Add gfx950 mfma instructions to ROCDL dialect (#123361)
[llvm-project.git] / llvm / lib / Target / ARM / ARMRegisterBankInfo.cpp
blob0493fd9e317d5c3e512f5a3d135b2e37ed63dcfb
1 //===- ARMRegisterBankInfo.cpp -----------------------------------*- C++ -*-==//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file
9 /// This file implements the targeting of the RegisterBankInfo class for ARM.
10 /// \todo This should be generated by TableGen.
11 //===----------------------------------------------------------------------===//
13 #include "ARMRegisterBankInfo.h"
14 #include "ARMInstrInfo.h" // For the register classes
15 #include "ARMSubtarget.h"
16 #include "llvm/CodeGen/MachineRegisterInfo.h"
17 #include "llvm/CodeGen/RegisterBank.h"
18 #include "llvm/CodeGen/RegisterBankInfo.h"
19 #include "llvm/CodeGen/TargetRegisterInfo.h"
21 #define GET_TARGET_REGBANK_IMPL
22 #include "ARMGenRegisterBank.inc"
24 using namespace llvm;
26 // FIXME: TableGen this.
27 // If it grows too much and TableGen still isn't ready to do the job, extract it
28 // into an ARMGenRegisterBankInfo.def (similar to AArch64).
29 namespace llvm {
30 namespace ARM {
31 enum PartialMappingIdx {
32 PMI_GPR,
33 PMI_SPR,
34 PMI_DPR,
35 PMI_Min = PMI_GPR,
38 const RegisterBankInfo::PartialMapping PartMappings[]{
39 // GPR Partial Mapping
40 {0, 32, GPRRegBank},
41 // SPR Partial Mapping
42 {0, 32, FPRRegBank},
43 // DPR Partial Mapping
44 {0, 64, FPRRegBank},
47 #ifndef NDEBUG
48 static bool checkPartMapping(const RegisterBankInfo::PartialMapping &PM,
49 unsigned Start, unsigned Length,
50 unsigned RegBankID) {
51 return PM.StartIdx == Start && PM.Length == Length &&
52 PM.RegBank->getID() == RegBankID;
55 static void checkPartialMappings() {
56 assert(
57 checkPartMapping(PartMappings[PMI_GPR - PMI_Min], 0, 32, GPRRegBankID) &&
58 "Wrong mapping for GPR");
59 assert(
60 checkPartMapping(PartMappings[PMI_SPR - PMI_Min], 0, 32, FPRRegBankID) &&
61 "Wrong mapping for SPR");
62 assert(
63 checkPartMapping(PartMappings[PMI_DPR - PMI_Min], 0, 64, FPRRegBankID) &&
64 "Wrong mapping for DPR");
66 #endif
68 enum ValueMappingIdx {
69 InvalidIdx = 0,
70 GPR3OpsIdx = 1,
71 SPR3OpsIdx = 4,
72 DPR3OpsIdx = 7,
75 const RegisterBankInfo::ValueMapping ValueMappings[] = {
76 // invalid
77 {nullptr, 0},
78 // 3 ops in GPRs
79 {&PartMappings[PMI_GPR - PMI_Min], 1},
80 {&PartMappings[PMI_GPR - PMI_Min], 1},
81 {&PartMappings[PMI_GPR - PMI_Min], 1},
82 // 3 ops in SPRs
83 {&PartMappings[PMI_SPR - PMI_Min], 1},
84 {&PartMappings[PMI_SPR - PMI_Min], 1},
85 {&PartMappings[PMI_SPR - PMI_Min], 1},
86 // 3 ops in DPRs
87 {&PartMappings[PMI_DPR - PMI_Min], 1},
88 {&PartMappings[PMI_DPR - PMI_Min], 1},
89 {&PartMappings[PMI_DPR - PMI_Min], 1}};
91 #ifndef NDEBUG
92 static bool
93 checkValueMapping(const RegisterBankInfo::ValueMapping &VM,
94 const RegisterBankInfo::PartialMapping *BreakDown) {
95 return VM.NumBreakDowns == 1 && VM.BreakDown == BreakDown;
98 static void checkValueMappings() {
99 assert(checkValueMapping(ValueMappings[GPR3OpsIdx],
100 &PartMappings[PMI_GPR - PMI_Min]) &&
101 "Wrong value mapping for 3 GPR ops instruction");
102 assert(checkValueMapping(ValueMappings[GPR3OpsIdx + 1],
