LLVM 22.0.0git
Utils.cpp
Go to the documentation of this file.
1//===- llvm/CodeGen/GlobalISel/Utils.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 This file implements the utility functions used by the GlobalISel
9/// pipeline.
10//===----------------------------------------------------------------------===//
11
13#include "llvm/ADT/APFloat.h"
14#include "llvm/ADT/APInt.h"
35#include "llvm/IR/Constants.h"
38#include <numeric>
39#include <optional>
40
41#define DEBUG_TYPE "globalisel-utils"
42
43using namespace llvm;
44using namespace MIPatternMatch;
45
47 const TargetInstrInfo &TII,
48 const RegisterBankInfo &RBI, Register Reg,
49 const TargetRegisterClass &RegClass) {
50 if (!RBI.constrainGenericRegister(Reg, RegClass, MRI))
51 return MRI.createVirtualRegister(&RegClass);
52
53 return Reg;
54}
55
57 const MachineFunction &MF, const TargetRegisterInfo &TRI,
59 const RegisterBankInfo &RBI, MachineInstr &InsertPt,
60 const TargetRegisterClass &RegClass, MachineOperand &RegMO) {
61 Register Reg = RegMO.getReg();
62 // Assume physical registers are properly constrained.
63 assert(Reg.isVirtual() && "PhysReg not implemented");
64
65 // Save the old register class to check whether
66 // the change notifications will be required.
67 // TODO: A better approach would be to pass
68 // the observers to constrainRegToClass().
69 auto *OldRegClass = MRI.getRegClassOrNull(Reg);
70 Register ConstrainedReg = constrainRegToClass(MRI, TII, RBI, Reg, RegClass);
71 // If we created a new virtual register because the class is not compatible
72 // then create a copy between the new and the old register.
73 if (ConstrainedReg != Reg) {
74 MachineBasicBlock::iterator InsertIt(&InsertPt);
75 MachineBasicBlock &MBB = *InsertPt.getParent();
76 // FIXME: The copy needs to have the classes constrained for its operands.
77 // Use operand's regbank to get the class for old register (Reg).
78 if (RegMO.isUse()) {
79 BuildMI(MBB, InsertIt, InsertPt.getDebugLoc(),
80 TII.get(TargetOpcode::COPY), ConstrainedReg)
81 .addReg(Reg);
82 } else {
83 assert(RegMO.isDef() && "Must be a definition");
84 BuildMI(MBB, std::next(InsertIt), InsertPt.getDebugLoc(),
85 TII.get(TargetOpcode::COPY), Reg)
86 .addReg(ConstrainedReg);
87 }
88 if (GISelChangeObserver *Observer = MF.getObserver()) {
89 Observer->changingInstr(*RegMO.getParent());
90 }
91 RegMO.setReg(ConstrainedReg);
92 if (GISelChangeObserver *Observer = MF.getObserver()) {
93 Observer->changedInstr(*RegMO.getParent());
94 }
95 } else if (OldRegClass != MRI.getRegClassOrNull(Reg)) {
96 if (GISelChangeObserver *Observer = MF.getObserver()) {
97 if (!RegMO.isDef()) {
98 MachineInstr *RegDef = MRI.getVRegDef(Reg);
99 Observer->changedInstr(*RegDef);
100 }
101 Observer->changingAllUsesOfReg(MRI, Reg);
102 Observer->finishedChangingAllUsesOfReg();
103 }
104 }
105 return ConstrainedReg;
106}
107
109 const MachineFunction &MF, const TargetRegisterInfo &TRI,
111 const RegisterBankInfo &RBI, MachineInstr &InsertPt, const MCInstrDesc &II,
112 MachineOperand &RegMO, unsigned OpIdx) {
113 Register Reg = RegMO.getReg();
114 // Assume physical registers are properly constrained.
115 assert(Reg.isVirtual() && "PhysReg not implemented");
116
117 const TargetRegisterClass *OpRC = TII.getRegClass(II, OpIdx, &TRI, MF);
118 // Some of the target independent instructions, like COPY, may not impose any
119 // register class constraints on some of their operands: If it's a use, we can
120 // skip constraining as the instruction defining the register would constrain
121 // it.
122
123 if (OpRC) {
124 // Obtain the RC from incoming regbank if it is a proper sub-class. Operands
125 // can have multiple regbanks for a superclass that combine different
126 // register types (E.g., AMDGPU's VGPR and AGPR). The regbank ambiguity
127 // resolved by targets during regbankselect should not be overridden.
128 if (const auto *SubRC = TRI.getCommonSubClass(
129 OpRC, TRI.getConstrainedRegClassForOperand(RegMO, MRI)))
130 OpRC = SubRC;
131
132 OpRC = TRI.getAllocatableClass(OpRC);
133 }
134
135 if (!OpRC) {
136 assert((!isTargetSpecificOpcode(II.getOpcode()) || RegMO.isUse()) &&
137 "Register class constraint is required unless either the "
138 "instruction is target independent or the operand is a use");
139 // FIXME: Just bailing out like this here could be not enough, unless we
140 // expect the users of this function to do the right thing for PHIs and
141 // COPY:
142 // v1 = COPY v0
143 // v2 = COPY v1
144 // v1 here may end up not being constrained at all. Please notice that to
145 // reproduce the issue we likely need a destination pattern of a selection
146 // rule producing such extra copies, not just an input GMIR with them as
147 // every existing target using selectImpl handles copies before calling it
148 // and they never reach this function.
149 return Reg;
150 }
151 return constrainOperandRegClass(MF, TRI, MRI, TII, RBI, InsertPt, *OpRC,
152 RegMO);
153}
154
156 const TargetInstrInfo &TII,
157 const TargetRegisterInfo &TRI,
158 const RegisterBankInfo &RBI) {
159 assert(!isPreISelGenericOpcode(I.getOpcode()) &&
160 "A selected instruction is expected");
161 MachineBasicBlock &MBB = *I.getParent();
164
165 for (unsigned OpI = 0, OpE = I.getNumExplicitOperands(); OpI != OpE; ++OpI) {
166 MachineOperand &MO = I.getOperand(OpI);
167
168 // There's nothing to be done on non-register operands.
169 if (!MO.isReg())
170 continue;
171
172 LLVM_DEBUG(dbgs() << "Converting operand: " << MO << '\n');
173 assert(MO.isReg() && "Unsupported non-reg operand");
174
175 Register Reg = MO.getReg();
176 // Physical registers don't need to be constrained.
177 if (Reg.isPhysical())
178 continue;
179
180 // Register operands with a value of 0 (e.g. predicate operands) don't need
181 // to be constrained.
182 if (Reg == 0)
183 continue;
184
185 // If the operand is a vreg, we should constrain its regclass, and only
186 // insert COPYs if that's impossible.
187 // constrainOperandRegClass does that for us.
188 constrainOperandRegClass(MF, TRI, MRI, TII, RBI, I, I.getDesc(), MO, OpI);
189
190 // Tie uses to defs as indicated in MCInstrDesc if this hasn't already been
191 // done.
192 if (MO.isUse()) {
193 int DefIdx = I.getDesc().getOperandConstraint(OpI, MCOI::TIED_TO);
194 if (DefIdx != -1 && !I.isRegTiedToUseOperand(DefIdx))
195 I.tieOperands(DefIdx, OpI);
196 }
197 }
198 return true;
199}
200
203 // Give up if either DstReg or SrcReg is a physical register.
204 if (DstReg.isPhysical() || SrcReg.isPhysical())
205 return false;
206 // Give up if the types don't match.
207 if (MRI.getType(DstReg) != MRI.getType(SrcReg))
208 return false;
209 // Replace if either DstReg has no constraints or the register
210 // constraints match.
211 const auto &DstRBC = MRI.getRegClassOrRegBank(DstReg);
212 if (!DstRBC || DstRBC == MRI.getRegClassOrRegBank(SrcReg))
213 return true;
214
215 // Otherwise match if the Src is already a regclass that is covered by the Dst
216 // RegBank.
217 return isa<const RegisterBank *>(DstRBC) && MRI.getRegClassOrNull(SrcReg) &&
218 cast<const RegisterBank *>(DstRBC)->covers(
219 *MRI.getRegClassOrNull(SrcReg));
220}
221
223 const MachineRegisterInfo &MRI) {
224 // Instructions without side-effects are dead iff they only define dead regs.
225 // This function is hot and this loop returns early in the common case,
226 // so only perform additional checks before this if absolutely necessary.
227 for (const auto &MO : MI.all_defs()) {
228 Register Reg = MO.getReg();
229 if (Reg.isPhysical() || !MRI.use_nodbg_empty(Reg))
230 return false;
231 }
232 return MI.wouldBeTriviallyDead();
233}
234
236 MachineFunction &MF,
237 const TargetPassConfig &TPC,
240 bool IsFatal = Severity == DS_Error &&
242 // Print the function name explicitly if we don't have a debug location (which
243 // makes the diagnostic less useful) or if we're going to emit a raw error.
244 if (!R.getLocation().isValid() || IsFatal)
245 R << (" (in function: " + MF.getName() + ")").str();
246
247 if (IsFatal)
248 reportFatalUsageError(Twine(R.getMsg()));
249 else
250 MORE.emit(R);
251}
252
257}
258
262 MF.getProperties().setFailedISel();
263 reportGISelDiagnostic(DS_Error, MF, TPC, MORE, R);
264}
265
268 const char *PassName, StringRef Msg,
269 const MachineInstr &MI) {
270 MachineOptimizationRemarkMissed R(PassName, "GISelFailure: ",
271 MI.getDebugLoc(), MI.getParent());
272 R << Msg;
273 // Printing MI is expensive; only do it if expensive remarks are enabled.
274 if (TPC.isGlobalISelAbortEnabled() || MORE.allowExtraAnalysis(PassName))
275 R << ": " << ore::MNV("Inst", MI);
276 reportGISelFailure(MF, TPC, MORE, R);
277}
278
279unsigned llvm::getInverseGMinMaxOpcode(unsigned MinMaxOpc) {
280 switch (MinMaxOpc) {
281 case TargetOpcode::G_SMIN:
282 return TargetOpcode::G_SMAX;
283 case TargetOpcode::G_SMAX:
284 return TargetOpcode::G_SMIN;
285 case TargetOpcode::G_UMIN:
286 return TargetOpcode::G_UMAX;
287 case TargetOpcode::G_UMAX:
288 return TargetOpcode::G_UMIN;
289 default:
290 llvm_unreachable("unrecognized opcode");
291 }
292}
293
294std::optional<APInt> llvm::getIConstantVRegVal(Register VReg,
295 const MachineRegisterInfo &MRI) {
296 std::optional<ValueAndVReg> ValAndVReg = getIConstantVRegValWithLookThrough(
297 VReg, MRI, /*LookThroughInstrs*/ false);
298 assert((!ValAndVReg || ValAndVReg->VReg == VReg) &&
299 "Value found while looking through instrs");
300 if (!ValAndVReg)
301 return std::nullopt;
302 return ValAndVReg->Value;
303}
304
306 const MachineRegisterInfo &MRI) {
307 MachineInstr *Const = MRI.getVRegDef(Reg);
308 assert((Const && Const->getOpcode() == TargetOpcode::G_CONSTANT) &&
309 "expected a G_CONSTANT on Reg");
310 return Const->getOperand(1).getCImm()->getValue();
311}
312
313std::optional<int64_t>
315 std::optional<APInt> Val = getIConstantVRegVal(VReg, MRI);
316 if (Val && Val->getBitWidth() <= 64)
317 return Val->getSExtValue();
318 return std::nullopt;
319}
320
321namespace {
322
323// This function is used in many places, and as such, it has some
324// micro-optimizations to try and make it as fast as it can be.
325//
326// - We use template arguments to avoid an indirect call caused by passing a
327// function_ref/std::function
328// - GetAPCstValue does not return std::optional<APInt> as that's expensive.
329// Instead it returns true/false and places the result in a pre-constructed
330// APInt.
331//
332// Please change this function carefully and benchmark your changes.
333template <bool (*IsConstantOpcode)(const MachineInstr *),
334 bool (*GetAPCstValue)(const MachineInstr *MI, APInt &)>
335std::optional<ValueAndVReg>
336getConstantVRegValWithLookThrough(Register VReg, const MachineRegisterInfo &MRI,
337 bool LookThroughInstrs = true,
338 bool LookThroughAnyExt = false) {
341
342 while ((MI = MRI.getVRegDef(VReg)) && !IsConstantOpcode(MI) &&
343 LookThroughInstrs) {
344 switch (MI->getOpcode()) {
345 case TargetOpcode::G_ANYEXT:
346 if (!LookThroughAnyExt)
347 return std::nullopt;
348 [[fallthrough]];
349 case TargetOpcode::G_TRUNC:
350 case TargetOpcode::G_SEXT:
351 case TargetOpcode::G_ZEXT:
352 SeenOpcodes.push_back(std::make_pair(
353 MI->getOpcode(),
354 MRI.getType(MI->getOperand(0).getReg()).getSizeInBits()));
355 VReg = MI->getOperand(1).getReg();
356 break;
357 case TargetOpcode::COPY:
358 VReg = MI->getOperand(1).getReg();
359 if (VReg.isPhysical())
360 return std::nullopt;
361 break;
362 case TargetOpcode::G_INTTOPTR:
363 VReg = MI->getOperand(1).getReg();
364 break;
365 default:
366 return std::nullopt;
367 }
368 }
369 if (!MI || !IsConstantOpcode(MI))
370 return std::nullopt;
371
372 APInt Val;
373 if (!GetAPCstValue(MI, Val))
374 return std::nullopt;
375 for (auto &Pair : reverse(SeenOpcodes)) {
376 switch (Pair.first) {
377 case TargetOpcode::G_TRUNC:
378 Val = Val.trunc(Pair.second);
379 break;
380 case TargetOpcode::G_ANYEXT:
381 case TargetOpcode::G_SEXT:
382 Val = Val.sext(Pair.second);
383 break;
384 case TargetOpcode::G_ZEXT:
385 Val = Val.zext(Pair.second);
386 break;
387 }
388 }
389
390 return ValueAndVReg{std::move(Val), VReg};
391}
392
393bool isIConstant(const MachineInstr *MI) {
394 if (!MI)
395 return false;
396 return MI->getOpcode() == TargetOpcode::G_CONSTANT;
397}
398
399bool isFConstant(const MachineInstr *MI) {
400 if (!MI)
401 return false;
402 return MI->getOpcode() == TargetOpcode::G_FCONSTANT;
403}
404
405bool isAnyConstant(const MachineInstr *MI) {
406 if (!MI)
407 return false;
408 unsigned Opc = MI->getOpcode();
409 return Opc == TargetOpcode::G_CONSTANT || Opc == TargetOpcode::G_FCONSTANT;
410}
411
412bool getCImmAsAPInt(const MachineInstr *MI, APInt &Result) {
413 const MachineOperand &CstVal = MI->getOperand(1);
414 if (!CstVal.isCImm())
415 return false;
416 Result = CstVal.getCImm()->getValue();
417 return true;
418}
419
420bool getCImmOrFPImmAsAPInt(const MachineInstr *MI, APInt &Result) {
421 const MachineOperand &CstVal = MI->getOperand(1);
422 if (CstVal.isCImm())
423 Result = CstVal.getCImm()->getValue();
424 else if (CstVal.isFPImm())
426 else
427 return false;
428 return true;
429}
430
431} // end anonymous namespace
432
434 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) {
435 return getConstantVRegValWithLookThrough<isIConstant, getCImmAsAPInt>(
436 VReg, MRI, LookThroughInstrs);
437}
438
440 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs,
441 bool LookThroughAnyExt) {
442 return getConstantVRegValWithLookThrough<isAnyConstant,
443 getCImmOrFPImmAsAPInt>(
444 VReg, MRI, LookThroughInstrs, LookThroughAnyExt);
445}
446
447std::optional<FPValueAndVReg> llvm::getFConstantVRegValWithLookThrough(
448 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) {
449 auto Reg =
450 getConstantVRegValWithLookThrough<isFConstant, getCImmOrFPImmAsAPInt>(
451 VReg, MRI, LookThroughInstrs);
452 if (!Reg)
453 return std::nullopt;
455 Reg->VReg};
456}
457
458const ConstantFP *
460 MachineInstr *MI = MRI.getVRegDef(VReg);
461 if (TargetOpcode::G_FCONSTANT != MI->getOpcode())
462 return nullptr;
463 return MI->getOperand(1).getFPImm();
464}
465
466std::optional<DefinitionAndSourceRegister>
468 Register DefSrcReg = Reg;
469 // This assumes that the code is in SSA form, so there should only be one
470 // definition.
