1 //===-- ARMRegisterInfo.td - ARM Register defs -------------*- tablegen -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 include "ARMSystemRegister.td"
11 //===----------------------------------------------------------------------===//
12 // Declarations that describe the ARM register file
13 //===----------------------------------------------------------------------===//
15 // Registers are identified with 4-bit ID numbers.
16 class ARMReg<bits<16> Enc, string n, list<Register> subregs = [],
17 list<string> altNames = []> : Register<n, altNames> {
19 let Namespace = "ARM";
20 let SubRegs = subregs;
21 // All bits of ARM registers with sub-registers are covered by sub-registers.
22 let CoveredBySubRegs = 1;
25 class ARMFReg<bits<16> Enc, string n> : Register<n> {
27 let Namespace = "ARM";
30 let Namespace = "ARM",
31 FallbackRegAltNameIndex = NoRegAltName in {
32 def RegNamesRaw : RegAltNameIndex;
35 // Subregister indices.
36 let Namespace = "ARM" in {
37 def qqsub_0 : SubRegIndex<256>;
38 def qqsub_1 : SubRegIndex<256, 256>;
40 // Note: Code depends on these having consecutive numbers.
41 def qsub_0 : SubRegIndex<128>;
42 def qsub_1 : SubRegIndex<128, 128>;
43 def qsub_2 : ComposedSubRegIndex<qqsub_1, qsub_0>;
44 def qsub_3 : ComposedSubRegIndex<qqsub_1, qsub_1>;
46 def dsub_0 : SubRegIndex<64>;
47 def dsub_1 : SubRegIndex<64, 64>;
48 def dsub_2 : ComposedSubRegIndex<qsub_1, dsub_0>;
49 def dsub_3 : ComposedSubRegIndex<qsub_1, dsub_1>;
50 def dsub_4 : ComposedSubRegIndex<qsub_2, dsub_0>;
51 def dsub_5 : ComposedSubRegIndex<qsub_2, dsub_1>;
52 def dsub_6 : ComposedSubRegIndex<qsub_3, dsub_0>;
53 def dsub_7 : ComposedSubRegIndex<qsub_3, dsub_1>;
55 def ssub_0 : SubRegIndex<32>;
56 def ssub_1 : SubRegIndex<32, 32>;
57 def ssub_2 : ComposedSubRegIndex<dsub_1, ssub_0>;
58 def ssub_3 : ComposedSubRegIndex<dsub_1, ssub_1>;
59 def ssub_4 : ComposedSubRegIndex<dsub_2, ssub_0>;
60 def ssub_5 : ComposedSubRegIndex<dsub_2, ssub_1>;
61 def ssub_6 : ComposedSubRegIndex<dsub_3, ssub_0>;
62 def ssub_7 : ComposedSubRegIndex<dsub_3, ssub_1>;
63 def ssub_8 : ComposedSubRegIndex<dsub_4, ssub_0>;
64 def ssub_9 : ComposedSubRegIndex<dsub_4, ssub_1>;
65 def ssub_10 : ComposedSubRegIndex<dsub_5, ssub_0>;
66 def ssub_11 : ComposedSubRegIndex<dsub_5, ssub_1>;
67 def ssub_12 : ComposedSubRegIndex<dsub_6, ssub_0>;
68 def ssub_13 : ComposedSubRegIndex<dsub_6, ssub_1>;
70 def gsub_0 : SubRegIndex<32>;
71 def gsub_1 : SubRegIndex<32, 32>;
72 // Let TableGen synthesize the remaining 12 ssub_* indices.
73 // We don't need to name them.
77 def R0 : ARMReg< 0, "r0">, DwarfRegNum<[0]>;
78 def R1 : ARMReg< 1, "r1">, DwarfRegNum<[1]>;
79 def R2 : ARMReg< 2, "r2">, DwarfRegNum<[2]>;
80 def R3 : ARMReg< 3, "r3">, DwarfRegNum<[3]>;
81 def R4 : ARMReg< 4, "r4">, DwarfRegNum<[4]>;
82 def R5 : ARMReg< 5, "r5">, DwarfRegNum<[5]>;
83 def R6 : ARMReg< 6, "r6">, DwarfRegNum<[6]>;
84 def R7 : ARMReg< 7, "r7">, DwarfRegNum<[7]>;
85 // These require 32-bit instructions.
