_4 + (sizeof(ARM_GPR_AO_4)/sizeof(unsigned));
282 I = ARM_GPR_AO_3 + (sizeof(ARM_GPR_AO_3)/sizeof(unsigned));
284 if (Subtarget.isR9Reserved())
285 I = ARM_GPR_AO_2 + (sizeof(ARM_GPR_AO_2)/sizeof(unsigned));
286 else if (Subtarget.isThumb())
287 I = ARM_GPR_AO_5 + (sizeof(ARM_GPR_AO_5)/sizeof(unsigned));
289 I = ARM_GPR_AO_1 + (sizeof(ARM_GPR_AO_1)/sizeof(unsigned));
292 // Mac OS X requires FP not to be clobbered for backtracing purpose.
293 return (Subtarget.isTargetDarwin() || RI->hasFP(MF)) ? I-1 : I;
298 // Thumb registers are R0-R7 normally. Some instructions can still use
299 // the general GPR register class above (MOV, e.g.)
300 def tGPR : RegisterClass<"ARM", [i32], 32, [R0, R1, R2, R3, R4, R5, R6, R7]> {
301 let MethodProtos = [{
302 iterator allocation_order_begin(const MachineFunction &MF) const;
303 iterator allocation_order_end(const MachineFunction &MF) const;
305 let MethodBodies = [{
306 static const unsigned THUMB_tGPR_AO[] = {
307 ARM::R0, ARM::R1, ARM::R2, ARM::R3,
308 ARM::R4, ARM::R5, ARM::R6, ARM::R7 };
310 // FP is R7, only low registers available.
312 tGPRClass::allocation_order_begin(const MachineFunction &MF) const {
313 return THUMB_tGPR_AO;
317 tGPRClass::allocation_order_end(const MachineFunction &MF) const {
318 const TargetMachine &TM = MF.getTarget();
319 const TargetRegisterInfo *RI = TM.getRegisterInfo();
320 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
321 tGPRClass::iterator I =
322 THUMB_tGPR_AO + (sizeof(THUMB_tGPR_AO)/sizeof(unsigned));
323 // Mac OS X requires FP not to be clobbered for backtracing purpose.
324 return (Subtarget.isTargetDarwin() || RI->hasFP(MF)) ? I-1 : I;
329 // Scalar single precision floating point register class..
330 def SPR : RegisterClass<"ARM", [f32], 32, [S0, S1, S2, S3, S4, S5, S6, S7, S8,
331 S9, S10, S11, S12, S13, S14, S15, S16, S17, S18, S19, S20, S21, S22,
332 S23, S24, S25, S26, S27, S28, S29, S30, S31]>;
334 // Subset of SPR which can be used as a source of NEON scalars for 16-bit
336 def SPR_8 : RegisterClass<"ARM", [f32], 32,
337 [S0, S1, S2, S3, S4, S5, S6, S7,
338 S8, S9, S10, S11, S12, S13, S14, S15]>;
340 // Scalar double precision floating point / generic 64-bit vector register
342 // ARM requires only word alignment for double. It's more performant if it
343 // is double-word alignment though.