103 &PartMappings[PMI_GPR - PMI_Min]) &&
104 "Wrong value mapping for 3 GPR ops instruction");
105 assert(checkValueMapping(ValueMappings[GPR3OpsIdx + 2],
106 &PartMappings[PMI_GPR - PMI_Min]) &&
107 "Wrong value mapping for 3 GPR ops instruction");
109 assert(checkValueMapping(ValueMappings[SPR3OpsIdx],
110 &PartMappings[PMI_SPR - PMI_Min]) &&
111 "Wrong value mapping for 3 SPR ops instruction");
112 assert(checkValueMapping(ValueMappings[SPR3OpsIdx + 1],
113 &PartMappings[PMI_SPR - PMI_Min]) &&
114 "Wrong value mapping for 3 SPR ops instruction");
115 assert(checkValueMapping(ValueMappings[SPR3OpsIdx + 2],
116 &PartMappings[PMI_SPR - PMI_Min]) &&
117 "Wrong value mapping for 3 SPR ops instruction");
119 assert(checkValueMapping(ValueMappings[DPR3OpsIdx],
120 &PartMappings[PMI_DPR - PMI_Min]) &&
121 "Wrong value mapping for 3 DPR ops instruction");
122 assert(checkValueMapping(ValueMappings[DPR3OpsIdx + 1],
123 &PartMappings[PMI_DPR - PMI_Min]) &&
124 "Wrong value mapping for 3 DPR ops instruction");
125 assert(checkValueMapping(ValueMappings[DPR3OpsIdx + 2],
126 &PartMappings[PMI_DPR - PMI_Min]) &&
127 "Wrong value mapping for 3 DPR ops instruction");
129 #endif
130 } // end namespace arm
131 } // end namespace llvm
133 ARMRegisterBankInfo::ARMRegisterBankInfo(const TargetRegisterInfo &TRI) {
134 // We have only one set of register banks, whatever the subtarget
135 // is. Therefore, the initialization of the RegBanks table should be
136 // done only once. Indeed the table of all register banks
137 // (ARM::RegBanks) is unique in the compiler. At some point, it
138 // will get tablegen'ed and the whole constructor becomes empty.
139 static llvm::once_flag InitializeRegisterBankFlag;
141 static auto InitializeRegisterBankOnce = [&]() {
142 const RegisterBank &RBGPR = getRegBank(ARM::GPRRegBankID);
143 (void)RBGPR;
144 assert(&ARM::GPRRegBank == &RBGPR && "The order in RegBanks is messed up");
146 // Initialize the GPR bank.
147 assert(RBGPR.covers(*TRI.getRegClass(ARM::GPRRegClassID)) &&
148 "Subclass not added?");
149 assert(RBGPR.covers(*TRI.getRegClass(ARM::GPRwithAPSRRegClassID)) &&
150 "Subclass not added?");
151 assert(RBGPR.covers(*TRI.getRegClass(ARM::GPRnopcRegClassID)) &&
152 "Subclass not added?");
153 assert(RBGPR.covers(*TRI.getRegClass(ARM::rGPRRegClassID)) &&
154 "Subclass not added?");
155 assert(RBGPR.covers(*TRI.getRegClass(ARM::tGPRRegClassID)) &&
156 "Subclass not added?");
157 assert(RBGPR.covers(*TRI.getRegClass(ARM::tcGPRRegClassID)) &&
158 "Subclass not added?");
159 assert(RBGPR.covers(*TRI.getRegClass(ARM::tGPROdd_and_tcGPRRegClassID)) &&
160 "Subclass not added?");
161 assert(getMaximumSize(RBGPR.getID()) == 32 &&
162 "GPRs should hold up to 32-bit");
164 #ifndef NDEBUG
165 ARM::checkPartialMappings();
166 ARM::checkValueMappings();
167 #endif
170 llvm::call_once(InitializeRegisterBankFlag, InitializeRegisterBankOnce);
173 const RegisterBankInfo::InstructionMapping &
174 ARMRegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
175 auto Opc = MI.getOpcode();
177 // Try the default logic for non-generic instructions that are either copies
178 // or already have some operands assigned to banks.