471 auto DefIt = MRI.def_begin(Reg);
472 if (DefIt == MRI.def_end())
473 return {};
474 MachineOperand &DefOpnd = *DefIt;
475 MachineInstr *DefMI = DefOpnd.getParent();
476 auto DstTy = MRI.getType(DefOpnd.getReg());
477 if (!DstTy.isValid())
478 return std::nullopt;
479 unsigned Opc = DefMI->getOpcode();
480 while (Opc == TargetOpcode::COPY || isPreISelGenericOptimizationHint(Opc)) {
481 Register SrcReg = DefMI->getOperand(1).getReg();
482 auto SrcTy = MRI.getType(SrcReg);
483 if (!SrcTy.isValid())
484 break;
485 DefMI = MRI.getVRegDef(SrcReg);
486 DefSrcReg = SrcReg;
487 Opc = DefMI->getOpcode();
488 }
489 return DefinitionAndSourceRegister{DefMI, DefSrcReg};
490}
491
493 const MachineRegisterInfo &MRI) {
494 std::optional<DefinitionAndSourceRegister> DefSrcReg =
496 return DefSrcReg ? DefSrcReg->MI : nullptr;
497}
498
500 const MachineRegisterInfo &MRI) {
501 std::optional<DefinitionAndSourceRegister> DefSrcReg =
503 return DefSrcReg ? DefSrcReg->Reg : Register();
504}
505
506void llvm::extractParts(Register Reg, LLT Ty, int NumParts,
508 MachineIRBuilder &MIRBuilder,
510 for (int i = 0; i < NumParts; ++i)
511 VRegs.push_back(MRI.createGenericVirtualRegister(Ty));
512 MIRBuilder.buildUnmerge(VRegs, Reg);
513}
514
515bool llvm::extractParts(Register Reg, LLT RegTy, LLT MainTy, LLT &LeftoverTy,
517 SmallVectorImpl<Register> &LeftoverRegs,
518 MachineIRBuilder &MIRBuilder,
520 assert(!LeftoverTy.isValid() && "this is an out argument");
521
522 unsigned RegSize = RegTy.getSizeInBits();
523 unsigned MainSize = MainTy.getSizeInBits();
524 unsigned NumParts = RegSize / MainSize;
525 unsigned LeftoverSize = RegSize - NumParts * MainSize;
526
527 // Use an unmerge when possible.
528 if (LeftoverSize == 0) {
529 for (unsigned I = 0; I < NumParts; ++I)
530 VRegs.push_back(MRI.createGenericVirtualRegister(MainTy));
531 MIRBuilder.buildUnmerge(VRegs, Reg);
532 return true;
533 }
534
535 // Try to use unmerge for irregular vector split where possible
536 // For example when splitting a <6 x i32> into <4 x i32> with <2 x i32>
537 // leftover, it becomes:
538 // <2 x i32> %2, <2 x i32>%3, <2 x i32> %4 = G_UNMERGE_VALUE <6 x i32> %1
539 // <4 x i32> %5 = G_CONCAT_VECTOR <2 x i32> %2, <2 x i32> %3
540 if (RegTy.isVector() && MainTy.isVector()) {
541 unsigned RegNumElts = RegTy.getNumElements();
542 unsigned MainNumElts = MainTy.getNumElements();
543 unsigned LeftoverNumElts = RegNumElts % MainNumElts;
544 // If can unmerge to LeftoverTy, do it
545 if (MainNumElts % LeftoverNumElts == 0 &&
546 RegNumElts % LeftoverNumElts == 0 &&
547 RegTy.getScalarSizeInBits() == MainTy.getScalarSizeInBits() &&
548 LeftoverNumElts > 1) {
549 LeftoverTy = LLT::fixed_vector(LeftoverNumElts, RegTy.getElementType());
550
551 // Unmerge the SrcReg to LeftoverTy vectors
552 SmallVector<Register, 4> UnmergeValues;
553 extractParts(Reg, LeftoverTy, RegNumElts / LeftoverNumElts, UnmergeValues,
554 MIRBuilder, MRI);
555
556 // Find how many LeftoverTy makes one MainTy
557 unsigned LeftoverPerMain = MainNumElts / LeftoverNumElts;
558 unsigned NumOfLeftoverVal =
559 ((RegNumElts % MainNumElts) / LeftoverNumElts);
560
561 // Create as many MainTy as possible using unmerged value
562 SmallVector<Register, 4> MergeValues;
563 for (unsigned I = 0; I < UnmergeValues.size() - NumOfLeftoverVal; I++) {
564 MergeValues.push_back(UnmergeValues[I]);
565 if (MergeValues.size() == LeftoverPerMain) {
566 VRegs.push_back(
567 MIRBuilder.buildMergeLikeInstr(MainTy, MergeValues).getReg(0));
568 MergeValues.clear();
569 }
570 }
571 // Populate LeftoverRegs with the leftovers
572 for (unsigned I = UnmergeValues.size() - NumOfLeftoverVal;
573 I < UnmergeValues.size(); I++) {
574 LeftoverRegs.push_back(UnmergeValues[I]);
575 }
576 return true;
577 }
578 }
579 // Perform irregular split. Leftover is last element of RegPieces.
580 if (MainTy.isVector()) {
581 SmallVector<Register, 8> RegPieces;
582 extractVectorParts(Reg, MainTy.getNumElements(), RegPieces, MIRBuilder,
583 MRI);
584 for (unsigned i = 0; i < RegPieces.size() - 1; ++i)
585 VRegs.push_back(RegPieces[i]);
586 LeftoverRegs.push_back(RegPieces[RegPieces.size() - 1]);
587 LeftoverTy = MRI.getType(LeftoverRegs[0]);
588 return true;
589 }
590
591 LeftoverTy = LLT::scalar(LeftoverSize);
592 // For irregular sizes, extract the individual parts.
593 for (unsigned I = 0; I != NumParts; ++I) {
594 Register NewReg = MRI.createGenericVirtualRegister(MainTy);
595 VRegs.push_back(NewReg);
596 MIRBuilder.buildExtract(NewReg, Reg, MainSize * I);
597 }
598
599 for (unsigned Offset = MainSize * NumParts; Offset < RegSize;
600 Offset += LeftoverSize) {
601 Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy);
602 LeftoverRegs.push_back(NewReg);
603 MIRBuilder.buildExtract(NewReg, Reg, Offset);
604 }
605
606 return true;
607}
608
609void llvm::extractVectorParts(Register Reg, unsigned NumElts,
611 MachineIRBuilder &MIRBuilder,
613 LLT RegTy = MRI.getType(Reg);
614 assert(RegTy.isVector() && "Expected a vector type");
615
616 LLT EltTy = RegTy.getElementType();
617 LLT NarrowTy = (NumElts == 1) ? EltTy : LLT::fixed_vector(NumElts, EltTy);
618 unsigned RegNumElts = RegTy.getNumElements();
619 unsigned LeftoverNumElts = RegNumElts % NumElts;
620 unsigned NumNarrowTyPieces = RegNumElts / NumElts;
621
622 // Perfect split without leftover
623 if (LeftoverNumElts == 0)
624 return extractParts(Reg, NarrowTy, NumNarrowTyPieces, VRegs, MIRBuilder,
625 MRI);
626
627 // Irregular split. Provide direct access to all elements for artifact
628 // combiner using unmerge to elements. Then build vectors with NumElts
629 // elements. Remaining element(s) will be (used to build vector) Leftover.
631 extractParts(Reg, EltTy, RegNumElts, Elts, MIRBuilder, MRI);
632
633 unsigned Offset = 0;
634 // Requested sub-vectors of NarrowTy.
635 for (unsigned i = 0; i < NumNarrowTyPieces; ++i, Offset += NumElts) {
636 ArrayRef<Register> Pieces(&Elts[Offset], NumElts);
637 VRegs.push_back(MIRBuilder.buildMergeLikeInstr(NarrowTy, Pieces).getReg(0));
638 }
639
640 // Leftover element(s).
641 if (LeftoverNumElts == 1) {
642 VRegs.push_back(Elts[Offset]);
643 } else {
644 LLT LeftoverTy = LLT::fixed_vector(LeftoverNumElts, EltTy);
645 ArrayRef<Register> Pieces(&Elts[Offset], LeftoverNumElts);
646 VRegs.push_back(
647 MIRBuilder.buildMergeLikeInstr(LeftoverTy, Pieces).getReg(0));
648 }
649}
650
652 const MachineRegisterInfo &MRI) {
654 return DefMI && DefMI->getOpcode() == Opcode ? DefMI : nullptr;
655}
656
657APFloat llvm::getAPFloatFromSize(double Val, unsigned Size) {
658 if (Size == 32)
659 return APFloat(float(Val));
660 if (Size == 64)
661 return APFloat(Val);
662 if (Size != 16)
663 llvm_unreachable("Unsupported FPConstant size");
664 bool Ignored;
665 APFloat APF(Val);
666 APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &Ignored);
667 return APF;
668}
669
670std::optional<APInt> llvm::ConstantFoldBinOp(unsigned Opcode,
671 const Register Op1,
672 const Register Op2,
673 const MachineRegisterInfo &MRI) {
674 auto MaybeOp2Cst = getAnyConstantVRegValWithLookThrough(Op2, MRI, false);
675 if (!MaybeOp2Cst)
676 return std::nullopt;
677
678 auto MaybeOp1Cst = getAnyConstantVRegValWithLookThrough(Op1, MRI, false);
679 if (!MaybeOp1Cst)
680 return std::nullopt;
681
682 const APInt &C1 = MaybeOp1Cst->Value;
683 const APInt &C2 = MaybeOp2Cst->Value;
684 switch (Opcode) {
685 default:
686 break;
687 case TargetOpcode::G_ADD:
688 return C1 + C2;
689 case TargetOpcode::G_PTR_ADD:
690 // Types can be of different width here.
691 // Result needs to be the same width as C1, so trunc or sext C2.
692 return C1 + C2.sextOrTrunc(C1.getBitWidth());
693 case TargetOpcode::G_AND:
694 return C1 & C2;
695 case TargetOpcode::G_ASHR:
696 return C1.ashr(C2);
697 case TargetOpcode::G_LSHR:
698 return C1.lshr(C2);
699 case TargetOpcode::G_MUL:
700 return C1 * C2;
701 case TargetOpcode::G_OR:
702 return C1 | C2;
703 case TargetOpcode::G_SHL:
704 return C1 << C2;
705 case TargetOpcode::G_SUB:
706 return C1 - C2;
707 case TargetOpcode::G_XOR:
708 return C1 ^ C2;
709 case TargetOpcode::G_UDIV:
710 if (!C2.getBoolValue())
711 break;
712 return C1.udiv(C2);
713 case TargetOpcode::G_SDIV:
714 if (!C2.getBoolValue())
715 break;
716 return C1.sdiv(C2);
717 case TargetOpcode::G_UREM:
718 if (!C2.getBoolValue())
719 break;
720 return C1.urem(C2);
721 case TargetOpcode::G_SREM:
722 if (!C2.getBoolValue())
723 break;
724 return C1.srem(C2);
725 case TargetOpcode::G_SMIN:
726 return APIntOps::smin(C1, C2);
727 case TargetOpcode::G_SMAX:
728 return APIntOps::smax(C1, C2);
729 case TargetOpcode::G_UMIN:
730 return APIntOps::umin(C1, C2);
731 case TargetOpcode::G_UMAX:
732 return APIntOps::umax(C1, C2);
733 }
734
735 return std::nullopt;
736}
737
738std::optional<APFloat>
739llvm::ConstantFoldFPBinOp(unsigned Opcode, const Register Op1,
740 const Register Op2, const MachineRegisterInfo &MRI) {
741 const ConstantFP *Op2Cst = getConstantFPVRegVal(Op2, MRI);
742 if (!Op2Cst)
743 return std::nullopt;
744
745 const ConstantFP *Op1Cst = getConstantFPVRegVal(Op1, MRI);
746 if (!Op1Cst)
747 return std::nullopt;
748
749 APFloat C1 = Op1Cst->getValueAPF();
750 const APFloat &C2 = Op2Cst->getValueAPF();
751 switch (Opcode) {
752 case TargetOpcode::G_FADD:
753 C1.add(C2, APFloat::rmNearestTiesToEven);
754 return C1;
755 case TargetOpcode::G_FSUB:
756 C1.subtract(C2, APFloat::rmNearestTiesToEven);
757 return C1;
758 case TargetOpcode::G_FMUL:
759 C1.multiply(C2, APFloat::rmNearestTiesToEven);
760 return C1;
761 case TargetOpcode::G_FDIV:
762 C1.divide(C2, APFloat::rmNearestTiesToEven);
763 return C1;
764 case TargetOpcode::G_FREM:
765 C1.mod(C2);
766 return C1;
767 case TargetOpcode::G_FCOPYSIGN:
768 C1.copySign(C2);
769 return C1;
770 case TargetOpcode::G_FMINNUM:
771 return minnum(C1, C2);
772 case TargetOpcode::G_FMAXNUM:
773 return maxnum(C1, C2);
774 case TargetOpcode::G_FMINIMUM:
775 return minimum(C1, C2);
776 case TargetOpcode::G_FMAXIMUM:
777 return maximum(C1, C2);
778 case TargetOpcode::G_FMINNUM_IEEE:
779 case TargetOpcode::G_FMAXNUM_IEEE:
780 // FIXME: These operations were unfortunately named. fminnum/fmaxnum do not
781 // follow the IEEE behavior for signaling nans and follow libm's fmin/fmax,
782 // and currently there isn't a nice wrapper in APFloat for the version with
783 // correct snan handling.