86 let CostPerUse = 1 in {
87 def R8 : ARMReg< 8, "r8">, DwarfRegNum<[8]>;
88 def R9 : ARMReg< 9, "r9">, DwarfRegNum<[9]>;
89 def R10 : ARMReg<10, "r10">, DwarfRegNum<[10]>;
90 def R11 : ARMReg<11, "r11">, DwarfRegNum<[11]>;
91 def R12 : ARMReg<12, "r12">, DwarfRegNum<[12]>;
92 let RegAltNameIndices = [RegNamesRaw] in {
93 def SP : ARMReg<13, "sp", [], ["r13"]>, DwarfRegNum<[13]>;
94 def LR : ARMReg<14, "lr", [], ["r14"]>, DwarfRegNum<[14]>;
95 def PC : ARMReg<15, "pc", [], ["r15"]>, DwarfRegNum<[15]>;
100 def S0 : ARMFReg< 0, "s0">; def S1 : ARMFReg< 1, "s1">;
101 def S2 : ARMFReg< 2, "s2">; def S3 : ARMFReg< 3, "s3">;
102 def S4 : ARMFReg< 4, "s4">; def S5 : ARMFReg< 5, "s5">;
103 def S6 : ARMFReg< 6, "s6">; def S7 : ARMFReg< 7, "s7">;
104 def S8 : ARMFReg< 8, "s8">; def S9 : ARMFReg< 9, "s9">;
105 def S10 : ARMFReg<10, "s10">; def S11 : ARMFReg<11, "s11">;
106 def S12 : ARMFReg<12, "s12">; def S13 : ARMFReg<13, "s13">;
107 def S14 : ARMFReg<14, "s14">; def S15 : ARMFReg<15, "s15">;
108 def S16 : ARMFReg<16, "s16">; def S17 : ARMFReg<17, "s17">;
109 def S18 : ARMFReg<18, "s18">; def S19 : ARMFReg<19, "s19">;
110 def S20 : ARMFReg<20, "s20">; def S21 : ARMFReg<21, "s21">;
111 def S22 : ARMFReg<22, "s22">; def S23 : ARMFReg<23, "s23">;
112 def S24 : ARMFReg<24, "s24">; def S25 : ARMFReg<25, "s25">;
113 def S26 : ARMFReg<26, "s26">; def S27 : ARMFReg<27, "s27">;
114 def S28 : ARMFReg<28, "s28">; def S29 : ARMFReg<29, "s29">;
115 def S30 : ARMFReg<30, "s30">; def S31 : ARMFReg<31, "s31">;
117 // Aliases of the F* registers used to hold 64-bit fp values (doubles)
118 let SubRegIndices = [ssub_0, ssub_1] in {
119 def D0 : ARMReg< 0, "d0", [S0, S1]>, DwarfRegNum<[256]>;
120 def D1 : ARMReg< 1, "d1", [S2, S3]>, DwarfRegNum<[257]>;
121 def D2 : ARMReg< 2, "d2", [S4, S5]>, DwarfRegNum<[258]>;
122 def D3 : ARMReg< 3, "d3", [S6, S7]>, DwarfRegNum<[259]>;
123 def D4 : ARMReg< 4, "d4", [S8, S9]>, DwarfRegNum<[260]>;
124 def D5 : ARMReg< 5, "d5", [S10, S11]>, DwarfRegNum<[261]>;
125 def D6 : ARMReg< 6, "d6", [S12, S13]>, DwarfRegNum<[262]>;
126 def D7 : ARMReg< 7, "d7", [S14, S15]>, DwarfRegNum<[263]>;
127 def D8 : ARMReg< 8, "d8", [S16, S17]>, DwarfRegNum<[264]>;
128 def D9 : ARMReg< 9, "d9", [S18, S19]>, DwarfRegNum<[265]>;
129 def D10 : ARMReg<10, "d10", [S20, S21]>, DwarfRegNum<[266]>;
130 def D11 : ARMReg<11, "d11", [S22, S23]>, DwarfRegNum<[267]>;
131 def D12 : ARMReg<12, "d12", [S24, S25]>, DwarfRegNum<[268]>;
132 def D13 : ARMReg<13, "d13", [S26, S27]>, DwarfRegNum<[269]>;
133 def D14 : ARMReg<14, "d14", [S28, S29]>, DwarfRegNum<[270]>;
134 def D15 : ARMReg<15, "d15", [S30, S31]>, DwarfRegNum<[271]>;
137 // VFP3 defines 16 additional double registers
138 def D16 : ARMFReg<16, "d16">, DwarfRegNum<[272]>;
139 def D17 : ARMFReg<17, "d17">, DwarfRegNum<[273]>;
