1 //===- ARMRegisterInfo.td - ARM Register defs --------------*- tablegen -*-===//
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; // In a Q reg.
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 // These require 32-bit instructions.
62 let CostPerUse = 1 in {
63 def R8 : ARMReg< 8, "r8">, DwarfRegNum<[8]>;
64 def R9 : ARMReg< 9, "r9">, DwarfRegNum<[9]>;
65 def R10 : ARMReg<10, "r10">, DwarfRegNum<[10]>;
66 def R11 : ARMReg<11, "r11">, DwarfRegNum<[11]>;
67 def R12 : ARMReg<12, "r12">, DwarfRegNum<[12]>;
68 def SP : ARMReg<13, "sp">, DwarfRegNum<[13]>;
69 def LR : ARMReg<14, "lr">, DwarfRegNum<[14]>;
70 def PC : ARMReg<15, "pc">, DwarfRegNum<[15]>;
74 def S0 : ARMFReg< 0, "s0">; def S1 : ARMFReg< 1, "s1">;
75 def S2 : ARMFReg< 2, "s2">; def S3 : ARMFReg< 3, "s3">;
76 def S4 : ARMFReg< 4, "s4">; def S5 : ARMFReg< 5, "s5">;
77 def S6 : ARMFReg< 6, "s6">; def S7 : ARMFReg< 7, "s7">;
78 def S8 : ARMFReg< 8, "s8">; def S9 : ARMFReg< 9, "s9">;
79 def S10 : ARMFReg<10, "s10">; def S11 : ARMFReg<11, "s11">;
80 def S12 : ARMFReg<12, "s12">; def S13 : ARMFReg<13, "s13">;
81 def S14 : ARMFReg<14, "s14">; def S15 : ARMFReg<15, "s15">;
82 def S16 : ARMFReg<16, "s16">; def S17 : ARMFReg<17, "s17">;
83 def S18 : ARMFReg<18, "s18">; def S19 : ARMFReg<19, "s19">;
84 def S20 : ARMFReg<20, "s20">; def S21 : ARMFReg<21, "s21">;
85 def S22 : ARMFReg<22, "s22">; def S23 : ARMFReg<23, "s23">;
86 def S24 : ARMFReg<24, "s24">; def S25 : ARMFReg<25, "s25">;
87 def S26 : ARMFReg<26, "s26">; def S27 : ARMFReg<27, "s27">;
88 def S28 : ARMFReg<28, "s28">; def S29 : ARMFReg<29, "s29">;
89 def S30 : ARMFReg<30, "s30">; def S31 : ARMFReg<31, "s31">;
91 // Aliases of the F* registers used to hold 64-bit fp values (doubles)
92 let SubRegIndices = [ssub_0, ssub_1] in {
93 def D0 : ARMReg< 0, "d0", [S0, S1]>, DwarfRegNum<[256]>;
94 def D1 : ARMReg< 1, "d1", [S2, S3]>, DwarfRegNum<[257]>;
95 def D2 : ARMReg< 2, "d2", [S4, S5]>, DwarfRegNum<[258]>;
96 def D3 : ARMReg< 3, "d3", [S6, S7]>, DwarfRegNum<[259]>;
97 def D4 : ARMReg< 4, "d4", [S8, S9]>, DwarfRegNum<[260]>;
98 def D5 : ARMReg< 5, "d5", [S10, S11]>, DwarfRegNum<[261]>;
99 def D6 : ARMReg< 6, "d6", [S12, S13]>, DwarfRegNum<[262]>;
100 def D7 : ARMReg< 7, "d7", [S14, S15]>, DwarfRegNum<[263]>;
101 def D8 : ARMReg< 8, "d8", [S16, S17]>, DwarfRegNum<[264]>;
102 def D9 : ARMReg< 9, "d9", [S18, S19]>, DwarfRegNum<[265]>;
103 def D10 : ARMReg<10, "d10", [S20, S21]>, DwarfRegNum<[266]>;