784 break;
785 default:
786 break;
787 }
788
789 return std::nullopt;
790}
791
793llvm::ConstantFoldVectorBinop(unsigned Opcode, const Register Op1,
794 const Register Op2,
795 const MachineRegisterInfo &MRI) {
796 auto *SrcVec2 = getOpcodeDef<GBuildVector>(Op2, MRI);
797 if (!SrcVec2)
798 return SmallVector<APInt>();
799
800 auto *SrcVec1 = getOpcodeDef<GBuildVector>(Op1, MRI);
801 if (!SrcVec1)
802 return SmallVector<APInt>();
803
804 SmallVector<APInt> FoldedElements;
805 for (unsigned Idx = 0, E = SrcVec1->getNumSources(); Idx < E; ++Idx) {
806 auto MaybeCst = ConstantFoldBinOp(Opcode, SrcVec1->getSourceReg(Idx),
807 SrcVec2->getSourceReg(Idx), MRI);
808 if (!MaybeCst)
809 return SmallVector<APInt>();
810 FoldedElements.push_back(*MaybeCst);
811 }
812 return FoldedElements;
813}
814
816 bool SNaN) {
817 const MachineInstr *DefMI = MRI.getVRegDef(Val);
818 if (!DefMI)
819 return false;
820
821 const TargetMachine& TM = DefMI->getMF()->getTarget();
822 if (DefMI->getFlag(MachineInstr::FmNoNans) || TM.Options.NoNaNsFPMath)
823 return true;
824
825 // If the value is a constant, we can obviously see if it is a NaN or not.
826 if (const ConstantFP *FPVal = getConstantFPVRegVal(Val, MRI)) {
827 return !FPVal->getValueAPF().isNaN() ||
828 (SNaN && !FPVal->getValueAPF().isSignaling());
829 }
830
831 if (DefMI->getOpcode() == TargetOpcode::G_BUILD_VECTOR) {
832 for (const auto &Op : DefMI->uses())
833 if (!isKnownNeverNaN(Op.getReg(), MRI, SNaN))
834 return false;
835 return true;
836 }
837
838 switch (DefMI->getOpcode()) {
839 default:
840 break;
841 case TargetOpcode::G_FADD:
842 case TargetOpcode::G_FSUB:
843 case TargetOpcode::G_FMUL:
844 case TargetOpcode::G_FDIV:
845 case TargetOpcode::G_FREM:
846 case TargetOpcode::G_FSIN:
847 case TargetOpcode::G_FCOS:
848 case TargetOpcode::G_FTAN:
849 case TargetOpcode::G_FACOS:
850 case TargetOpcode::G_FASIN:
851 case TargetOpcode::G_FATAN:
852 case TargetOpcode::G_FATAN2:
853 case TargetOpcode::G_FCOSH:
854 case TargetOpcode::G_FSINH:
855 case TargetOpcode::G_FTANH:
856 case TargetOpcode::G_FMA:
857 case TargetOpcode::G_FMAD:
858 if (SNaN)
859 return true;
860
861 // TODO: Need isKnownNeverInfinity
862 return false;
863 case TargetOpcode::G_FMINNUM_IEEE:
864 case TargetOpcode::G_FMAXNUM_IEEE: {
865 if (SNaN)
866 return true;
867 // This can return a NaN if either operand is an sNaN, or if both operands
868 // are NaN.
869 return (isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI) &&
873 }
874 case TargetOpcode::G_FMINNUM:
875 case TargetOpcode::G_FMAXNUM: {
876 // Only one needs to be known not-nan, since it will be returned if the
877 // other ends up being one.
878 return isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI, SNaN) ||
880 }
881 }
882
883 if (SNaN) {
884 // FP operations quiet. For now, just handle the ones inserted during
885 // legalization.
886 switch (DefMI->getOpcode()) {
887 case TargetOpcode::G_FPEXT:
888 case TargetOpcode::G_FPTRUNC:
889 case TargetOpcode::G_FCANONICALIZE:
890 return true;
891 default:
892 return false;
893 }
894 }
895
896 return false;
897}
898
900 const MachinePointerInfo &MPO) {
901 auto PSV = dyn_cast_if_present<const PseudoSourceValue *>(MPO.V);
902 if (auto FSPV = dyn_cast_or_null<FixedStackPseudoSourceValue>(PSV)) {
903 MachineFrameInfo &MFI = MF.getFrameInfo();
904 return commonAlignment(MFI.getObjectAlign(FSPV->getFrameIndex()),
905 MPO.Offset);
906 }
907
908 if (const Value *V = dyn_cast_if_present<const Value *>(MPO.V)) {
909 const Module *M = MF.getFunction().getParent();
910 return V->getPointerAlignment(M->getDataLayout());
911 }
912
913 return Align(1);
914}
915
917 const TargetInstrInfo &TII,
918 MCRegister PhysReg,
919 const TargetRegisterClass &RC,
920 const DebugLoc &DL, LLT RegTy) {
921 MachineBasicBlock &EntryMBB = MF.front();
923 Register LiveIn = MRI.getLiveInVirtReg(PhysReg);
924 if (LiveIn) {
925 MachineInstr *Def = MRI.getVRegDef(LiveIn);
926 if (Def) {
927 // FIXME: Should the verifier check this is in the entry block?
928 assert(Def->getParent() == &EntryMBB && "live-in copy not in entry block");
929 return LiveIn;
930 }
931
932 // It's possible the incoming argument register and copy was added during
933 // lowering, but later deleted due to being/becoming dead. If this happens,
934 // re-insert the copy.
935 } else {
936 // The live in register was not present, so add it.
937 LiveIn = MF.addLiveIn(PhysReg, &RC);
938 if (RegTy.isValid())
939 MRI.setType(LiveIn, RegTy);
940 }
941
942 BuildMI(EntryMBB, EntryMBB.begin(), DL, TII.get(TargetOpcode::COPY), LiveIn)
943 .addReg(PhysReg);
944 if (!EntryMBB.isLiveIn(PhysReg))
945 EntryMBB.addLiveIn(PhysReg);
946 return LiveIn;
947}
948
949std::optional<APInt> llvm::ConstantFoldExtOp(unsigned Opcode,
950 const Register Op1, uint64_t Imm,
951 const MachineRegisterInfo &MRI) {
952 auto MaybeOp1Cst = getIConstantVRegVal(Op1, MRI);
953 if (MaybeOp1Cst) {
954 switch (Opcode) {
955 default:
956 break;
957 case TargetOpcode::G_SEXT_INREG: {
958 LLT Ty = MRI.getType(Op1);
959 return MaybeOp1Cst->trunc(Imm).sext(Ty.getScalarSizeInBits());
960 }
961 }
962 }
963 return std::nullopt;
964}
965
966std::optional<APInt> llvm::ConstantFoldCastOp(unsigned Opcode, LLT DstTy,
967 const Register Op0,
968 const MachineRegisterInfo &MRI) {
969 std::optional<APInt> Val = getIConstantVRegVal(Op0, MRI);
970 if (!Val)
971 return Val;
972
973 const unsigned DstSize = DstTy.getScalarSizeInBits();
974
975 switch (Opcode) {
976 case TargetOpcode::G_SEXT:
977 return Val->sext(DstSize);
978 case TargetOpcode::G_ZEXT:
979 case TargetOpcode::G_ANYEXT:
980 // TODO: DAG considers target preference when constant folding any_extend.
981 return Val->zext(DstSize);
982 default:
983 break;
984 }
985
986 llvm_unreachable("unexpected cast opcode to constant fold");
987}
988
989std::optional<APFloat>
990llvm::ConstantFoldIntToFloat(unsigned Opcode, LLT DstTy, Register Src,
991 const MachineRegisterInfo &MRI) {
992 assert(Opcode == TargetOpcode::G_SITOFP || Opcode == TargetOpcode::G_UITOFP);
993 if (auto MaybeSrcVal = getIConstantVRegVal(Src, MRI)) {
994 APFloat DstVal(getFltSemanticForLLT(DstTy));
995 DstVal.convertFromAPInt(*MaybeSrcVal, Opcode == TargetOpcode::G_SITOFP,
996 APFloat::rmNearestTiesToEven);
997 return DstVal;
998 }
999 return std::nullopt;
1000}
1001
1002std::optional<SmallVector<unsigned>>
1004 std::function<unsigned(APInt)> CB) {
1005 LLT Ty = MRI.getType(Src);
1006 SmallVector<unsigned> FoldedCTLZs;
1007 auto tryFoldScalar = [&](Register R) -> std::optional<unsigned> {
1008 auto MaybeCst = getIConstantVRegVal(R, MRI);
1009 if (!MaybeCst)
1010 return std::nullopt;
1011 return CB(*MaybeCst);
1012 };
1013 if (Ty.isVector()) {
1014 // Try to constant fold each element.
1015 auto *BV = getOpcodeDef<GBuildVector>(Src, MRI);
1016 if (!BV)
1017 return std::nullopt;
1018 for (unsigned SrcIdx = 0; SrcIdx < BV->getNumSources(); ++SrcIdx) {
1019 if (auto MaybeFold = tryFoldScalar(BV->getSourceReg(SrcIdx))) {
1020 FoldedCTLZs.emplace_back(*MaybeFold);
1021 continue;
1022 }
1023 return std::nullopt;
1024 }
1025 return FoldedCTLZs;
1026 }
1027 if (auto MaybeCst = tryFoldScalar(Src)) {
1028 FoldedCTLZs.emplace_back(*MaybeCst);
1029 return FoldedCTLZs;
1030 }
1031 return std::nullopt;
1032}
1033
1034std::optional<SmallVector<APInt>>
1035llvm::ConstantFoldICmp(unsigned Pred, const Register Op1, const Register Op2,
1036 unsigned DstScalarSizeInBits, unsigned ExtOp,
1037 const MachineRegisterInfo &MRI) {
1038 assert(ExtOp == TargetOpcode::G_SEXT || ExtOp == TargetOpcode::G_ZEXT ||
1039 ExtOp == TargetOpcode::G_ANYEXT);
1040
1041 const LLT Ty = MRI.getType(Op1);
1042
1043 auto GetICmpResultCst = [&](bool IsTrue) {
1044 if (IsTrue)
1045 return ExtOp == TargetOpcode::G_SEXT
1046 ? APInt::getAllOnes(DstScalarSizeInBits)
1047 : APInt::getOneBitSet(DstScalarSizeInBits, 0);
1048 return APInt::getZero(DstScalarSizeInBits);
1049 };
1050
1051 auto TryFoldScalar = [&](Register LHS, Register RHS) -> std::optional<APInt> {
1052 auto RHSCst = getIConstantVRegVal(RHS, MRI);
1053 if (!RHSCst)
1054 return std::nullopt;
1055 auto LHSCst = getIConstantVRegVal(LHS, MRI);
1056 if (!LHSCst)
1057 return std::nullopt;
1058
1059 switch (Pred) {
1060 case CmpInst::Predicate::ICMP_EQ:
1061 return GetICmpResultCst(LHSCst->eq(*RHSCst));
1062 case CmpInst::Predicate::ICMP_NE:
1063 return GetICmpResultCst(LHSCst->ne(*RHSCst));
1064 case CmpInst::Predicate::ICMP_UGT:
1065 return GetICmpResultCst(LHSCst->ugt(*RHSCst));
1066 case CmpInst::Predicate::ICMP_UGE:
1067 return GetICmpResultCst(LHSCst->uge(*RHSCst));
1068 case CmpInst::Predicate::ICMP_ULT:
1069 return GetICmpResultCst(LHSCst->ult(*RHSCst));
1070 case CmpInst::Predicate::ICMP_ULE:
1071 return GetICmpResultCst(LHSCst->ule(*RHSCst));
1072 case CmpInst::Predicate::ICMP_SGT:
1073 return GetICmpResultCst(LHSCst->sgt(*RHSCst));
1074 case CmpInst::Predicate::ICMP_SGE:
1075 return GetICmpResultCst(LHSCst->sge(*RHSCst));
1076 case CmpInst::Predicate::ICMP_SLT:
1077 return GetICmpResultCst(LHSCst->slt(*RHSCst));
1078 case CmpInst::Predicate::ICMP_SLE:
1079 return GetICmpResultCst(LHSCst->sle(*RHSCst));
1080 default:
1081 return std::nullopt;
1082 }
1083 };
1084
1085 SmallVector<APInt> FoldedICmps;
1086
1087 if (Ty.isVector()) {
1088 // Try to constant fold each element.