140 def D18 : ARMFReg<18, "d18">, DwarfRegNum<[274]>;
141 def D19 : ARMFReg<19, "d19">, DwarfRegNum<[275]>;
142 def D20 : ARMFReg<20, "d20">, DwarfRegNum<[276]>;
143 def D21 : ARMFReg<21, "d21">, DwarfRegNum<[277]>;
144 def D22 : ARMFReg<22, "d22">, DwarfRegNum<[278]>;
145 def D23 : ARMFReg<23, "d23">, DwarfRegNum<[279]>;
146 def D24 : ARMFReg<24, "d24">, DwarfRegNum<[280]>;
147 def D25 : ARMFReg<25, "d25">, DwarfRegNum<[281]>;
148 def D26 : ARMFReg<26, "d26">, DwarfRegNum<[282]>;
149 def D27 : ARMFReg<27, "d27">, DwarfRegNum<[283]>;
150 def D28 : ARMFReg<28, "d28">, DwarfRegNum<[284]>;
151 def D29 : ARMFReg<29, "d29">, DwarfRegNum<[285]>;
152 def D30 : ARMFReg<30, "d30">, DwarfRegNum<[286]>;
153 def D31 : ARMFReg<31, "d31">, DwarfRegNum<[287]>;
155 // Advanced SIMD (NEON) defines 16 quad-word aliases
156 let SubRegIndices = [dsub_0, dsub_1] in {
157 def Q0 : ARMReg< 0, "q0", [D0, D1]>;
158 def Q1 : ARMReg< 1, "q1", [D2, D3]>;
159 def Q2 : ARMReg< 2, "q2", [D4, D5]>;
160 def Q3 : ARMReg< 3, "q3", [D6, D7]>;
161 def Q4 : ARMReg< 4, "q4", [D8, D9]>;
162 def Q5 : ARMReg< 5, "q5", [D10, D11]>;
163 def Q6 : ARMReg< 6, "q6", [D12, D13]>;
164 def Q7 : ARMReg< 7, "q7", [D14, D15]>;
166 let SubRegIndices = [dsub_0, dsub_1] in {
167 def Q8 : ARMReg< 8, "q8", [D16, D17]>;
168 def Q9 : ARMReg< 9, "q9", [D18, D19]>;
169 def Q10 : ARMReg<10, "q10", [D20, D21]>;
170 def Q11 : ARMReg<11, "q11", [D22, D23]>;
171 def Q12 : ARMReg<12, "q12", [D24, D25]>;
172 def Q13 : ARMReg<13, "q13", [D26, D27]>;
173 def Q14 : ARMReg<14, "q14", [D28, D29]>;
174 def Q15 : ARMReg<15, "q15", [D30, D31]>;
177 // Current Program Status Register.
178 // We model fpscr with two registers: FPSCR models the control bits and will be
179 // reserved. FPSCR_NZCV models the flag bits and will be unreserved. APSR_NZCV
180 // models the APSR when it's accessed by some special instructions. In such cases
181 // it has the same encoding as PC.
182 def CPSR : ARMReg<0, "cpsr">;
183 def APSR : ARMReg<1, "apsr">;
184 def APSR_NZCV : ARMReg<15, "apsr_nzcv">;
185 def SPSR : ARMReg<2, "spsr">;
186 def FPSCR : ARMReg<3, "fpscr">;
187 def FPSCR_NZCV : ARMReg<3, "fpscr_nzcv"> {
188 let Aliases = [FPSCR];
190 def ITSTATE : ARMReg<4, "itstate">;
192 // Special Registers - only available in privileged mode.
193 def FPSID : ARMReg<0, "fpsid">;
194 def MVFR2 : ARMReg<5, "mvfr2">;
195 def MVFR1 : ARMReg<6, "mvfr1">;
196 def MVFR0 : ARMReg<7, "mvfr0">;
197 def FPEXC : ARMReg<8, "fpexc">;
198 def FPINST : ARMReg<9, "fpinst">;
199 def FPINST2 : ARMReg<10, "fpinst2">;
200 // These encodings aren't actual instruction encodings, their encoding depends
201 // on the instruction they are used in and for VPR 32 was chosen such that it
202 // always comes last in spr_reglist_with_vpr.