179 if (!isPreISelGenericOpcode(Opc) || Opc == TargetOpcode::G_PHI) {
180 const InstructionMapping &Mapping = getInstrMappingImpl(MI);
181 if (Mapping.isValid())
182 return Mapping;
185 using namespace TargetOpcode;
187 const MachineFunction &MF = *MI.getParent()->getParent();
188 const MachineRegisterInfo &MRI = MF.getRegInfo();
189 unsigned NumOperands = MI.getNumOperands();
190 const ValueMapping *OperandsMapping = &ARM::ValueMappings[ARM::GPR3OpsIdx];
192 switch (Opc) {
193 case G_ADD:
194 case G_SUB: {
195 // Integer operations where the source and destination are in the
196 // same register class.
197 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
198 OperandsMapping = Ty.getSizeInBits() == 64
199 ? &ARM::ValueMappings[ARM::DPR3OpsIdx]
200 : &ARM::ValueMappings[ARM::GPR3OpsIdx];
201 break;
203 case G_MUL:
204 case G_AND:
205 case G_OR:
206 case G_XOR:
207 case G_LSHR:
208 case G_ASHR:
209 case G_SHL:
210 case G_SDIV:
211 case G_UDIV:
212 case G_SEXT:
213 case G_ZEXT:
214 case G_ANYEXT:
215 case G_PTR_ADD:
216 case G_INTTOPTR:
217 case G_PTRTOINT:
218 case G_CTLZ:
219 // FIXME: We're abusing the fact that everything lives in a GPR for now; in
220 // the real world we would use different mappings.
221 OperandsMapping = &ARM::ValueMappings[ARM::GPR3OpsIdx];
222 break;
223 case G_TRUNC: {
224 // In some cases we may end up with a G_TRUNC from a 64-bit value to a
225 // 32-bit value. This isn't a real floating point trunc (that would be a
226 // G_FPTRUNC). Instead it is an integer trunc in disguise, which can appear
227 // because the legalizer doesn't distinguish between integer and floating
228 // point values so it may leave some 64-bit integers un-narrowed. Until we
229 // have a more principled solution that doesn't let such things sneak all
230 // the way to this point, just map the source to a DPR and the destination
231 // to a GPR.
232 LLT LargeTy = MRI.getType(MI.getOperand(1).getReg());
233 OperandsMapping =
234 LargeTy.getSizeInBits() <= 32
235 ? &ARM::ValueMappings[ARM::GPR3OpsIdx]
236 : getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx],
237 &ARM::ValueMappings[ARM::DPR3OpsIdx]});
238 break;
240 case G_LOAD:
241 case G_STORE: {
242 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
243 OperandsMapping =
244 Ty.getSizeInBits() == 64
245 ? getOperandsMapping({&ARM::ValueMappings[ARM::DPR3OpsIdx],
246 &ARM::ValueMappings[ARM::GPR3OpsIdx]})
247 : &ARM::ValueMappings[ARM::GPR3OpsIdx];
248 break;
250 case G_FADD:
251 case G_FSUB:
252 case G_FMUL:
253 case G_FDIV:
254 case G_FNEG: {
255 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
256 OperandsMapping =Ty.getSizeInBits() == 64
257 ? &ARM::ValueMappings[ARM::DPR3OpsIdx]
258 : &ARM::ValueMappings[ARM::SPR3OpsIdx];
259 break;
261 case G_FMA: {
262 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
263 OperandsMapping =
264 Ty.getSizeInBits() == 64
265 ? getOperandsMapping({&ARM::ValueMappings[ARM::DPR3OpsIdx],
266 &ARM::ValueMappings[ARM::DPR3OpsIdx],
267 &ARM::ValueMappings[ARM::DPR3OpsIdx],
268 &ARM::ValueMappings[ARM::DPR3OpsIdx]})
269 : getOperandsMapping({&ARM::ValueMappings[ARM::SPR3OpsIdx],
270 &ARM::ValueMappings[ARM::SPR3OpsIdx],
271 &ARM::ValueMappings[ARM::SPR3OpsIdx],
272 &ARM::ValueMappings[ARM::SPR3OpsIdx]});
273 break;
275 case G_FPEXT: {
276 LLT ToTy = MRI.getType(MI.getOperand(0).getReg());
277 LLT FromTy = MRI.getType(MI.getOperand(1).getReg());
278 if (ToTy.getSizeInBits() == 64 && FromTy.getSizeInBits() == 32)
279 OperandsMapping =
280 getOperandsMapping({&ARM::ValueMappings[ARM::DPR3OpsIdx],
281 &ARM::ValueMappings[ARM::SPR3OpsIdx]});
282 break;
284 case G_FPTRUNC: {