1089 auto *BV1 = getOpcodeDef<GBuildVector>(Op1, MRI);
1090 auto *BV2 = getOpcodeDef<GBuildVector>(Op2, MRI);
1091 if (!BV1 || !BV2)
1092 return std::nullopt;
1093 assert(BV1->getNumSources() == BV2->getNumSources() && "Invalid vectors");
1094 for (unsigned I = 0; I < BV1->getNumSources(); ++I) {
1095 if (auto MaybeFold =
1096 TryFoldScalar(BV1->getSourceReg(I), BV2->getSourceReg(I))) {
1097 FoldedICmps.emplace_back(*MaybeFold);
1098 continue;
1099 }
1100 return std::nullopt;
1101 }
1102 return FoldedICmps;
1103 }
1104
1105 if (auto MaybeCst = TryFoldScalar(Op1, Op2)) {
1106 FoldedICmps.emplace_back(*MaybeCst);
1107 return FoldedICmps;
1108 }
1109
1110 return std::nullopt;
1111}
1112
1114 GISelValueTracking *VT) {
1115 std::optional<DefinitionAndSourceRegister> DefSrcReg =
1117 if (!DefSrcReg)
1118 return false;
1119
1120 const MachineInstr &MI = *DefSrcReg->MI;
1121 const LLT Ty = MRI.getType(Reg);
1122
1123 switch (MI.getOpcode()) {
1124 case TargetOpcode::G_CONSTANT: {
1125 unsigned BitWidth = Ty.getScalarSizeInBits();
1126 const ConstantInt *CI = MI.getOperand(1).getCImm();
1127 return CI->getValue().zextOrTrunc(BitWidth).isPowerOf2();
1128 }
1129 case TargetOpcode::G_SHL: {
1130 // A left-shift of a constant one will have exactly one bit set because
1131 // shifting the bit off the end is undefined.
1132
1133 // TODO: Constant splat
1134 if (auto ConstLHS = getIConstantVRegVal(MI.getOperand(1).getReg(), MRI)) {
1135 if (*ConstLHS == 1)
1136 return true;
1137 }
1138
1139 break;
1140 }
1141 case TargetOpcode::G_LSHR: {
1142 if (auto ConstLHS = getIConstantVRegVal(MI.getOperand(1).getReg(), MRI)) {
1143 if (ConstLHS->isSignMask())
1144 return true;
1145 }
1146
1147 break;
1148 }
1149 case TargetOpcode::G_BUILD_VECTOR: {
1150 // TODO: Probably should have a recursion depth guard since you could have
1151 // bitcasted vector elements.
1152 for (const MachineOperand &MO : llvm::drop_begin(MI.operands()))
1153 if (!isKnownToBeAPowerOfTwo(MO.getReg(), MRI, VT))
1154 return false;
1155
1156 return true;
1157 }
1158 case TargetOpcode::G_BUILD_VECTOR_TRUNC: {
1159 // Only handle constants since we would need to know if number of leading
1160 // zeros is greater than the truncation amount.
1161 const unsigned BitWidth = Ty.getScalarSizeInBits();
1162 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) {
1163 auto Const = getIConstantVRegVal(MO.getReg(), MRI);
1164 if (!Const || !Const->zextOrTrunc(BitWidth).isPowerOf2())
1165 return false;
1166 }
1167
1168 return true;
1169 }
1170 default:
1171 break;
1172 }
1173
1174 if (!VT)
1175 return false;
1176
1177 // More could be done here, though the above checks are enough
1178 // to handle some common cases.
1179
1180 // Fall back to computeKnownBits to catch other known cases.
1181 KnownBits Known = VT->getKnownBits(Reg);
1182 return (Known.countMaxPopulation() == 1) && (Known.countMinPopulation() == 1);
1183}
1184
1187}
1188
1189LLT llvm::getLCMType(LLT OrigTy, LLT TargetTy) {
1190 if (OrigTy.getSizeInBits() == TargetTy.getSizeInBits())
1191 return OrigTy;
1192
1193 if (OrigTy.isVector() && TargetTy.isVector()) {
1194 LLT OrigElt = OrigTy.getElementType();
1195 LLT TargetElt = TargetTy.getElementType();
1196
1197 // TODO: The docstring for this function says the intention is to use this
1198 // function to build MERGE/UNMERGE instructions. It won't be the case that
1199 // we generate a MERGE/UNMERGE between fixed and scalable vector types. We
1200 // could implement getLCMType between the two in the future if there was a
1201 // need, but it is not worth it now as this function should not be used in
1202 // that way.
1203 assert(((OrigTy.isScalableVector() && !TargetTy.isFixedVector()) ||
1204 (OrigTy.isFixedVector() && !TargetTy.isScalableVector())) &&
1205 "getLCMType not implemented between fixed and scalable vectors.");
1206
1207 if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) {
1208 int GCDMinElts = std::gcd(OrigTy.getElementCount().getKnownMinValue(),
1209 TargetTy.getElementCount().getKnownMinValue());
1210 // Prefer the original element type.
1212 TargetTy.getElementCount().getKnownMinValue());
1213 return LLT::vector(Mul.divideCoefficientBy(GCDMinElts),
1214 OrigTy.getElementType());
1215 }
1216 unsigned LCM = std::lcm(OrigTy.getSizeInBits().getKnownMinValue(),
1217 TargetTy.getSizeInBits().getKnownMinValue());
1218 return LLT::vector(
1219 ElementCount::get(LCM / OrigElt.getSizeInBits(), OrigTy.isScalable()),
1220 OrigElt);
1221 }
1222
1223 // One type is scalar, one type is vector
1224 if (OrigTy.isVector() || TargetTy.isVector()) {
1225 LLT VecTy = OrigTy.isVector() ? OrigTy : TargetTy;
1226 LLT ScalarTy = OrigTy.isVector() ? TargetTy : OrigTy;
1227 LLT EltTy = VecTy.getElementType();
1228 LLT OrigEltTy = OrigTy.isVector() ? OrigTy.getElementType() : OrigTy;
1229
1230 // Prefer scalar type from OrigTy.
1231 if (EltTy.getSizeInBits() == ScalarTy.getSizeInBits())
1232 return LLT::vector(VecTy.getElementCount(), OrigEltTy);
1233
1234 // Different size scalars. Create vector with the same total size.
1235 // LCM will take fixed/scalable from VecTy.
1236 unsigned LCM = std::lcm(EltTy.getSizeInBits().getFixedValue() *
1238 ScalarTy.getSizeInBits().getFixedValue());
1239 // Prefer type from OrigTy
1240 return LLT::vector(ElementCount::get(LCM / OrigEltTy.getSizeInBits(),
1241 VecTy.getElementCount().isScalable()),
1242 OrigEltTy);
1243 }
1244
1245 // At this point, both types are scalars of different size
1246 unsigned LCM = std::lcm(OrigTy.getSizeInBits().getFixedValue(),
1247 TargetTy.getSizeInBits().getFixedValue());
1248 // Preserve pointer types.
1249 if (LCM == OrigTy.getSizeInBits())
1250 return OrigTy;
1251 if (LCM == TargetTy.getSizeInBits())
1252 return TargetTy;
1253 return LLT::scalar(LCM);
1254}
1255
1256LLT llvm::getCoverTy(LLT OrigTy, LLT TargetTy) {
1257
1258 if ((OrigTy.isScalableVector() && TargetTy.isFixedVector()) ||
1259 (OrigTy.isFixedVector() && TargetTy.isScalableVector()))
1261 "getCoverTy not implemented between fixed and scalable vectors.");
1262
1263 if (!OrigTy.isVector() || !TargetTy.isVector() || OrigTy == TargetTy ||
1264 (OrigTy.getScalarSizeInBits() != TargetTy.getScalarSizeInBits()))
1265 return getLCMType(OrigTy, TargetTy);
1266
1267 unsigned OrigTyNumElts = OrigTy.getElementCount().getKnownMinValue();
1268 unsigned TargetTyNumElts = TargetTy.getElementCount().getKnownMinValue();
1269 if (OrigTyNumElts % TargetTyNumElts == 0)
1270 return OrigTy;
1271
1272 unsigned NumElts = alignTo(OrigTyNumElts, TargetTyNumElts);
1274 OrigTy.getElementType());
1275}
1276
1277LLT llvm::getGCDType(LLT OrigTy, LLT TargetTy) {
1278 if (OrigTy.getSizeInBits() == TargetTy.getSizeInBits())
1279 return OrigTy;
1280
1281 if (OrigTy.isVector() && TargetTy.isVector()) {
1282 LLT OrigElt = OrigTy.getElementType();
1283
1284 // TODO: The docstring for this function says the intention is to use this
1285 // function to build MERGE/UNMERGE instructions. It won't be the case that
1286 // we generate a MERGE/UNMERGE between fixed and scalable vector types. We
1287 // could implement getGCDType between the two in the future if there was a
1288 // need, but it is not worth it now as this function should not be used in
1289 // that way.
1290 assert(((OrigTy.isScalableVector() && !TargetTy.isFixedVector()) ||
1291 (OrigTy.isFixedVector() && !TargetTy.isScalableVector())) &&
1292 "getGCDType not implemented between fixed and scalable vectors.");
1293
1294 unsigned GCD = std::gcd(OrigTy.getSizeInBits().getKnownMinValue(),
1295 TargetTy.getSizeInBits().getKnownMinValue());
1296 if (GCD == OrigElt.getSizeInBits())
1298 OrigElt);
1299
1300 // Cannot produce original element type, but both have vscale in common.
1301 if (GCD < OrigElt.getSizeInBits())
1303 GCD);
1304
1305 return LLT::vector(
1307 OrigTy.isScalable()),
1308 OrigElt);
1309 }
1310
1311 // If one type is vector and the element size matches the scalar size, then
1312 // the gcd is the scalar type.
1313 if (OrigTy.isVector() &&
1314 OrigTy.getElementType().getSizeInBits() == TargetTy.getSizeInBits())
1315 return OrigTy.getElementType();
1316 if (TargetTy.isVector() &&
1317 TargetTy.getElementType().getSizeInBits() == OrigTy.getSizeInBits())
1318 return OrigTy;
1319
1320 // At this point, both types are either scalars of different type or one is a
1321 // vector and one is a scalar. If both types are scalars, the GCD type is the
1322 // GCD between the two scalar sizes. If one is vector and one is scalar, then
1323 // the GCD type is the GCD between the scalar and the vector element size.
1324 LLT OrigScalar = OrigTy.getScalarType();
1325 LLT TargetScalar = TargetTy.getScalarType();
1326 unsigned GCD = std::gcd(OrigScalar.getSizeInBits().getFixedValue(),
1327 TargetScalar.getSizeInBits().getFixedValue());
1328 return LLT::scalar(GCD);
1329}
1330
1332 assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR &&
1333 "Only G_SHUFFLE_VECTOR can have a splat index!");
1334 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask();
1335 auto FirstDefinedIdx = find_if(Mask, [](int Elt) { return Elt >= 0; });
1336
1337 // If all elements are undefined, this shuffle can be considered a splat.
1338 // Return 0 for better potential for callers to simplify.
1339 if (FirstDefinedIdx == Mask.end())
1340 return 0;
1341
1342 // Make sure all remaining elements are either undef or the same
1343 // as the first non-undef value.
1344 int SplatValue = *FirstDefinedIdx;
1345 if (any_of(make_range(std::next(FirstDefinedIdx), Mask.end()),
1346 [&SplatValue](int Elt) { return Elt >= 0 && Elt != SplatValue; }))
1347 return std::nullopt;
1348
1349 return SplatValue;
1350}
1351
1352static bool isBuildVectorOp(unsigned Opcode) {
1353 return Opcode == TargetOpcode::G_BUILD_VECTOR ||
1354 Opcode == TargetOpcode::G_BUILD_VECTOR_TRUNC;
1355}
1356
1357namespace {
1358
1359std::optional<ValueAndVReg> getAnyConstantSplat(Register VReg,
1360 const MachineRegisterInfo &MRI,
1361 bool AllowUndef) {
1363 if (!MI)
1364 return std::nullopt;
1365
1366 bool isConcatVectorsOp = MI->getOpcode() == TargetOpcode::G_CONCAT_VECTORS;
1367 if (!isBuildVectorOp(MI->getOpcode()) && !isConcatVectorsOp)
1368 return std::nullopt;
1369
1370 std::optional<ValueAndVReg> SplatValAndReg;
1371 for (MachineOperand &Op : MI->uses()) {
1372 Register Element = Op.getReg();
1373 // If we have a G_CONCAT_VECTOR, we recursively look into the
1374 // vectors that we're concatenating to see if they're splats.
1375 auto ElementValAndReg =
1376 isConcatVectorsOp
1377 ? getAnyConstantSplat(Element, MRI, AllowUndef)
1379
1380 // If AllowUndef, treat undef as value that will result in a constant splat.
1381 if (!ElementValAndReg) {
1382 if (AllowUndef && isa<GImplicitDef>(MRI.getVRegDef(Element)))
1383 continue;
1384 return std::nullopt;
1385 }
1386
1387 // Record splat value
1388 if (!SplatValAndReg)
1389 SplatValAndReg = ElementValAndReg;
1390
1391 // Different constant than the one already recorded, not a constant splat.