203 def VPR : ARMReg<32, "vpr">;
205 : ARMReg<2, "fpscr_nzcvqc">;
206 def P0 : ARMReg<13, "p0">;
207 def FPCXTNS : ARMReg<14, "fpcxtns">;
208 def FPCXTS : ARMReg<15, "fpcxts">;
210 def ZR : ARMReg<15, "zr">, DwarfRegNum<[15]>;
214 // pc == Program Counter
215 // lr == Link Register
216 // sp == Stack Pointer
217 // r12 == ip (scratch)
218 // r7 == Frame Pointer (thumb-style backtraces)
219 // r9 == May be reserved as Thread Register
220 // r11 == Frame Pointer (arm-style backtraces)
221 // r10 == Stack Limit
223 def GPR : RegisterClass<"ARM", [i32], 32, (add (sequence "R%u", 0, 12),
225 // Allocate LR as the first CSR since it is always saved anyway.
226 // For Thumb1 mode, we don't want to allocate hi regs at all, as we don't
227 // know how to spill them. If we make our prologue/epilogue code smarter at
228 // some point, we can go back to using the above allocation orders for the
229 // Thumb1 instructions that know how to use hi regs.
230 let AltOrders = [(add LR, GPR), (trunc GPR, 8),
231 (add (trunc GPR, 8), R12, LR, (shl GPR, 8))];
232 let AltOrderSelect = [{
233 return MF.getSubtarget<ARMSubtarget>().getGPRAllocationOrder(MF);
235 let DiagnosticString = "operand must be a register in range [r0, r15]";
238 // GPRs without the PC. Some ARM instructions do not allow the PC in
239 // certain operand slots, particularly as the destination. Primarily
240 // useful for disassembly.
241 def GPRnopc : RegisterClass<"ARM", [i32], 32, (sub GPR, PC)> {
242 let AltOrders = [(add LR, GPRnopc), (trunc GPRnopc, 8),
243 (add (trunc GPRnopc, 8), R12, LR, (shl GPRnopc, 8))];
244 let AltOrderSelect = [{
245 return MF.getSubtarget<ARMSubtarget>().getGPRAllocationOrder(MF);
247 let DiagnosticString = "operand must be a register in range [r0, r14]";
250 // GPRs without the PC but with APSR. Some instructions allow accessing the
251 // APSR, while actually encoding PC in the register field. This is useful
252 // for assembly and disassembly only.
253 def GPRwithAPSR : RegisterClass<"ARM", [i32], 32, (add (sub GPR, PC), APSR_NZCV)> {
254 let AltOrders = [(add LR, GPRnopc), (trunc GPRnopc, 8)];
255 let AltOrderSelect = [{
256 return 1 + MF.getSubtarget<ARMSubtarget>().isThumb1Only();
258 let DiagnosticString = "operand must be a register in range [r0, r14] or apsr_nzcv";
261 // GPRs without the PC and SP registers but with APSR. Used by CLRM instruction.
262 def GPRwithAPSRnosp : RegisterClass<"ARM", [i32], 32, (add (sequence "R%u", 0, 12), LR, APSR)> {
263 let isAllocatable = 0;
266 def GPRwithZR : RegisterClass<"ARM", [i32], 32, (add (sub GPR, PC), ZR)> {
267 let AltOrders = [(add LR, GPRwithZR), (trunc GPRwithZR, 8)];
268 let AltOrderSelect = [{
269 return 1 + MF.getSubtarget<ARMSubtarget>().isThumb1Only();
271 let DiagnosticString = "operand must be a register in range [r0, r14] or zr";
274 def GPRwithZRnosp : RegisterClass<"ARM", [i32], 32, (sub GPRwithZR, SP)> {
275 let AltOrders = [(add LR, GPRwithZRnosp), (trunc GPRwithZRnosp, 8)];
276 let AltOrderSelect = [{
277 return 1 + MF.getSubtarget<ARMSubtarget>().isThumb1Only();
279 let DiagnosticString = "operand must be a register in range [r0, r12] or r14 or zr";
282 // GPRsp - Only the SP is legal. Used by Thumb1 instructions that want the
283 // implied SP argument list.
284 // FIXME: It would be better to not use this at all and refactor the
285 // instructions to not have SP an an explicit argument. That makes
286 // frame index resolution a bit trickier, though.