285 LLT ToTy = MRI.getType(MI.getOperand(0).getReg());
286 LLT FromTy = MRI.getType(MI.getOperand(1).getReg());
287 if (ToTy.getSizeInBits() == 32 && FromTy.getSizeInBits() == 64)
288 OperandsMapping =
289 getOperandsMapping({&ARM::ValueMappings[ARM::SPR3OpsIdx],
290 &ARM::ValueMappings[ARM::DPR3OpsIdx]});
291 break;
293 case G_FPTOSI:
294 case G_FPTOUI: {
295 LLT ToTy = MRI.getType(MI.getOperand(0).getReg());
296 LLT FromTy = MRI.getType(MI.getOperand(1).getReg());
297 if ((FromTy.getSizeInBits() == 32 || FromTy.getSizeInBits() == 64) &&
298 ToTy.getSizeInBits() == 32)
299 OperandsMapping =
300 FromTy.getSizeInBits() == 64
301 ? getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx],
302 &ARM::ValueMappings[ARM::DPR3OpsIdx]})
303 : getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx],
304 &ARM::ValueMappings[ARM::SPR3OpsIdx]});
305 break;
307 case G_SITOFP:
308 case G_UITOFP: {
309 LLT ToTy = MRI.getType(MI.getOperand(0).getReg());
310 LLT FromTy = MRI.getType(MI.getOperand(1).getReg());
311 if (FromTy.getSizeInBits() == 32 &&
312 (ToTy.getSizeInBits() == 32 || ToTy.getSizeInBits() == 64))
313 OperandsMapping =
314 ToTy.getSizeInBits() == 64
315 ? getOperandsMapping({&ARM::ValueMappings[ARM::DPR3OpsIdx],
316 &ARM::ValueMappings[ARM::GPR3OpsIdx]})
317 : getOperandsMapping({&ARM::ValueMappings[ARM::SPR3OpsIdx],
318 &ARM::ValueMappings[ARM::GPR3OpsIdx]});
319 break;
321 case G_FCONSTANT: {
322 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
323 OperandsMapping = getOperandsMapping(
324 {Ty.getSizeInBits() == 64 ? &ARM::ValueMappings[ARM::DPR3OpsIdx]
325 : &ARM::ValueMappings[ARM::SPR3OpsIdx],
326 nullptr});
327 break;
329 case G_CONSTANT:
330 case G_FRAME_INDEX:
331 case G_GLOBAL_VALUE:
332 case G_CONSTANT_POOL:
333 OperandsMapping =
334 getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx], nullptr});
335 break;
336 case G_SELECT: {
337 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
338 (void)Ty;
339 LLT Ty2 = MRI.getType(MI.getOperand(1).getReg());
340 (void)Ty2;
341 assert(Ty.getSizeInBits() == 32 && "Unsupported size for G_SELECT");
342 assert(Ty2.getSizeInBits() == 1 && "Unsupported size for G_SELECT");
343 OperandsMapping =
344 getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx],
345 &ARM::ValueMappings[ARM::GPR3OpsIdx],
346 &ARM::ValueMappings[ARM::GPR3OpsIdx],
347 &ARM::ValueMappings[ARM::GPR3OpsIdx]});
348 break;
350 case G_ICMP: {
351 LLT Ty2 = MRI.getType(MI.getOperand(2).getReg());
352 (void)Ty2;
353 assert(Ty2.getSizeInBits() == 32 && "Unsupported size for G_ICMP");
354 OperandsMapping =
355 getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx], nullptr,
356 &ARM::ValueMappings[ARM::GPR3OpsIdx],
357 &ARM::ValueMappings[ARM::GPR3OpsIdx]});
358 break;
360 case G_FCMP: {
361 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
362 (void)Ty;
363 LLT Ty1 = MRI.getType(MI.getOperand(2).getReg());
364 LLT Ty2 = MRI.getType(MI.getOperand(3).getReg());
365 (void)Ty2;
366 assert(Ty.getSizeInBits() == 1 && "Unsupported size for G_FCMP");
367 assert(Ty1.getSizeInBits() == Ty2.getSizeInBits() &&
368 "Mismatched operand sizes for G_FCMP");
370 unsigned Size = Ty1.getSizeInBits();
371 assert((Size == 32 || Size == 64) && "Unsupported size for G_FCMP");
373 auto FPRValueMapping = Size == 32 ? &ARM::ValueMappings[ARM::SPR3OpsIdx]
374 : &ARM::ValueMappings[ARM::DPR3OpsIdx];
375 OperandsMapping =
376 getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx], nullptr,
377 FPRValueMapping, FPRValueMapping});
378 break;
380 case G_MERGE_VALUES: {
381 // We only support G_MERGE_VALUES for creating a double precision floating
382 // point value out of two GPRs.