1392 if (SplatValAndReg->Value != ElementValAndReg->Value)
1393 return std::nullopt;
1394 }
1395
1396 return SplatValAndReg;
1397}
1398
1399} // end anonymous namespace
1400
1402 const MachineRegisterInfo &MRI,
1403 int64_t SplatValue, bool AllowUndef) {
1404 if (auto SplatValAndReg = getAnyConstantSplat(Reg, MRI, AllowUndef))
1405 return SplatValAndReg->Value.getSExtValue() == SplatValue;
1406
1407 return false;
1408}
1409
1411 const MachineRegisterInfo &MRI,
1412 APInt SplatValue, bool AllowUndef) {
1413 if (auto SplatValAndReg = getAnyConstantSplat(Reg, MRI, AllowUndef)) {
1414 if (SplatValAndReg->Value.getBitWidth() < SplatValue.getBitWidth())
1415 return APInt::isSameValue(
1416 SplatValAndReg->Value.sext(SplatValue.getBitWidth()), SplatValue);
1417 return APInt::isSameValue(
1418 SplatValAndReg->Value,
1419 SplatValue.sext(SplatValAndReg->Value.getBitWidth()));
1420 }
1421
1422 return false;
1423}
1424
1426 const MachineRegisterInfo &MRI,
1427 int64_t SplatValue, bool AllowUndef) {
1428 return isBuildVectorConstantSplat(MI.getOperand(0).getReg(), MRI, SplatValue,
1429 AllowUndef);
1430}
1431
1433 const MachineRegisterInfo &MRI,
1434 APInt SplatValue, bool AllowUndef) {
1435 return isBuildVectorConstantSplat(MI.getOperand(0).getReg(), MRI, SplatValue,
1436 AllowUndef);
1437}
1438
1439std::optional<APInt>
1441 if (auto SplatValAndReg =
1442 getAnyConstantSplat(Reg, MRI, /* AllowUndef */ false)) {
1443 if (std::optional<ValueAndVReg> ValAndVReg =
1444 getIConstantVRegValWithLookThrough(SplatValAndReg->VReg, MRI))
1445 return ValAndVReg->Value;
1446 }
1447
1448 return std::nullopt;
1449}
1450
1451std::optional<APInt>
1453 const MachineRegisterInfo &MRI) {
1454 return getIConstantSplatVal(MI.getOperand(0).getReg(), MRI);
1455}
1456
1457std::optional<int64_t>
1459 const MachineRegisterInfo &MRI) {
1460 if (auto SplatValAndReg =
1461 getAnyConstantSplat(Reg, MRI, /* AllowUndef */ false))
1462 return getIConstantVRegSExtVal(SplatValAndReg->VReg, MRI);
1463 return std::nullopt;
1464}
1465
1466std::optional<int64_t>
1468 const MachineRegisterInfo &MRI) {
1469 return getIConstantSplatSExtVal(MI.getOperand(0).getReg(), MRI);
1470}
1471
1472std::optional<FPValueAndVReg>
1474 bool AllowUndef) {
1475 if (auto SplatValAndReg = getAnyConstantSplat(VReg, MRI, AllowUndef))
1476 return getFConstantVRegValWithLookThrough(SplatValAndReg->VReg, MRI);
1477 return std::nullopt;
1478}
1479
1481 const MachineRegisterInfo &MRI,
1482 bool AllowUndef) {
1483 return isBuildVectorConstantSplat(MI, MRI, 0, AllowUndef);
1484}
1485
1487 const MachineRegisterInfo &MRI,
1488 bool AllowUndef) {
1489 return isBuildVectorConstantSplat(MI, MRI, -1, AllowUndef);
1490}
1491
1492std::optional<RegOrConstant>
1494 unsigned Opc = MI.getOpcode();
1495 if (!isBuildVectorOp(Opc))
1496 return std::nullopt;
1497 if (auto Splat = getIConstantSplatSExtVal(MI, MRI))
1498 return RegOrConstant(*Splat);
1499 auto Reg = MI.getOperand(1).getReg();
1500 if (any_of(drop_begin(MI.operands(), 2),
1501 [&Reg](const MachineOperand &Op) { return Op.getReg() != Reg; }))
1502 return std::nullopt;
1503 return RegOrConstant(Reg);
1504}
1505
1507 const MachineRegisterInfo &MRI,
1508 bool AllowFP = true,
1509 bool AllowOpaqueConstants = true) {
1510 switch (MI.getOpcode()) {
1511 case TargetOpcode::G_CONSTANT:
1512 case TargetOpcode::G_IMPLICIT_DEF:
1513 return true;
1514 case TargetOpcode::G_FCONSTANT:
1515 return AllowFP;
1516 case TargetOpcode::G_GLOBAL_VALUE:
1517 case TargetOpcode::G_FRAME_INDEX:
1518 case TargetOpcode::G_BLOCK_ADDR:
1519 case TargetOpcode::G_JUMP_TABLE:
1520 return AllowOpaqueConstants;
1521 default:
1522 return false;
1523 }
1524}
1525
1527 const MachineRegisterInfo &MRI) {
1528 Register Def = MI.getOperand(0).getReg();
1529 if (auto C = getIConstantVRegValWithLookThrough(Def, MRI))
1530 return true;
1531 GBuildVector *BV = dyn_cast<GBuildVector>(&MI);
1532 if (!BV)
1533 return false;
1534 for (unsigned SrcIdx = 0; SrcIdx < BV->getNumSources(); ++SrcIdx) {
1536 getOpcodeDef<GImplicitDef>(BV->getSourceReg(SrcIdx), MRI))
1537 continue;
1538 return false;
1539 }
1540 return true;
1541}
1542
1544 const MachineRegisterInfo &MRI,
1545 bool AllowFP, bool AllowOpaqueConstants) {
1546 if (isConstantScalar(MI, MRI, AllowFP, AllowOpaqueConstants))
1547 return true;
1548
1549 if (!isBuildVectorOp(MI.getOpcode()))
1550 return false;
1551
1552 const unsigned NumOps = MI.getNumOperands();
1553 for (unsigned I = 1; I != NumOps; ++I) {
1554 const MachineInstr *ElementDef = MRI.getVRegDef(MI.getOperand(I).getReg());
1555 if (!isConstantScalar(*ElementDef, MRI, AllowFP, AllowOpaqueConstants))
1556 return false;
1557 }
1558
1559 return true;
1560}
1561
1562std::optional<APInt>
1564 const MachineRegisterInfo &MRI) {
1565 Register Def = MI.getOperand(0).getReg();
1566 if (auto C = getIConstantVRegValWithLookThrough(Def, MRI))
1567 return C->Value;
1568 auto MaybeCst = getIConstantSplatSExtVal(MI, MRI);
1569 if (!MaybeCst)
1570 return std::nullopt;
1571 const unsigned ScalarSize = MRI.getType(Def).getScalarSizeInBits();
1572 return APInt(ScalarSize, *MaybeCst, true);
1573}
1574
1575std::optional<APFloat>
1577 const MachineRegisterInfo &MRI) {
1578 Register Def = MI.getOperand(0).getReg();
1579 if (auto FpConst = getFConstantVRegValWithLookThrough(Def, MRI))
1580 return FpConst->Value;
1581 auto MaybeCstFP = getFConstantSplat(Def, MRI, /*allowUndef=*/false);
1582 if (!MaybeCstFP)
1583 return std::nullopt;
1584 return MaybeCstFP->Value;
1585}
1586
1588 const MachineRegisterInfo &MRI, bool AllowUndefs) {
1589 switch (MI.getOpcode()) {
1590 case TargetOpcode::G_IMPLICIT_DEF:
1591 return AllowUndefs;
1592 case TargetOpcode::G_CONSTANT:
1593 return MI.getOperand(1).getCImm()->isNullValue();
1594 case TargetOpcode::G_FCONSTANT: {
1595 const ConstantFP *FPImm = MI.getOperand(1).getFPImm();
1596 return FPImm->isZero() && !FPImm->isNegative();
1597 }
1598 default:
1599 if (!AllowUndefs) // TODO: isBuildVectorAllZeros assumes undef is OK already
1600 return false;
1601 return isBuildVectorAllZeros(MI, MRI);
1602 }
1603}
1604
1606 const MachineRegisterInfo &MRI,
1607 bool AllowUndefs) {
1608 switch (MI.getOpcode()) {
1609 case TargetOpcode::G_IMPLICIT_DEF:
1610 return AllowUndefs;
1611 case TargetOpcode::G_CONSTANT:
1612 return MI.getOperand(1).getCImm()->isAllOnesValue();
1613 default:
1614 if (!AllowUndefs) // TODO: isBuildVectorAllOnes assumes undef is OK already
1615 return false;
1616 return isBuildVectorAllOnes(MI, MRI);
1617 }
1618}
1619
1621 const MachineRegisterInfo &MRI, Register Reg,
1622 std::function<bool(const Constant *ConstVal)> Match, bool AllowUndefs) {
1623
1624 const MachineInstr *Def = getDefIgnoringCopies(Reg, MRI);
1625 if (AllowUndefs && Def->getOpcode() == TargetOpcode::G_IMPLICIT_DEF)
1626 return Match(nullptr);
1627
1628 // TODO: Also handle fconstant
1629 if (Def->getOpcode() == TargetOpcode::G_CONSTANT)
1630 return Match(Def->getOperand(1).getCImm());
1631
1632 if (Def->getOpcode() != TargetOpcode::G_BUILD_VECTOR)
1633 return false;
1634
1635 for (unsigned I = 1, E = Def->getNumOperands(); I != E; ++I) {
1636 Register SrcElt = Def->getOperand(I).getReg();
1637 const MachineInstr *SrcDef = getDefIgnoringCopies(SrcElt, MRI);
1638 if (AllowUndefs && SrcDef->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) {
1639 if (!Match(nullptr))
1640 return false;
1641 continue;
1642 }
1643
1644 if (SrcDef->getOpcode() != TargetOpcode::G_CONSTANT ||
1645 !Match(SrcDef->getOperand(1).getCImm()))
1646 return false;
1647 }
1648
1649 return true;
1650}
1651
1652bool llvm::isConstTrueVal(const TargetLowering &TLI, int64_t Val, bool IsVector,
1653 bool IsFP) {
1654 switch (TLI.getBooleanContents(IsVector, IsFP)) {
1655 case TargetLowering::UndefinedBooleanContent:
1656 return Val & 0x1;
1657 case TargetLowering::ZeroOrOneBooleanContent:
1658 return Val == 1;
1659 case TargetLowering::ZeroOrNegativeOneBooleanContent:
1660 return Val == -1;
1661 }
1662 llvm_unreachable("Invalid boolean contents");
1663}
1664
1665bool llvm::isConstFalseVal(const TargetLowering &TLI, int64_t Val,
1666 bool IsVector, bool IsFP) {
1667 switch (TLI.getBooleanContents(IsVector, IsFP)) {
1668 case TargetLowering::UndefinedBooleanContent:
1669 return ~Val & 0x1;
1670 case TargetLowering::ZeroOrOneBooleanContent:
1671 case TargetLowering::ZeroOrNegativeOneBooleanContent:
1672 return Val == 0;
1673 }
1674 llvm_unreachable("Invalid boolean contents");
1675}
1676
1677int64_t llvm::getICmpTrueVal(const TargetLowering &TLI, bool IsVector,
1678 bool IsFP) {
1679 switch (TLI.getBooleanContents(IsVector, IsFP)) {
1680 case TargetLowering::UndefinedBooleanContent:
1681 case TargetLowering::ZeroOrOneBooleanContent:
1682 return 1;
1683 case TargetLowering::ZeroOrNegativeOneBooleanContent:
1684 return -1;
1685 }
1686 llvm_unreachable("Invalid boolean contents");
1687}
1688
1690 LostDebugLocObserver *LocObserver,
1691 SmallInstListTy &DeadInstChain) {
1692 for (MachineOperand &Op : MI.uses()) {
1693 if (Op.isReg() && Op.getReg().isVirtual())
1694 DeadInstChain.insert(MRI.getVRegDef(Op.getReg()));
1695 }
1696 LLVM_DEBUG(dbgs() << MI << "Is dead; erasing.\n");
1697 DeadInstChain.remove(&MI);
1698 MI.eraseFromParent();
1699 if (LocObserver)
1700 LocObserver->checkpoint(false);
1701}
1702
1705 LostDebugLocObserver *LocObserver) {
1706 SmallInstListTy DeadInstChain;
1707 for (MachineInstr *MI : DeadInstrs)
1708 saveUsesAndErase(*MI, MRI, LocObserver, DeadInstChain);
1709
1710 while (!DeadInstChain.empty()) {
1711 MachineInstr *Inst = DeadInstChain.pop_back_val();
1712 if (!isTriviallyDead(*Inst, MRI))
1713 continue;
1714 saveUsesAndErase(*Inst, MRI, LocObserver, DeadInstChain);
1715 }
1716}
1717
1719 LostDebugLocObserver *LocObserver) {
1720 return eraseInstrs({&MI}, MRI, LocObserver);
1721}
1722
1724 for (auto &Def : MI.defs()) {
1725 assert(Def.isReg() && "Must be a reg");
1726
1728 for (auto &MOUse : MRI.use_operands(Def.getReg())) {
1729 MachineInstr *DbgValue = MOUse.getParent();
1730 // Ignore partially formed DBG_VALUEs.
1731 if (DbgValue->isNonListDebugValue() && DbgValue->getNumOperands() == 4) {
1732 DbgUsers.push_back(&MOUse);
1733 }
1734 }
1735
1736 if (!DbgUsers.empty()) {
1738 }
1739 }
1740}
1741
1743 switch (Opc) {
1744 case TargetOpcode::G_FABS:
1745 case TargetOpcode::G_FADD:
1746 case TargetOpcode::G_FCANONICALIZE:
1747 case TargetOpcode::G_FCEIL:
1748 case TargetOpcode::G_FCONSTANT:
1749 case TargetOpcode::G_FCOPYSIGN:
1750 case TargetOpcode::G_FCOS:
1751 case TargetOpcode::G_FDIV:
1752 case TargetOpcode::G_FEXP2:
1753 case TargetOpcode::G_FEXP:
1754 case TargetOpcode::G_FFLOOR:
1755 case TargetOpcode::G_FLOG10:
1756 case TargetOpcode::G_FLOG2:
1757 case TargetOpcode::G_FLOG:
1758 case TargetOpcode::G_FMA:
1759 case TargetOpcode::G_FMAD:
1760 case TargetOpcode::G_FMAXIMUM:
1761 case TargetOpcode::G_FMAXNUM:
1762 case TargetOpcode::G_FMAXNUM_IEEE:
1763 case TargetOpcode::G_FMINIMUM:
1764 case TargetOpcode::G_FMINNUM:
1765 case TargetOpcode::G_FMINNUM_IEEE:
1766 case TargetOpcode::G_FMUL:
1767 case TargetOpcode::G_FNEARBYINT:
1768 case TargetOpcode::G_FNEG:
1769 case TargetOpcode::G_FPEXT:
1770 case TargetOpcode::G_FPOW:
1771 case TargetOpcode::G_FPTRUNC:
1772 case TargetOpcode::G_FREM:
1773 case TargetOpcode::G_FRINT:
1774 case TargetOpcode::G_FSIN:
1775 case TargetOpcode::G_FTAN:
1776 case TargetOpcode::G_FACOS:
1777 case TargetOpcode::G_FASIN:
1778 case TargetOpcode::G_FATAN:
1779 case TargetOpcode::G_FATAN2:
1780 case TargetOpcode::G_FCOSH:
1781 case TargetOpcode::G_FSINH:
1782 case TargetOpcode::G_FTANH:
1783 case TargetOpcode::G_FSQRT:
1784 case TargetOpcode::G_FSUB:
1785 case TargetOpcode::G_INTRINSIC_ROUND:
1786 case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
1787 case TargetOpcode::G_INTRINSIC_TRUNC:
1788 return true;
1789 default:
1790 return false;
1791 }
1792}
1793
1794/// Shifts return poison if shiftwidth is larger than the bitwidth.