287 def GPRsp : RegisterClass<"ARM", [i32], 32, (add SP)> {
288 let DiagnosticString = "operand must be a register sp";
291 // GPRlr - Only LR is legal. Used by ARMv8.1-M Low Overhead Loop instructions
292 // where LR is the only legal loop counter register.
293 def GPRlr : RegisterClass<"ARM", [i32], 32, (add LR)>;
295 // restricted GPR register class. Many Thumb2 instructions allow the full
296 // register range for operands, but have undefined behaviours when PC
297 // or SP (R13 or R15) are used. The ARM ISA refers to these operands
298 // via the BadReg() pseudo-code description.
299 def rGPR : RegisterClass<"ARM", [i32], 32, (sub GPR, SP, PC)> {
300 let AltOrders = [(add LR, rGPR), (trunc rGPR, 8),
301 (add (trunc rGPR, 8), R12, LR, (shl rGPR, 8))];
302 let AltOrderSelect = [{
303 return MF.getSubtarget<ARMSubtarget>().getGPRAllocationOrder(MF);
305 let DiagnosticType = "rGPR";
308 // Thumb registers are R0-R7 normally. Some instructions can still use
309 // the general GPR register class above (MOV, e.g.)
310 def tGPR : RegisterClass<"ARM", [i32], 32, (trunc GPR, 8)> {
311 let DiagnosticString = "operand must be a register in range [r0, r7]";
314 // Thumb registers R0-R7 and the PC. Some instructions like TBB or THH allow
315 // the PC to be used as a destination operand as well.
316 def tGPRwithpc : RegisterClass<"ARM", [i32], 32, (add tGPR, PC)>;
318 // The high registers in thumb mode, R8-R15.
319 def hGPR : RegisterClass<"ARM", [i32], 32, (sub GPR, tGPR)> {
320 let DiagnosticString = "operand must be a register in range [r8, r15]";
323 // For tail calls, we can't use callee-saved registers, as they are restored
324 // to the saved value before the tail call, which would clobber a call address.
325 // Note, getMinimalPhysRegClass(R0) returns tGPR because of the names of
326 // this class and the preceding one(!) This is what we want.
327 def tcGPR : RegisterClass<"ARM", [i32], 32, (add R0, R1, R2, R3, R12)> {
328 let AltOrders = [(and tcGPR, tGPR)];
329 let AltOrderSelect = [{
330 return MF.getSubtarget<ARMSubtarget>().isThumb1Only();
334 def tGPROdd : RegisterClass<"ARM", [i32], 32, (add R1, R3, R5, R7, R9, R11)> {
335 let AltOrders = [(and tGPROdd, tGPR)];
336 let AltOrderSelect = [{
337 return MF.getSubtarget<ARMSubtarget>().isThumb1Only();
339 let DiagnosticString =
340 "operand must be an odd-numbered register in range [r1,r11]";
343 def tGPREven : RegisterClass<"ARM", [i32], 32, (add R0, R2, R4, R6, R8, R10, R12, LR)> {
344 let AltOrders = [(and tGPREven, tGPR)];
345 let AltOrderSelect = [{
346 return MF.getSubtarget<ARMSubtarget>().isThumb1Only();
348 let DiagnosticString = "operand must be an even-numbered register";
351 // Condition code registers.
352 def CCR : RegisterClass<"ARM", [i32], 32, (add CPSR)> {
353 let CopyCost = -1; // Don't allow copying of status registers.
354 let isAllocatable = 0;
357 // MVE Condition code register.
358 def VCCR : RegisterClass<"ARM", [i32, v16i1, v8i1, v4i1], 32, (add VPR)> {
359 // let CopyCost = -1; // Don't allow copying of status registers.
362 // FPSCR, when the flags at the top of it are used as the input or
363 // output to an instruction such as MVE VADC.
364 def cl_FPSCR_NZCV : RegisterClass<"ARM", [i32], 32, (add FPSCR_NZCV)>;
366 // Scalar single precision floating point register class..
367 // FIXME: Allocation order changed to s0, s2, ... or s0, s4, ... as a quick hack
368 // to avoid partial-write dependencies on D or Q (depending on platform)
369 // registers (S registers are renamed as portions of D/Q registers).