344 def DPR : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
345 [D0, D1, D2, D3, D4, D5, D6, D7,
346 D8, D9, D10, D11, D12, D13, D14, D15,
347 D16, D17, D18, D19, D20, D21, D22, D23,
348 D24, D25, D26, D27, D28, D29, D30, D31]> {
349 let MethodProtos = [{
350 iterator allocation_order_begin(const MachineFunction &MF) const;
351 iterator allocation_order_end(const MachineFunction &MF) const;
353 let MethodBodies = [{
355 static const unsigned ARM_DPR_VFP2[] = {
356 ARM::D0, ARM::D1, ARM::D2, ARM::D3,
357 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
358 ARM::D8, ARM::D9, ARM::D10, ARM::D11,
359 ARM::D12, ARM::D13, ARM::D14, ARM::D15 };
361 static const unsigned ARM_DPR_VFP3[] = {
362 ARM::D0, ARM::D1, ARM::D2, ARM::D3,
363 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
364 ARM::D8, ARM::D9, ARM::D10, ARM::D11,
365 ARM::D12, ARM::D13, ARM::D14, ARM::D15,
366 ARM::D16, ARM::D17, ARM::D18, ARM::D19,
367 ARM::D20, ARM::D21, ARM::D22, ARM::D23,
368 ARM::D24, ARM::D25, ARM::D26, ARM::D27,
369 ARM::D28, ARM::D29, ARM::D30, ARM::D31 };
371 DPRClass::allocation_order_begin(const MachineFunction &MF) const {
372 const TargetMachine &TM = MF.getTarget();
373 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
374 if (Subtarget.hasVFP3())
380 DPRClass::allocation_order_end(const MachineFunction &MF) const {
381 const TargetMachine &TM = MF.getTarget();
382 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>();
383 if (Subtarget.hasVFP3())
384 return ARM_DPR_VFP3 + (sizeof(ARM_DPR_VFP3)/sizeof(unsigned));
386 return ARM_DPR_VFP2 + (sizeof(ARM_DPR_VFP2)/sizeof(unsigned));
391 // Subset of DPR that are accessible with VFP2 (and so that also have
392 // 32-bit SPR subregs).
393 def DPR_VFP2 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
394 [D0, D1, D2, D3, D4, D5, D6, D7,
395 D8, D9, D10, D11, D12, D13, D14, D15]> {
396 let SubRegClasses = [(SPR ssub_0, ssub_1)];
399 // Subset of DPR which can be used as a source of NEON scalars for 16-bit
401 def DPR_8 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
402 [D0, D1, D2, D3, D4, D5, D6, D7]> {
403 let SubRegClasses = [(SPR_8 ssub_0, ssub_1)];
406 // Generic 128-bit vector register class.
407 def QPR : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64], 128,
408 [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7,
409 Q8, Q9, Q10, Q11, Q12, Q13, Q14, Q15]> {
410 let SubRegClasses = [(DPR dsub_0, dsub_1)];
413 // Subset of QPR that have 32-bit SPR subregs.
414 def QPR_VFP2 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
416 [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]> {
417 let SubRegClasses = [(SPR ssub_0, ssub_1, ssub_2, ssub_3),
418 (DPR_VFP2 dsub_0, dsub_1)];
421 // Subset of QPR that have DPR_8 and SPR_8 subregs.
422 def QPR_8 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
425 let SubRegClasses = [(SPR_8 ssub_0, ssub_1, ssub_2, ssub_3),
426 (DPR_8 dsub_0, dsub_1)];
429 // Pseudo 256-bit vector register class to model pairs of Q registers
430 // (4 consecutive D registers).
431 def QQPR : RegisterClass<"ARM", [v4i64],
433 [QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7]> {
434 let SubRegClasses = [(DPR dsub_0, dsub_1, dsub_2, dsub_3),
435 (QPR qsub_0, qsub_1)];
438 // Subset of QQPR that have 32-bit SPR subregs.
439 def QQPR_VFP2 : RegisterClass<"ARM", [v4i64],
441 [QQ0, QQ1, QQ2, QQ3]> {
442 let SubRegClasses = [(SPR ssub_0, ssub_1, ssub_2, ssub_3),
443 (DPR_VFP2 dsub_0, dsub_1, dsub_2, dsub_3),
444 (QPR_VFP2 qsub_0, qsub_1)];
448 // Pseudo 512-bit vector register class to model 4 consecutive Q registers
449 // (8 consecutive D registers).
450 def QQQQPR : RegisterClass<"ARM", [v8i64],
452 [QQQQ0, QQQQ1, QQQQ2, QQQQ3]> {
453 let SubRegClasses = [(DPR dsub_0, dsub_1, dsub_2, dsub_3,
454 dsub_4, dsub_5, dsub_6, dsub_7),
455 (QPR qsub_0, qsub_1, qsub_2, qsub_3)];
458 // Condition code registers.
459 def CCR : RegisterClass<"ARM", [i32], 32, [CPSR]>;