1795static bool shiftAmountKnownInRange(Register ShiftAmount,
1796 const MachineRegisterInfo &MRI) {
1797 LLT Ty = MRI.getType(ShiftAmount);
1798
1799 if (Ty.isScalableVector())
1800 return false; // Can't tell, just return false to be safe
1801
1802 if (Ty.isScalar()) {
1803 std::optional<ValueAndVReg> Val =
1805 if (!Val)
1806 return false;
1807 return Val->Value.ult(Ty.getScalarSizeInBits());
1808 }
1809
1810 GBuildVector *BV = getOpcodeDef<GBuildVector>(ShiftAmount, MRI);
1811 if (!BV)
1812 return false;
1813
1814 unsigned Sources = BV->getNumSources();
1815 for (unsigned I = 0; I < Sources; ++I) {
1816 std::optional<ValueAndVReg> Val =
1818 if (!Val)
1819 return false;
1820 if (!Val->Value.ult(Ty.getScalarSizeInBits()))
1821 return false;
1822 }
1823
1824 return true;
1825}
1826
1827namespace {
1828enum class UndefPoisonKind {
1829 PoisonOnly = (1 << 0),
1830 UndefOnly = (1 << 1),
1832};
1833}
1834
1836 return (unsigned(Kind) & unsigned(UndefPoisonKind::PoisonOnly)) != 0;
1837}
1838
1840 return (unsigned(Kind) & unsigned(UndefPoisonKind::UndefOnly)) != 0;
1841}
1842
1844 bool ConsiderFlagsAndMetadata,
1845 UndefPoisonKind Kind) {
1846 MachineInstr *RegDef = MRI.getVRegDef(Reg);
1847
1848 if (ConsiderFlagsAndMetadata && includesPoison(Kind))
1849 if (auto *GMI = dyn_cast<GenericMachineInstr>(RegDef))
1850 if (GMI->hasPoisonGeneratingFlags())
1851 return true;
1852
1853 // Check whether opcode is a poison/undef-generating operation.
1854 switch (RegDef->getOpcode()) {
1855 case TargetOpcode::G_BUILD_VECTOR:
1856 case TargetOpcode::G_CONSTANT_FOLD_BARRIER:
1857 return false;
1858 case TargetOpcode::G_SHL:
1859 case TargetOpcode::G_ASHR:
1860 case TargetOpcode::G_LSHR:
1861 return includesPoison(Kind) &&
1863 case TargetOpcode::G_FPTOSI:
1864 case TargetOpcode::G_FPTOUI:
1865 // fptosi/ui yields poison if the resulting value does not fit in the
1866 // destination type.
1867 return true;
1868 case TargetOpcode::G_CTLZ:
1869 case TargetOpcode::G_CTTZ:
1870 case TargetOpcode::G_ABS:
1871 case TargetOpcode::G_CTPOP:
1872 case TargetOpcode::G_BSWAP:
1873 case TargetOpcode::G_BITREVERSE:
1874 case TargetOpcode::G_FSHL:
1875 case TargetOpcode::G_FSHR:
1876 case TargetOpcode::G_SMAX:
1877 case TargetOpcode::G_SMIN:
1878 case TargetOpcode::G_SCMP:
1879 case TargetOpcode::G_UMAX:
1880 case TargetOpcode::G_UMIN:
1881 case TargetOpcode::G_UCMP:
1882 case TargetOpcode::G_PTRMASK:
1883 case TargetOpcode::G_SADDO:
1884 case TargetOpcode::G_SSUBO:
1885 case TargetOpcode::G_UADDO:
1886 case TargetOpcode::G_USUBO:
1887 case TargetOpcode::G_SMULO:
1888 case TargetOpcode::G_UMULO:
1889 case TargetOpcode::G_SADDSAT:
1890 case TargetOpcode::G_UADDSAT:
1891 case TargetOpcode::G_SSUBSAT:
1892 case TargetOpcode::G_USUBSAT:
1893 return false;
1894 case TargetOpcode::G_SSHLSAT:
1895 case TargetOpcode::G_USHLSAT:
1896 return includesPoison(Kind) &&
1898 case TargetOpcode::G_INSERT_VECTOR_ELT: {
1899 GInsertVectorElement *Insert = cast<GInsertVectorElement>(RegDef);
1900 if (includesPoison(Kind)) {
1901 std::optional<ValueAndVReg> Index =
1902 getIConstantVRegValWithLookThrough(Insert->getIndexReg(), MRI);
1903 if (!Index)
1904 return true;
1905 LLT VecTy = MRI.getType(Insert->getVectorReg());
1906 return Index->Value.uge(VecTy.getElementCount().getKnownMinValue());
1907 }
1908 return false;
1909 }
1910 case TargetOpcode::G_EXTRACT_VECTOR_ELT: {
1911 GExtractVectorElement *Extract = cast<GExtractVectorElement>(RegDef);
1912 if (includesPoison(Kind)) {
1913 std::optional<ValueAndVReg> Index =
1915 if (!Index)
1916 return true;
1917 LLT VecTy = MRI.getType(Extract->getVectorReg());
1918 return Index->Value.uge(VecTy.getElementCount().getKnownMinValue());
1919 }
1920 return false;
1921 }
1922 case TargetOpcode::G_SHUFFLE_VECTOR: {
1923 GShuffleVector *Shuffle = cast<GShuffleVector>(RegDef);
1924 ArrayRef<int> Mask = Shuffle->getMask();
1925 return includesPoison(Kind) && is_contained(Mask, -1);
1926 }
1927 case TargetOpcode::G_FNEG:
1928 case TargetOpcode::G_PHI:
1929 case TargetOpcode::G_SELECT:
1930 case TargetOpcode::G_UREM:
1931 case TargetOpcode::G_SREM:
1932 case TargetOpcode::G_FREEZE:
1933 case TargetOpcode::G_ICMP:
1934 case TargetOpcode::G_FCMP:
1935 case TargetOpcode::G_FADD:
1936 case TargetOpcode::G_FSUB:
1937 case TargetOpcode::G_FMUL:
1938 case TargetOpcode::G_FDIV:
1939 case TargetOpcode::G_FREM:
1940 case TargetOpcode::G_PTR_ADD:
1941 return false;
1942 default:
1943 return !isa<GCastOp>(RegDef) && !isa<GBinOp>(RegDef);
1944 }
1945}
1946
1948 const MachineRegisterInfo &MRI,
1949 unsigned Depth,
1950 UndefPoisonKind Kind) {
1952 return false;
1953
1954 MachineInstr *RegDef = MRI.getVRegDef(Reg);
1955
1956 switch (RegDef->getOpcode()) {
1957 case TargetOpcode::G_FREEZE:
1958 return true;
1959 case TargetOpcode::G_IMPLICIT_DEF:
1960 return !includesUndef(Kind);
1961 case TargetOpcode::G_CONSTANT:
1962 case TargetOpcode::G_FCONSTANT:
1963 return true;
1964 case TargetOpcode::G_BUILD_VECTOR: {
1965 GBuildVector *BV = cast<GBuildVector>(RegDef);
1966 unsigned NumSources = BV->getNumSources();
1967 for (unsigned I = 0; I < NumSources; ++I)
1969 Depth + 1, Kind))
1970 return false;
1971 return true;
1972 }
1973 case TargetOpcode::G_PHI: {
1974 GPhi *Phi = cast<GPhi>(RegDef);
1975 unsigned NumIncoming = Phi->getNumIncomingValues();
1976 for (unsigned I = 0; I < NumIncoming; ++I)
1977 if (!::isGuaranteedNotToBeUndefOrPoison(Phi->getIncomingValue(I), MRI,
1978 Depth + 1, Kind))
1979 return false;
1980 return true;
1981 }
1982 default: {
1983 auto MOCheck = [&](const MachineOperand &MO) {
1984 if (!MO.isReg())
1985 return true;
1986 return ::isGuaranteedNotToBeUndefOrPoison(MO.getReg(), MRI, Depth + 1,
1987 Kind);
1988 };
1989 return !::canCreateUndefOrPoison(Reg, MRI,
1990 /*ConsiderFlagsAndMetadata=*/true, Kind) &&
1991 all_of(RegDef->uses(), MOCheck);
1992 }
1993 }
1994}
1995
1997 bool ConsiderFlagsAndMetadata) {
1998 return ::canCreateUndefOrPoison(Reg, MRI, ConsiderFlagsAndMetadata,
2000}
2001
2003 bool ConsiderFlagsAndMetadata = true) {
2004 return ::canCreateUndefOrPoison(Reg, MRI, ConsiderFlagsAndMetadata,
2006}
2007
2009 const MachineRegisterInfo &MRI,
2010 unsigned Depth) {
2011 return ::isGuaranteedNotToBeUndefOrPoison(Reg, MRI, Depth,
2013}
2014
2016 const MachineRegisterInfo &MRI,
2017 unsigned Depth) {
2018 return ::isGuaranteedNotToBeUndefOrPoison(Reg, MRI, Depth,
2020}
2021
2023 const MachineRegisterInfo &MRI,
2024 unsigned Depth) {
2025 return ::isGuaranteedNotToBeUndefOrPoison(Reg, MRI, Depth,
2027}
2028
2030 if (Ty.isVector())
2031 return VectorType::get(IntegerType::get(C, Ty.getScalarSizeInBits()),
2032 Ty.getElementCount());
2033 return IntegerType::get(C, Ty.getSizeInBits());
2034}
2035
2037 switch (MI.getOpcode()) {
2038 default:
2039 return false;
2040 case TargetOpcode::G_ASSERT_ALIGN:
2041 case TargetOpcode::G_ASSERT_SEXT:
2042 case TargetOpcode::G_ASSERT_ZEXT:
2043 return true;
2044 }
2045}
2046
2048 assert(Kind == GIConstantKind::Scalar && "Expected scalar constant");
2049
2050 return Value;
2051}
2052
2053std::optional<GIConstant>
2056
2057 if (GSplatVector *Splat = dyn_cast<GSplatVector>(Constant)) {
2058 std::optional<ValueAndVReg> MayBeConstant =
2060 if (!MayBeConstant)
2061 return std::nullopt;
2062 return GIConstant(MayBeConstant->Value, GIConstantKind::ScalableVector);
2063 }
2064
2065 if (GBuildVector *Build = dyn_cast<GBuildVector>(Constant)) {
2066 SmallVector<APInt> Values;
2067 unsigned NumSources = Build->getNumSources();
2068 for (unsigned I = 0; I < NumSources; ++I) {
2069 Register SrcReg = Build->getSourceReg(I);
2070 std::optional<ValueAndVReg> MayBeConstant =
2072 if (!MayBeConstant)
2073 return std::nullopt;
2074 Values.push_back(MayBeConstant->Value);
2075 }
2076 return GIConstant(Values);
2077 }
2078
2079 std::optional<ValueAndVReg> MayBeConstant =
2081 if (!MayBeConstant)
2082 return std::nullopt;
2083
2084 return GIConstant(MayBeConstant->Value, GIConstantKind::Scalar);
2085}
2086
2088 assert(Kind == GFConstantKind::Scalar && "Expected scalar constant");
2089
2090 return Values[0];
2091}
2092
2093std::optional<GFConstant>
2096
2097 if (GSplatVector *Splat = dyn_cast<GSplatVector>(Constant)) {
2098 std::optional<FPValueAndVReg> MayBeConstant =
2100 if (!MayBeConstant)
2101 return std::nullopt;
2102 return GFConstant(MayBeConstant->Value, GFConstantKind::ScalableVector);
2103 }
2104
2105 if (GBuildVector *Build = dyn_cast<GBuildVector>(Constant)) {
2106 SmallVector<APFloat> Values;
2107 unsigned NumSources = Build->getNumSources();
2108 for (unsigned I = 0; I < NumSources; ++I) {
2109 Register SrcReg = Build->getSourceReg(I);
2110 std::optional<FPValueAndVReg> MayBeConstant =
2112 if (!MayBeConstant)
2113 return std::nullopt;
2114 Values.push_back(MayBeConstant->Value);
2115 }
2116 return GFConstant(Values);
2117 }
2118
2119 std::optional<FPValueAndVReg> MayBeConstant =
2121 if (!MayBeConstant)
2122 return std::nullopt;
2123
2124 return GFConstant(MayBeConstant->Value, GFConstantKind::Scalar);
2125}
unsigned const MachineRegisterInfo * MRI
MachineInstrBuilder MachineInstrBuilder & DefMI
unsigned RegSize
assert(UImm &&(UImm !=~static_cast< T >(0)) &&"Invalid immediate!")
This file declares a class to represent arbitrary precision floating point values and provide a varie...
This file implements a class to represent arbitrary precision integral constant values and operations...
MachineBasicBlock & MBB
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static bool canCreateUndefOrPoison(Register Reg, const MachineRegisterInfo &MRI, bool ConsiderFlagsAndMetadata, UndefPoisonKind Kind)
Definition: Utils.cpp:1843
static bool isGuaranteedNotToBeUndefOrPoison(Register Reg, const MachineRegisterInfo &MRI, unsigned Depth, UndefPoisonKind Kind)
Definition: Utils.cpp:1947
static bool includesPoison(UndefPoisonKind Kind)
Definition: Utils.cpp:1835
static bool includesUndef(UndefPoisonKind Kind)
Definition: Utils.cpp:1839
static void reportGISelDiagnostic(DiagnosticSeverity Severity, MachineFunction &MF, const TargetPassConfig &TPC, MachineOptimizationRemarkEmitter &MORE, MachineOptimizationRemarkMissed &R)
Definition: Utils.cpp:235
static bool shiftAmountKnownInRange(Register ShiftAmount, const MachineRegisterInfo &MRI)
Shifts return poison if shiftwidth is larger than the bitwidth.