104 def D11 : ARMReg<11, "d11", [S22, S23]>, DwarfRegNum<[267]>;
105 def D12 : ARMReg<12, "d12", [S24, S25]>, DwarfRegNum<[268]>;
106 def D13 : ARMReg<13, "d13", [S26, S27]>, DwarfRegNum<[269]>;
107 def D14 : ARMReg<14, "d14", [S28, S29]>, DwarfRegNum<[270]>;
108 def D15 : ARMReg<15, "d15", [S30, S31]>, DwarfRegNum<[271]>;
111 // VFP3 defines 16 additional double registers
112 def D16 : ARMFReg<16, "d16">, DwarfRegNum<[272]>;
113 def D17 : ARMFReg<17, "d17">, DwarfRegNum<[273]>;
114 def D18 : ARMFReg<18, "d18">, DwarfRegNum<[274]>;
115 def D19 : ARMFReg<19, "d19">, DwarfRegNum<[275]>;
116 def D20 : ARMFReg<20, "d20">, DwarfRegNum<[276]>;
117 def D21 : ARMFReg<21, "d21">, DwarfRegNum<[277]>;
118 def D22 : ARMFReg<22, "d22">, DwarfRegNum<[278]>;
119 def D23 : ARMFReg<23, "d23">, DwarfRegNum<[279]>;
120 def D24 : ARMFReg<24, "d24">, DwarfRegNum<[280]>;
121 def D25 : ARMFReg<25, "d25">, DwarfRegNum<[281]>;
122 def D26 : ARMFReg<26, "d26">, DwarfRegNum<[282]>;
123 def D27 : ARMFReg<27, "d27">, DwarfRegNum<[283]>;
124 def D28 : ARMFReg<28, "d28">, DwarfRegNum<[284]>;
125 def D29 : ARMFReg<29, "d29">, DwarfRegNum<[285]>;
126 def D30 : ARMFReg<30, "d30">, DwarfRegNum<[286]>;
127 def D31 : ARMFReg<31, "d31">, DwarfRegNum<[287]>;
129 // Advanced SIMD (NEON) defines 16 quad-word aliases
130 let SubRegIndices = [dsub_0, dsub_1],
131 CompositeIndices = [(ssub_2 dsub_1, ssub_0),
132 (ssub_3 dsub_1, ssub_1)] in {
133 def Q0 : ARMReg< 0, "q0", [D0, D1]>;
134 def Q1 : ARMReg< 1, "q1", [D2, D3]>;
135 def Q2 : ARMReg< 2, "q2", [D4, D5]>;
136 def Q3 : ARMReg< 3, "q3", [D6, D7]>;
137 def Q4 : ARMReg< 4, "q4", [D8, D9]>;
138 def Q5 : ARMReg< 5, "q5", [D10, D11]>;
139 def Q6 : ARMReg< 6, "q6", [D12, D13]>;
140 def Q7 : ARMReg< 7, "q7", [D14, D15]>;
142 let SubRegIndices = [dsub_0, dsub_1] in {
143 def Q8 : ARMReg< 8, "q8", [D16, D17]>;
144 def Q9 : ARMReg< 9, "q9", [D18, D19]>;
145 def Q10 : ARMReg<10, "q10", [D20, D21]>;
146 def Q11 : ARMReg<11, "q11", [D22, D23]>;
147 def Q12 : ARMReg<12, "q12", [D24, D25]>;
148 def Q13 : ARMReg<13, "q13", [D26, D27]>;
149 def Q14 : ARMReg<14, "q14", [D28, D29]>;
150 def Q15 : ARMReg<15, "q15", [D30, D31]>;
153 // Current Program Status Register.
154 def CPSR : ARMReg<0, "cpsr">;
155 def APSR : ARMReg<1, "apsr">;
156 def SPSR : ARMReg<2, "spsr">;
157 def FPSCR : ARMReg<3, "fpscr">;
158 def ITSTATE : ARMReg<4, "itstate">;
160 // Special Registers - only available in privileged mode.