383 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
384 LLT Ty1 = MRI.getType(MI.getOperand(1).getReg());
385 LLT Ty2 = MRI.getType(MI.getOperand(2).getReg());
386 if (Ty.getSizeInBits() != 64 || Ty1.getSizeInBits() != 32 ||
387 Ty2.getSizeInBits() != 32)
388 return getInvalidInstructionMapping();
389 OperandsMapping =
390 getOperandsMapping({&ARM::ValueMappings[ARM::DPR3OpsIdx],
391 &ARM::ValueMappings[ARM::GPR3OpsIdx],
392 &ARM::ValueMappings[ARM::GPR3OpsIdx]});
393 break;
395 case G_UNMERGE_VALUES: {
396 // We only support G_UNMERGE_VALUES for splitting a double precision
397 // floating point value into two GPRs.
398 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
399 LLT Ty1 = MRI.getType(MI.getOperand(1).getReg());
400 LLT Ty2 = MRI.getType(MI.getOperand(2).getReg());
401 if (Ty.getSizeInBits() != 32 || Ty1.getSizeInBits() != 32 ||
402 Ty2.getSizeInBits() != 64)
403 return getInvalidInstructionMapping();
404 OperandsMapping =
405 getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx],
406 &ARM::ValueMappings[ARM::GPR3OpsIdx],
407 &ARM::ValueMappings[ARM::DPR3OpsIdx]});
408 break;
410 case G_BR:
411 OperandsMapping = getOperandsMapping({nullptr});
412 break;
413 case G_BRCOND:
414 OperandsMapping =
415 getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx], nullptr});
416 break;
417 case DBG_VALUE: {
418 SmallVector<const ValueMapping *, 4> OperandBanks(NumOperands);
419 const MachineOperand &MaybeReg = MI.getOperand(0);
420 if (MaybeReg.isReg() && MaybeReg.getReg()) {
421 unsigned Size = MRI.getType(MaybeReg.getReg()).getSizeInBits();
422 if (Size > 32 && Size != 64)
423 return getInvalidInstructionMapping();
424 OperandBanks[0] = Size == 64 ? &ARM::ValueMappings[ARM::DPR3OpsIdx]
425 : &ARM::ValueMappings[ARM::GPR3OpsIdx];
427 OperandsMapping = getOperandsMapping(OperandBanks);
428 break;
430 case G_GET_FPENV:
431 case G_SET_FPENV:
432 case G_GET_FPMODE:
433 OperandsMapping =
434 getOperandsMapping({&ARM::ValueMappings[ARM::GPR3OpsIdx], nullptr});
435 break;
436 case G_RESET_FPENV:
437 OperandsMapping = getOperandsMapping({nullptr});
438 break;
439 default:
440 return getInvalidInstructionMapping();
443 #ifndef NDEBUG
444 for (unsigned i = 0; i < NumOperands; i++) {
445 for (const auto &Mapping : OperandsMapping[i]) {
446 assert(
447 (Mapping.RegBank->getID() != ARM::FPRRegBankID ||
448 MF.getSubtarget<ARMSubtarget>().hasVFP2Base()) &&
449 "Trying to use floating point register bank on target without vfp");
452 #endif
454 return getInstructionMapping(DefaultMappingID, /*Cost=*/1, OperandsMapping,
455 NumOperands);