Definition: Utils.cpp:1795
bool canCreatePoison(Register Reg, const MachineRegisterInfo &MRI, bool ConsiderFlagsAndMetadata=true)
Definition: Utils.cpp:2002
static bool isBuildVectorOp(unsigned Opcode)
Definition: Utils.cpp:1352
static bool isConstantScalar(const MachineInstr &MI, const MachineRegisterInfo &MRI, bool AllowFP=true, bool AllowOpaqueConstants=true)
Definition: Utils.cpp:1506
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
uint64_t Size
This contains common code to allow clients to notify changes to machine instr.
Provides analysis for querying information about KnownBits during GISel passes.
Declares convenience wrapper classes for interpreting MachineInstr instances as specific generic oper...
const HexagonInstrInfo * TII
IRTranslator LLVM IR MI
Tracks DebugLocs between checkpoints and verifies that they are transferred.
#define I(x, y, z)
Definition: MD5.cpp:58
Contains matchers for matching SSA Machine Instructions.
This file declares the MachineIRBuilder class.
===- MachineOptimizationRemarkEmitter.h - Opt Diagnostics -*- C++ -*-—===//
Register const TargetRegisterInfo * TRI
MachineInstr unsigned OpIdx
uint64_t IntrinsicInst * II
#define LLVM_DEBUG(...)
Definition: Debug.h:119
This file describes how to lower LLVM code to machine code.
Target-Independent Code Generator Pass Configuration Options pass.
UndefPoisonKind
static const char PassName[]
Value * RHS
Value * LHS
BinaryOperator * Mul
Class recording the (high level) value of a variable.
opStatus divide(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1208
void copySign(const APFloat &RHS)
Definition: APFloat.h:1302
LLVM_ABI opStatus convert(const fltSemantics &ToSemantics, roundingMode RM, bool *losesInfo)
Definition: APFloat.cpp:6057
opStatus subtract(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1190
opStatus add(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1181
opStatus convertFromAPInt(const APInt &Input, bool IsSigned, roundingMode RM)
Definition: APFloat.h:1347
opStatus multiply(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1199
APInt bitcastToAPInt() const
Definition: APFloat.h:1353
opStatus mod(const APFloat &RHS)
Definition: APFloat.h:1226
Class for arbitrary precision integers.
Definition: APInt.h:78
LLVM_ABI APInt udiv(const APInt &RHS) const
Unsigned division operation.
Definition: APInt.cpp:1573
static APInt getAllOnes(unsigned numBits)
Return an APInt of a specified width with all bits set.
Definition: APInt.h:234
LLVM_ABI APInt zext(unsigned width) const
Zero extend to a new width.
Definition: APInt.cpp:1012
LLVM_ABI APInt zextOrTrunc(unsigned width) const
Zero extend or truncate to width.
Definition: APInt.cpp:1033
LLVM_ABI APInt trunc(unsigned width) const
Truncate to new width.
Definition: APInt.cpp:936
LLVM_ABI APInt urem(const APInt &RHS) const
Unsigned remainder operation.
Definition: APInt.cpp:1666
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1488
LLVM_ABI APInt sdiv(const APInt &RHS) const
Signed division function for APInt.
Definition: APInt.cpp:1644
LLVM_ABI APInt sextOrTrunc(unsigned width) const
Sign extend or truncate to width.
Definition: APInt.cpp:1041
APInt ashr(unsigned ShiftAmt) const
Arithmetic right-shift function.
Definition: APInt.h:827
LLVM_ABI APInt srem(const APInt &RHS) const
Function for signed remainder operation.
Definition: APInt.cpp:1736
LLVM_ABI APInt sext(unsigned width) const
Sign extend to a new width.
Definition: APInt.cpp:985
bool isPowerOf2() const
Check if this APInt's value is a power of two greater than zero.
Definition: APInt.h:440
static bool isSameValue(const APInt &I1, const APInt &I2)
Determine if two APInts have the same value, after zero-extending one of them (if needed!...
Definition: APInt.h:553
static APInt getZero(unsigned numBits)
Get the '0' value for the specified bit-width.
Definition: APInt.h:200
static APInt getOneBitSet(unsigned numBits, unsigned BitNo)
Return an APInt with exactly one bit set in the result.
Definition: APInt.h:239
APInt lshr(unsigned shiftAmt) const
Logical right-shift function.
Definition: APInt.h:851
Represent the analysis usage information of a pass.
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition: ArrayRef.h:41
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:277
const APFloat & getValueAPF() const
Definition: Constants.h:320
bool isNegative() const
Return true if the sign bit is set.
Definition: Constants.h:327
bool isZero() const
Return true if the value is positive or negative zero.
Definition: Constants.h:324
This is the shared class of boolean and integer constants.
Definition: Constants.h:87
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:154
This is an important base class in LLVM.
Definition: Constant.h:43
This class represents an Operation in the Expression.
A debug info location.
Definition: DebugLoc.h:124
static constexpr ElementCount getFixed(ScalarTy MinVal)
Definition: TypeSize.h:312
static constexpr ElementCount get(ScalarTy MinVal, bool Scalable)
Definition: TypeSize.h:318
Represents a G_BUILD_VECTOR.
Represents an extract vector element.
An floating-point-like constant.
Definition: Utils.h:689
static LLVM_ABI std::optional< GFConstant > getConstant(Register Const, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:2094
LLVM_ABI APFloat getScalarValue() const
Returns the value, if this constant is a scalar.
Definition: Utils.cpp:2087
An integer-like constant.
Definition: Utils.h:650
LLVM_ABI APInt getScalarValue() const
Returns the value, if this constant is a scalar.
Definition: Utils.cpp:2047
static LLVM_ABI std::optional< GIConstant > getConstant(Register Const, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:2054
Abstract class that contains various methods for clients to notify about changes.
KnownBits getKnownBits(Register R)
void insert(MachineInstr *I)
Add the specified instruction to the worklist if it isn't already in it.
Definition: GISelWorkList.h:74
MachineInstr * pop_back_val()
bool empty() const
Definition: GISelWorkList.h:38
void remove(const MachineInstr *I)
Remove I from the worklist if it exists.
Definition: GISelWorkList.h:83
Represents an insert vector element.
Register getSourceReg(unsigned I) const
Returns the I'th source register.
unsigned getNumSources() const
Returns the number of source registers.
Represents a G_PHI.
Represents a G_SHUFFLE_VECTOR.
ArrayRef< int > getMask() const
Represents a splat vector.
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:663
static LLVM_ABI IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition: Type.cpp:319
constexpr bool isScalableVector() const
Returns true if the LLT is a scalable vector.
Definition: LowLevelType.h:182
constexpr unsigned getScalarSizeInBits() const
Definition: LowLevelType.h:265
constexpr bool isScalar() const
Definition: LowLevelType.h:147
static constexpr LLT vector(ElementCount EC, unsigned ScalarSizeInBits)
Get a low-level vector of some number of elements and element width.
Definition: LowLevelType.h:65
static constexpr LLT scalar(unsigned SizeInBits)
Get a low-level scalar or aggregate "bag of bits".
Definition: LowLevelType.h:43
constexpr bool isValid() const
Definition: LowLevelType.h:146
constexpr uint16_t getNumElements() const
Returns the number of elements in a vector LLT.
Definition: LowLevelType.h:160
constexpr bool isVector() const
Definition: LowLevelType.h:149
constexpr bool isScalable() const
Returns true if the LLT is a scalable vector.
Definition: LowLevelType.h:171
constexpr TypeSize getSizeInBits() const
Returns the total size of the type. Must only be called on sized types.
Definition: LowLevelType.h:191
constexpr LLT getElementType() const
Returns the vector's element type. Only valid for vector types.
Definition: LowLevelType.h:278
constexpr ElementCount getElementCount() const
Definition: LowLevelType.h:184
static constexpr LLT fixed_vector(unsigned NumElements, unsigned ScalarSizeInBits)
Get a low-level fixed-width vector of some number of elements and element width.
Definition: LowLevelType.h:101
constexpr bool isFixedVector() const
Returns true if the LLT is a fixed vector.
Definition: LowLevelType.h:178
constexpr LLT getScalarType() const
Definition: LowLevelType.h:206
static constexpr LLT scalarOrVector(ElementCount EC, LLT ScalarTy)
Definition: LowLevelType.h:125
This is an important class for using LLVM in a threaded context.
Definition: LLVMContext.h:68
void checkpoint(bool CheckDebugLocs=true)
Call this to indicate that it's a good point to assess whether locations have been lost.
Describe properties that are true of each instruction in the target description file.
Definition: MCInstrDesc.h:199
Wrapper class representing physical registers. Should be passed by value.
Definition: MCRegister.h:33
void addLiveIn(MCRegister PhysReg, LaneBitmask LaneMask=LaneBitmask::getAll())
Adds the specified register as a live in.
const MachineFunction * getParent() const
Return the MachineFunction containing this basic block.
LLVM_ABI bool isLiveIn(MCRegister Reg, LaneBitmask LaneMask=LaneBitmask::getAll()) const
Return true if the specified register is in the live in set.
The MachineFrameInfo class represents an abstract stack frame until prolog/epilog code is inserted.
Align getObjectAlign(int ObjectIdx) const
Return the alignment of the specified stack object.
StringRef getName() const
getName - Return the name of the corresponding LLVM function.
GISelChangeObserver * getObserver() const
MachineFrameInfo & getFrameInfo()
getFrameInfo - Return the frame info object for the current function.
MachineRegisterInfo & getRegInfo()
getRegInfo - Return information about the registers currently in use.
Function & getFunction()
Return the LLVM function that this machine code represents.
const MachineFunctionProperties & getProperties() const
Get the function properties.
const MachineBasicBlock & front() const
Register addLiveIn(MCRegister PReg, const TargetRegisterClass *RC)
addLiveIn - Add the specified physical register as a live-in value and create a corresponding virtual...
const TargetMachine & getTarget() const
getTarget - Return the target machine this machine code is compiled with
Helper class to build MachineInstr.
MachineInstrBuilder buildUnmerge(ArrayRef< LLT > Res, const SrcOp &Op)
Build and insert Res0, ... = G_UNMERGE_VALUES Op.
MachineInstrBuilder buildExtract(const DstOp &Res, const SrcOp &Src, uint64_t Index)
Build and insert Res0, ... = G_EXTRACT Src, Idx0.
MachineInstrBuilder buildMergeLikeInstr(const DstOp &Res, ArrayRef< Register > Ops)
Build and insert Res = G_MERGE_VALUES Op0, ... or Res = G_BUILD_VECTOR Op0, ... or Res = G_CONCAT_VEC...
Register getReg(unsigned Idx) const
Get the register for the operand index.
const MachineInstrBuilder & addReg(Register RegNo, unsigned flags=0, unsigned SubReg=0) const
Add a new virtual register operand.
Representation of each machine instruction.
Definition: MachineInstr.h:72
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition: MachineInstr.h:587
const MachineBasicBlock * getParent() const
Definition: MachineInstr.h:359
bool getFlag(MIFlag Flag) const
Return whether an MI flag is set.
Definition: MachineInstr.h:409
mop_range uses()
Returns all operands which may be register uses.
Definition: MachineInstr.h:731
LLVM_ABI const MachineFunction * getMF() const
Return the function that contains the basic block that this instruction belongs to.
const DebugLoc & getDebugLoc() const
Returns the debug location id of this MachineInstr.
Definition: MachineInstr.h:511
const MachineOperand & getOperand(unsigned i) const
Definition: MachineInstr.h:595
MachineOperand class - Representation of each machine instruction operand.
const ConstantInt * getCImm() const
bool isCImm() const
isCImm - Test if this is a MO_CImmediate operand.
bool isReg() const
isReg - Tests if this is a MO_Register operand.
LLVM_ABI void setReg(Register Reg)
Change the register this operand corresponds to.
MachineInstr * getParent()
getParent - Return the instruction that this operand belongs to.
Register getReg() const
getReg - Returns the register number.
const ConstantFP * getFPImm() const
bool isFPImm() const
isFPImm - Tests if this is a MO_FPImmediate operand.
Diagnostic information for missed-optimization remarks.
MachineRegisterInfo - Keep track of information for virtual and physical registers,...
A Module instance is used to store all the information related to an LLVM module.
Definition: Module.h:67
Represents a value which can be a Register or a constant.
Definition: Utils.h:407
Holds all the information related to register banks.
static const TargetRegisterClass * constrainGenericRegister(Register Reg, const TargetRegisterClass &RC, MachineRegisterInfo &MRI)
Constrain the (possibly generic) virtual register Reg to RC.
Wrapper class representing virtual and physical registers.
Definition: Register.h:19
constexpr bool isPhysical() const
Return true if the specified register number is in the physical register namespace.
Definition: Register.h:78
bool empty() const
Definition: SmallVector.h:82
size_t size() const
Definition: SmallVector.h:79
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: SmallVector.h:574
reference emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:938
void push_back(const T &Elt)
Definition: SmallVector.h:414
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1197
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:55
TargetInstrInfo - Interface to description of machine instruction set.
BooleanContent getBooleanContents(bool isVec, bool isFloat) const
For targets without i1 registers, this gives the nature of the high-bits of boolean values held in ty...
This class defines information used to lower LLVM code to legal SelectionDAG operators that the targe...
Primary interface to the complete machine description for the target machine.
Definition: TargetMachine.h:83
Target-Independent Code Generator Pass Configuration Options.
bool isGlobalISelAbortEnabled() const
Check whether or not GlobalISel should abort on error.
TargetRegisterInfo base class - We assume that the target defines a static array of TargetRegisterDes...
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:82
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
LLVM Value Representation.
Definition: Value.h:75
constexpr ScalarTy getFixedValue() const
Definition: TypeSize.h:203
constexpr bool isScalable() const
Returns whether the quantity is scaled by a runtime quantity (vscale).
Definition: TypeSize.h:172
constexpr LeafTy multiplyCoefficientBy(ScalarTy RHS) const
Definition: TypeSize.h:259
constexpr ScalarTy getKnownMinValue() const
Returns the minimum value this quantity can represent.
Definition: TypeSize.h:169
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
const APInt & smin(const APInt &A, const APInt &B)
Determine the smaller of two APInts considered to be signed.
Definition: APInt.h:2248
const APInt & smax(const APInt &A, const APInt &B)
Determine the larger of two APInts considered to be signed.
Definition: APInt.h:2253
const APInt & umin(const APInt &A, const APInt &B)
Determine the smaller of two APInts considered to be unsigned.
Definition: APInt.h:2258
const APInt & umax(const APInt &A, const APInt &B)
Determine the larger of two APInts considered to be unsigned.