370 def SPR : RegisterClass<"ARM", [f32], 32, (sequence "S%u", 0, 31)> {
371 let AltOrders = [(add (decimate SPR, 2), SPR),
372 (add (decimate SPR, 4),
374 (decimate (rotl SPR, 1), 4),
375 (decimate (rotl SPR, 1), 2))];
376 let AltOrderSelect = [{
377 return 1 + MF.getSubtarget<ARMSubtarget>().useStride4VFPs();
379 let DiagnosticString = "operand must be a register in range [s0, s31]";
382 def HPR : RegisterClass<"ARM", [f16], 32, (sequence "S%u", 0, 31)> {
383 let AltOrders = [(add (decimate HPR, 2), SPR),
384 (add (decimate HPR, 4),
386 (decimate (rotl HPR, 1), 4),
387 (decimate (rotl HPR, 1), 2))];
388 let AltOrderSelect = [{
389 return 1 + MF.getSubtarget<ARMSubtarget>().useStride4VFPs();
391 let DiagnosticString = "operand must be a register in range [s0, s31]";
394 // Subset of SPR which can be used as a source of NEON scalars for 16-bit
396 def SPR_8 : RegisterClass<"ARM", [f32], 32, (sequence "S%u", 0, 15)> {
397 let DiagnosticString = "operand must be a register in range [s0, s15]";
400 // Scalar double precision floating point / generic 64-bit vector register
402 // ARM requires only word alignment for double. It's more performant if it
403 // is double-word alignment though.
404 def DPR : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32, v4f16], 64,
405 (sequence "D%u", 0, 31)> {
406 // Allocate non-VFP2 registers D16-D31 first, and prefer even registers on
408 let AltOrders = [(rotl DPR, 16),
409 (add (decimate (rotl DPR, 16), 2), (rotl DPR, 16))];
410 let AltOrderSelect = [{
411 return 1 + MF.getSubtarget<ARMSubtarget>().useStride4VFPs();
413 let DiagnosticType = "DPR";
416 // Scalar single and double precision floating point and VPR register class,
417 // this is only used for parsing, don't use it anywhere else as the size and
418 // types don't match!
419 def FPWithVPR : RegisterClass<"ARM", [f32], 32, (add SPR, DPR, VPR)> {
420 let isAllocatable = 0;
423 // Subset of DPR that are accessible with VFP2 (and so that also have
424 // 32-bit SPR subregs).
425 def DPR_VFP2 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32, v4f16], 64,
427 let DiagnosticString = "operand must be a register in range [d0, d15]";
430 // Subset of DPR which can be used as a source of NEON scalars for 16-bit
432 def DPR_8 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32, v4f16], 64,
434 let DiagnosticString = "operand must be a register in range [d0, d7]";
437 // Generic 128-bit vector register class.
438 def QPR : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64, v8f16], 128,
439 (sequence "Q%u", 0, 15)> {
440 // Allocate non-VFP2 aliases Q8-Q15 first.
441 let AltOrders = [(rotl QPR, 8), (trunc QPR, 8)];
442 let AltOrderSelect = [{
443 return 1 + MF.getSubtarget<ARMSubtarget>().hasMVEIntegerOps();
445 let DiagnosticString = "operand must be a register in range [q0, q15]";
448 // Subset of QPR that have 32-bit SPR subregs.
449 def QPR_VFP2 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
450 128, (trunc QPR, 8)> {
451 let DiagnosticString = "operand must be a register in range [q0, q7]";
454 // Subset of QPR that have DPR_8 and SPR_8 subregs.
455 def QPR_8 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
456 128, (trunc QPR, 4)> {
457 let DiagnosticString = "operand must be a register in range [q0, q3]";
460 // MVE 128-bit vector register class. This class is only really needed for
461 // parsing assembly, since we still have to truncate the register set in the QPR
463 def MQPR : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64, v8f16],
464 128, (trunc QPR, 8)>;
466 // Pseudo-registers representing odd-even pairs of D registers. The even-odd
467 // pairs are already represented by the Q registers.
468 // These are needed by NEON instructions requiring two consecutive D registers.
469 // There is no D31_D0 register as that is always an UNPREDICTABLE encoding.
470 def TuplesOE2D : RegisterTuples<[dsub_0, dsub_1],
471 [(decimate (shl DPR, 1), 2),
472 (decimate (shl DPR, 2), 2)]>;
474 // Register class representing a pair of consecutive D registers.
475 // Use the Q registers for the even-odd pairs.