161 def FPSID : ARMReg<0, "fpsid">;
162 def FPEXC : ARMReg<8, "fpexc">;
166 // pc == Program Counter
167 // lr == Link Register
168 // sp == Stack Pointer
169 // r12 == ip (scratch)
170 // r7 == Frame Pointer (thumb-style backtraces)
171 // r9 == May be reserved as Thread Register
172 // r11 == Frame Pointer (arm-style backtraces)
173 // r10 == Stack Limit
175 def GPR : RegisterClass<"ARM", [i32], 32, (add (sequence "R%u", 0, 12),
177 // Allocate LR as the first CSR since it is always saved anyway.
178 // For Thumb1 mode, we don't want to allocate hi regs at all, as we don't
179 // know how to spill them. If we make our prologue/epilogue code smarter at
180 // some point, we can go back to using the above allocation orders for the
181 // Thumb1 instructions that know how to use hi regs.
182 let AltOrders = [(add LR, GPR), (trunc GPR, 8)];
183 let AltOrderSelect = [{
184 return 1 + MF.getTarget().getSubtarget<ARMSubtarget>().isThumb1Only();
188 // GPRs without the PC. Some ARM instructions do not allow the PC in
189 // certain operand slots, particularly as the destination. Primarily
190 // useful for disassembly.
191 def GPRnopc : RegisterClass<"ARM", [i32], 32, (sub GPR, PC)> {
192 let AltOrders = [(add LR, GPRnopc), (trunc GPRnopc, 8)];
193 let AltOrderSelect = [{
194 return 1 + MF.getTarget().getSubtarget<ARMSubtarget>().isThumb1Only();
198 // GPRsp - Only the SP is legal. Used by Thumb1 instructions that want the
199 // implied SP argument list.
200 // FIXME: It would be better to not use this at all and refactor the
201 // instructions to not have SP an an explicit argument. That makes
202 // frame index resolution a bit trickier, though.
203 def GPRsp : RegisterClass<"ARM", [i32], 32, (add SP)>;
205 // restricted GPR register class. Many Thumb2 instructions allow the full
206 // register range for operands, but have undefined behaviours when PC
207 // or SP (R13 or R15) are used. The ARM ISA refers to these operands
208 // via the BadReg() pseudo-code description.
209 def rGPR : RegisterClass<"ARM", [i32], 32, (sub GPR, SP, PC)> {
210 let AltOrders = [(add LR, rGPR), (trunc rGPR, 8)];
211 let AltOrderSelect = [{
212 return 1 + MF.getTarget().getSubtarget<ARMSubtarget>().isThumb1Only();
216 // Thumb registers are R0-R7 normally. Some instructions can still use
217 // the general GPR register class above (MOV, e.g.)
218 def tGPR : RegisterClass<"ARM", [i32], 32, (trunc GPR, 8)>;
220 // The high registers in thumb mode, R8-R15.
221 def hGPR : RegisterClass<"ARM", [i32], 32, (sub GPR, tGPR)>;
223 // For tail calls, we can't use callee-saved registers, as they are restored
224 // to the saved value before the tail call, which would clobber a call address.
225 // Note, getMinimalPhysRegClass(R0) returns tGPR because of the names of
226 // this class and the preceding one(!) This is what we want.
227 def tcGPR : RegisterClass<"ARM", [i32], 32, (add R0, R1, R2, R3, R9, R12)> {
228 let AltOrders = [(and tcGPR, tGPR)];
229 let AltOrderSelect = [{
230 return MF.getTarget().getSubtarget<ARMSubtarget>().isThumb1Only();
234 // Scalar single precision floating point register class..
235 def SPR : RegisterClass<"ARM", [f32], 32, (sequence "S%u", 0, 31)>;
237 // Subset of SPR which can be used as a source of NEON scalars for 16-bit
239 def SPR_8 : RegisterClass<"ARM", [f32], 32, (trunc SPR, 16)>;
241 // Scalar double precision floating point / generic 64-bit vector register
243 // ARM requires only word alignment for double. It's more performant if it
244 // is double-word alignment though.
245 def DPR : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
246 (sequence "D%u", 0, 31)> {
247 // Allocate non-VFP2 registers D16-D31 first.
248 let AltOrders = [(rotl DPR, 16)];
249 let AltOrderSelect = [{ return 1; }];
252 // Subset of DPR that are accessible with VFP2 (and so that also have
253 // 32-bit SPR subregs).