476 def DPair : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
477 128, (interleave QPR, TuplesOE2D)> {
478 // Allocate starting at non-VFP2 registers D16-D31 first.
479 // Prefer even-odd pairs as they are easier to copy.
480 let AltOrders = [(add (rotl QPR, 8), (rotl DPair, 16)),
481 (add (trunc QPR, 8), (trunc DPair, 16))];
482 let AltOrderSelect = [{
483 return 1 + MF.getSubtarget<ARMSubtarget>().hasMVEIntegerOps();
487 // Pseudo-registers representing even-odd pairs of GPRs from R1 to R13/SP.
488 // These are needed by instructions (e.g. ldrexd/strexd) requiring even-odd GPRs.
489 def Tuples2R : RegisterTuples<[gsub_0, gsub_1],
490 [(add R0, R2, R4, R6, R8, R10, R12),
491 (add R1, R3, R5, R7, R9, R11, SP)]>;
493 // Register class representing a pair of even-odd GPRs.
494 def GPRPair : RegisterClass<"ARM", [untyped], 64, (add Tuples2R)> {
495 let Size = 64; // 2 x 32 bits, we have no predefined type of that size.
498 // Pseudo-registers representing 3 consecutive D registers.
499 def Tuples3D : RegisterTuples<[dsub_0, dsub_1, dsub_2],
504 // 3 consecutive D registers.
505 def DTriple : RegisterClass<"ARM", [untyped], 64, (add Tuples3D)> {
506 let Size = 192; // 3 x 64 bits, we have no predefined type of that size.
509 // Pseudo 256-bit registers to represent pairs of Q registers. These should
510 // never be present in the emitted code.
511 // These are used for NEON load / store instructions, e.g., vld4, vst3.
512 def Tuples2Q : RegisterTuples<[qsub_0, qsub_1], [(shl QPR, 0), (shl QPR, 1)]>;
514 // Pseudo 256-bit vector register class to model pairs of Q registers
515 // (4 consecutive D registers).
516 def QQPR : RegisterClass<"ARM", [v4i64], 256, (add Tuples2Q)> {
517 // Allocate non-VFP2 aliases first.
518 let AltOrders = [(rotl QQPR, 8)];
519 let AltOrderSelect = [{ return 1; }];
522 // Tuples of 4 D regs that isn't also a pair of Q regs.
523 def TuplesOE4D : RegisterTuples<[dsub_0, dsub_1, dsub_2, dsub_3],
524 [(decimate (shl DPR, 1), 2),
525 (decimate (shl DPR, 2), 2),
526 (decimate (shl DPR, 3), 2),
527 (decimate (shl DPR, 4), 2)]>;
529 // 4 consecutive D registers.
530 def DQuad : RegisterClass<"ARM", [v4i64], 256,
531 (interleave Tuples2Q, TuplesOE4D)>;
533 // Pseudo 512-bit registers to represent four consecutive Q registers.
534 def Tuples2QQ : RegisterTuples<[qqsub_0, qqsub_1],
535 [(shl QQPR, 0), (shl QQPR, 2)]>;
537 // Pseudo 512-bit vector register class to model 4 consecutive Q registers
538 // (8 consecutive D registers).
539 def QQQQPR : RegisterClass<"ARM", [v8i64], 256, (add Tuples2QQ)> {
540 // Allocate non-VFP2 aliases first.
541 let AltOrders = [(rotl QQQQPR, 8)];
542 let AltOrderSelect = [{ return 1; }];
546 // Pseudo-registers representing 2-spaced consecutive D registers.
547 def Tuples2DSpc : RegisterTuples<[dsub_0, dsub_2],
551 // Spaced pairs of D registers.
552 def DPairSpc : RegisterClass<"ARM", [v2i64], 64, (add Tuples2DSpc)>;
554 def Tuples3DSpc : RegisterTuples<[dsub_0, dsub_2, dsub_4],
559 // Spaced triples of D registers.
560 def DTripleSpc : RegisterClass<"ARM", [untyped], 64, (add Tuples3DSpc)> {
561 let Size = 192; // 3 x 64 bits, we have no predefined type of that size.
564 def Tuples4DSpc : RegisterTuples<[dsub_0, dsub_2, dsub_4, dsub_6],
570 // Spaced quads of D registers.
571 def DQuadSpc : RegisterClass<"ARM", [v4i64], 64, (add Tuples3DSpc)>;