254 def DPR_VFP2 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
256 let SubRegClasses = [(SPR ssub_0, ssub_1)];
259 // Subset of DPR which can be used as a source of NEON scalars for 16-bit
261 def DPR_8 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64,
263 let SubRegClasses = [(SPR_8 ssub_0, ssub_1)];
266 // Generic 128-bit vector register class.
267 def QPR : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64], 128,
268 (sequence "Q%u", 0, 15)> {
269 let SubRegClasses = [(DPR dsub_0, dsub_1)];
270 // Allocate non-VFP2 aliases Q8-Q15 first.
271 let AltOrders = [(rotl QPR, 8)];
272 let AltOrderSelect = [{ return 1; }];
275 // Subset of QPR that have 32-bit SPR subregs.
276 def QPR_VFP2 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
277 128, (trunc QPR, 8)> {
278 let SubRegClasses = [(SPR ssub_0, ssub_1, ssub_2, ssub_3),
279 (DPR_VFP2 dsub_0, dsub_1)];
282 // Subset of QPR that have DPR_8 and SPR_8 subregs.
283 def QPR_8 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
284 128, (trunc QPR, 4)> {
285 let SubRegClasses = [(SPR_8 ssub_0, ssub_1, ssub_2, ssub_3),
286 (DPR_8 dsub_0, dsub_1)];
289 // Pseudo 256-bit registers to represent pairs of Q registers. These should
290 // never be present in the emitted code.
291 // These are used for NEON load / store instructions, e.g., vld4, vst3.
292 // NOTE: It's possible to define more QQ registers since technically the
293 // starting D register number doesn't have to be multiple of 4, e.g.,
294 // D1, D2, D3, D4 would be a legal quad, but that would make the subregister
296 def Tuples2Q : RegisterTuples<[qsub_0, qsub_1],
298 (decimate (shl QPR, 1), 2)]> {
299 let CompositeIndices = [(dsub_2 qsub_1, dsub_0), (dsub_3 qsub_1, dsub_1)];
302 // Pseudo 256-bit vector register class to model pairs of Q registers
303 // (4 consecutive D registers).
304 def QQPR : RegisterClass<"ARM", [v4i64], 256, (add Tuples2Q)> {
305 let SubRegClasses = [(DPR dsub_0, dsub_1, dsub_2, dsub_3),
306 (QPR qsub_0, qsub_1)];
307 // Allocate non-VFP2 aliases first.
308 let AltOrders = [(rotl QQPR, 4)];
309 let AltOrderSelect = [{ return 1; }];
312 // Pseudo 512-bit registers to represent four consecutive Q registers.
313 def Tuples2QQ : RegisterTuples<[qqsub_0, qqsub_1],
315 (decimate (shl QQPR, 1), 2)]> {
316 let CompositeIndices = [(qsub_2 qqsub_1, qsub_0), (qsub_3 qqsub_1, qsub_1),
317 (dsub_4 qqsub_1, dsub_0), (dsub_5 qqsub_1, dsub_1),
318 (dsub_6 qqsub_1, dsub_2), (dsub_7 qqsub_1, dsub_3)];
321 // Pseudo 512-bit vector register class to model 4 consecutive Q registers
322 // (8 consecutive D registers).
323 def QQQQPR : RegisterClass<"ARM", [v8i64], 256, (add Tuples2QQ)> {
324 let SubRegClasses = [(DPR dsub_0, dsub_1, dsub_2, dsub_3,
325 dsub_4, dsub_5, dsub_6, dsub_7),
326 (QPR qsub_0, qsub_1, qsub_2, qsub_3)];
327 // Allocate non-VFP2 aliases first.
328 let AltOrders = [(rotl QQQQPR, 2)];
329 let AltOrderSelect = [{ return 1; }];
332 // Condition code registers.
333 def CCR : RegisterClass<"ARM", [i32], 32, (add CPSR)> {
334 let CopyCost = -1; // Don't allow copying of status registers.
335 let isAllocatable = 0;