Definition: APInt.h:2263
@ C
The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
DiagnosticInfoMIROptimization::MachineArgument MNV
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
LLVM_ABI Register getFunctionLiveInPhysReg(MachineFunction &MF, const TargetInstrInfo &TII, MCRegister PhysReg, const TargetRegisterClass &RC, const DebugLoc &DL, LLT RegTy=LLT())
Return a virtual register corresponding to the incoming argument register PhysReg.
Definition: Utils.cpp:916
auto drop_begin(T &&RangeOrContainer, size_t N=1)
Return a range covering RangeOrContainer with the first N elements excluded.
Definition: STLExtras.h:338
LLVM_ABI std::optional< SmallVector< APInt > > ConstantFoldICmp(unsigned Pred, const Register Op1, const Register Op2, unsigned DstScalarSizeInBits, unsigned ExtOp, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:1035
@ Offset
Definition: DWP.cpp:477
LLVM_ABI bool isBuildVectorAllZeros(const MachineInstr &MI, const MachineRegisterInfo &MRI, bool AllowUndef=false)
Return true if the specified instruction is a G_BUILD_VECTOR or G_BUILD_VECTOR_TRUNC where all of the...
Definition: Utils.cpp:1480
LLVM_ABI Type * getTypeForLLT(LLT Ty, LLVMContext &C)
Get the type back from LLT.
Definition: Utils.cpp:2029
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1744
LLVM_ABI Register constrainOperandRegClass(const MachineFunction &MF, const TargetRegisterInfo &TRI, MachineRegisterInfo &MRI, const TargetInstrInfo &TII, const RegisterBankInfo &RBI, MachineInstr &InsertPt, const TargetRegisterClass &RegClass, MachineOperand &RegMO)
Constrain the Register operand OpIdx, so that it is now constrained to the TargetRegisterClass passed...
Definition: Utils.cpp:56
LLVM_ABI MachineInstr * getOpcodeDef(unsigned Opcode, Register Reg, const MachineRegisterInfo &MRI)
See if Reg is defined by an single def instruction that is Opcode.
Definition: Utils.cpp:651
LLVM_ABI const ConstantFP * getConstantFPVRegVal(Register VReg, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:459
MachineInstrBuilder BuildMI(MachineFunction &MF, const MIMetadata &MIMD, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
LLVM_ABI std::optional< APInt > getIConstantVRegVal(Register VReg, const MachineRegisterInfo &MRI)
If VReg is defined by a G_CONSTANT, return the corresponding value.
Definition: Utils.cpp:294
LLVM_ABI std::optional< APFloat > ConstantFoldIntToFloat(unsigned Opcode, LLT DstTy, Register Src, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:990
LLVM_ABI std::optional< APInt > getIConstantSplatVal(const Register Reg, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:1440
LLVM_ABI bool isAllOnesOrAllOnesSplat(const MachineInstr &MI, const MachineRegisterInfo &MRI, bool AllowUndefs=false)
Return true if the value is a constant -1 integer or a splatted vector of a constant -1 integer (with...
Definition: Utils.cpp:1605
LLVM_ABI const llvm::fltSemantics & getFltSemanticForLLT(LLT Ty)
Get the appropriate floating point arithmetic semantic based on the bit size of the given scalar LLT.
LLVM_ABI std::optional< APFloat > ConstantFoldFPBinOp(unsigned Opcode, const Register Op1, const Register Op2, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:739
LLVM_ABI void salvageDebugInfo(const MachineRegisterInfo &MRI, MachineInstr &MI)
Assuming the instruction MI is going to be deleted, attempt to salvage debug users of MI by writing t...
Definition: Utils.cpp:1723
LLVM_ABI bool constrainSelectedInstRegOperands(MachineInstr &I, const TargetInstrInfo &TII, const TargetRegisterInfo &TRI, const RegisterBankInfo &RBI)
Mutate the newly-selected instruction I to constrain its (possibly generic) virtual register operands...
Definition: Utils.cpp:155
bool isPreISelGenericOpcode(unsigned Opcode)
Check whether the given Opcode is a generic opcode that is not supposed to appear after ISel.
Definition: TargetOpcodes.h:30
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
LLVM_ABI std::optional< SmallVector< unsigned > > ConstantFoldCountZeros(Register Src, const MachineRegisterInfo &MRI, std::function< unsigned(APInt)> CB)
Tries to constant fold a counting-zero operation (G_CTLZ or G_CTTZ) on Src.
Definition: Utils.cpp:1003
LLVM_ABI std::optional< APInt > ConstantFoldExtOp(unsigned Opcode, const Register Op1, uint64_t Imm, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:949
LLVM_ABI std::optional< RegOrConstant > getVectorSplat(const MachineInstr &MI, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:1493
LLVM_READONLY APFloat maximum(const APFloat &A, const APFloat &B)
Implements IEEE 754-2019 maximum semantics.
Definition: APFloat.h:1643
LLVM_ABI std::optional< APInt > isConstantOrConstantSplatVector(MachineInstr &MI, const MachineRegisterInfo &MRI)
Determines if MI defines a constant integer or a splat vector of constant integers.
Definition: Utils.cpp:1563
LLVM_ABI bool isNullOrNullSplat(const MachineInstr &MI, const MachineRegisterInfo &MRI, bool AllowUndefs=false)
Return true if the value is a constant 0 integer or a splatted vector of a constant 0 integer (with n...
Definition: Utils.cpp:1587
LLVM_ABI MachineInstr * getDefIgnoringCopies(Register Reg, const MachineRegisterInfo &MRI)
Find the def instruction for Reg, folding away any trivial copies.
Definition: Utils.cpp:492
LLVM_ABI bool matchUnaryPredicate(const MachineRegisterInfo &MRI, Register Reg, std::function< bool(const Constant *ConstVal)> Match, bool AllowUndefs=false)
Attempt to match a unary predicate against a scalar/splat constant or every element of a constant G_B...
Definition: Utils.cpp:1620
bool isPreISelGenericOptimizationHint(unsigned Opcode)
Definition: TargetOpcodes.h:42
LLVM_ABI bool isGuaranteedNotToBeUndef(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Returns true if V cannot be undef, but may be poison.
LLVM_ABI bool isConstTrueVal(const TargetLowering &TLI, int64_t Val, bool IsVector, bool IsFP)
Returns true if given the TargetLowering's boolean contents information, the value Val contains a tru...
Definition: Utils.cpp:1652
LLVM_ABI LLVM_READNONE LLT getLCMType(LLT OrigTy, LLT TargetTy)
Return the least common multiple type of OrigTy and TargetTy, by changing the number of vector elemen...
Definition: Utils.cpp:1189
LLVM_ABI std::optional< int64_t > getIConstantVRegSExtVal(Register VReg, const MachineRegisterInfo &MRI)
If VReg is defined by a G_CONSTANT fits in int64_t returns it.
Definition: Utils.cpp:314
LLVM_ABI std::optional< APInt > ConstantFoldBinOp(unsigned Opcode, const Register Op1, const Register Op2, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:670
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1751
LLVM_ABI const APInt & getIConstantFromReg(Register VReg, const MachineRegisterInfo &MRI)
VReg is defined by a G_CONSTANT, return the corresponding value.
Definition: Utils.cpp:305
LLVM_READONLY APFloat maxnum(const APFloat &A, const APFloat &B)
Implements IEEE-754 2008 maxNum semantics.
Definition: APFloat.h:1598
LLVM_ABI bool isConstantOrConstantVector(const MachineInstr &MI, const MachineRegisterInfo &MRI, bool AllowFP=true, bool AllowOpaqueConstants=true)
Return true if the specified instruction is known to be a constant, or a vector of constants.
Definition: Utils.cpp:1543
constexpr unsigned MaxAnalysisRecursionDepth
Definition: ValueTracking.h:47
auto reverse(ContainerTy &&C)
Definition: STLExtras.h:428
LLVM_ABI bool canReplaceReg(Register DstReg, Register SrcReg, MachineRegisterInfo &MRI)
Check if DstReg can be replaced with SrcReg depending on the register constraints.
Definition: Utils.cpp:201
LLVM_ABI void saveUsesAndErase(MachineInstr &MI, MachineRegisterInfo &MRI, LostDebugLocObserver *LocObserver, SmallInstListTy &DeadInstChain)
Definition: Utils.cpp:1689
LLVM_ABI void reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, MachineOptimizationRemarkEmitter &MORE, MachineOptimizationRemarkMissed &R)
Report an ISel error as a missed optimization remark to the LLVMContext's diagnostic stream.
Definition: Utils.cpp:259
LLVM_ABI raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:207
LLVM_ABI std::optional< ValueAndVReg > getAnyConstantVRegValWithLookThrough(Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs=true, bool LookThroughAnyExt=false)
If VReg is defined by a statically evaluable chain of instructions rooted on a G_CONSTANT or G_FCONST...
Definition: Utils.cpp:439
LLVM_ABI bool isBuildVectorAllOnes(const MachineInstr &MI, const MachineRegisterInfo &MRI, bool AllowUndef=false)
Return true if the specified instruction is a G_BUILD_VECTOR or G_BUILD_VECTOR_TRUNC where all of the...
Definition: Utils.cpp:1486
LLVM_ABI bool canCreateUndefOrPoison(const Operator *Op, bool ConsiderFlagsAndMetadata=true)
canCreateUndefOrPoison returns true if Op can create undef or poison from non-undef & non-poison oper...
LLVM_ABI SmallVector< APInt > ConstantFoldVectorBinop(unsigned Opcode, const Register Op1, const Register Op2, const MachineRegisterInfo &MRI)
Tries to constant fold a vector binop with sources Op1 and Op2.
Definition: Utils.cpp:793
LLVM_ABI std::optional< FPValueAndVReg > getFConstantSplat(Register VReg, const MachineRegisterInfo &MRI, bool AllowUndef=true)
Returns a floating point scalar constant of a build vector splat if it exists.
Definition: Utils.cpp:1473
LLVM_ABI std::optional< APInt > ConstantFoldCastOp(unsigned Opcode, LLT DstTy, const Register Op0, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:966
LLVM_ABI void extractParts(Register Reg, LLT Ty, int NumParts, SmallVectorImpl< Register > &VRegs, MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI)
Helper function to split a wide generic register into bitwise blocks with the given Type (which impli...
Definition: Utils.cpp:506
LLVM_ABI void getSelectionDAGFallbackAnalysisUsage(AnalysisUsage &AU)
Modify analysis usage so it preserves passes required for the SelectionDAG fallback.
Definition: Utils.cpp:1185
LLVM_ABI LLVM_READNONE LLT getCoverTy(LLT OrigTy, LLT TargetTy)
Return smallest type that covers both OrigTy and TargetTy and is multiple of TargetTy.
Definition: Utils.cpp:1256
LLVM_READONLY APFloat minnum(const APFloat &A, const APFloat &B)
Implements IEEE-754 2008 minNum semantics.
Definition: APFloat.h:1579
LLVM_ABI unsigned getInverseGMinMaxOpcode(unsigned MinMaxOpc)
Returns the inverse opcode of MinMaxOpc, which is a generic min/max opcode like G_SMIN.
Definition: Utils.cpp:279
uint64_t alignTo(uint64_t Size, Align A)
Returns a multiple of A needed to store Size bytes.
Definition: Alignment.h:155
bool isTargetSpecificOpcode(unsigned Opcode)
Check whether the given Opcode is a target-specific opcode.
Definition: TargetOpcodes.h:36
LLVM_ABI bool isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Return true if this function can prove that V does not have undef bits and is never poison.
LLVM_ABI std::optional< FPValueAndVReg > getFConstantVRegValWithLookThrough(Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs=true)
If VReg is defined by a statically evaluable chain of instructions rooted on a G_FCONSTANT returns it...
Definition: Utils.cpp:447
LLVM_ABI bool isConstFalseVal(const TargetLowering &TLI, int64_t Val, bool IsVector, bool IsFP)
Definition: Utils.cpp:1665
LLVM_ABI std::optional< APFloat > isConstantOrConstantSplatVectorFP(MachineInstr &MI, const MachineRegisterInfo &MRI)
Determines if MI defines a float constant integer or a splat vector of float constant integers.
Definition: Utils.cpp:1576
constexpr unsigned BitWidth
Definition: BitmaskEnum.h:223
LLVM_ABI APFloat getAPFloatFromSize(double Val, unsigned Size)
Returns an APFloat from Val converted to the appropriate size.
Definition: Utils.cpp:657
LLVM_ABI bool isBuildVectorConstantSplat(const Register Reg, const MachineRegisterInfo &MRI, int64_t SplatValue, bool AllowUndef)
Return true if the specified register is defined by G_BUILD_VECTOR or G_BUILD_VECTOR_TRUNC where all ...
Definition: Utils.cpp:1401
LLVM_ABI void eraseInstr(MachineInstr &MI, MachineRegisterInfo &MRI, LostDebugLocObserver *LocObserver=nullptr)
Definition: Utils.cpp:1718
DiagnosticSeverity
Defines the different supported severity of a diagnostic.
@ DS_Warning
@ DS_Error
LLVM_ABI Register constrainRegToClass(MachineRegisterInfo &MRI, const TargetInstrInfo &TII, const RegisterBankInfo &RBI, Register Reg, const TargetRegisterClass &RegClass)
Try to constrain Reg to the specified register class.
Definition: Utils.cpp:46
LLVM_ABI int64_t getICmpTrueVal(const TargetLowering &TLI, bool IsVector, bool IsFP)
Returns an integer representing true, as defined by the TargetBooleanContents.
Definition: Utils.cpp:1677
LLVM_ABI bool isKnownNeverNaN(const Value *V, const SimplifyQuery &SQ, unsigned Depth=0)
Return true if the floating-point scalar value is not a NaN or if the floating-point vector value has...
LLVM_ABI std::optional< ValueAndVReg > getIConstantVRegValWithLookThrough(Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs=true)
If VReg is defined by a statically evaluable chain of instructions rooted on a G_CONSTANT returns its...
Definition: Utils.cpp:433
auto find_if(R &&Range, UnaryPredicate P)
Provide wrappers to std::find_if which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1777
LLVM_ABI bool isPreISelGenericFloatingPointOpcode(unsigned Opc)
Returns whether opcode Opc is a pre-isel generic floating-point opcode, having only floating-point op...
Definition: Utils.cpp:1742
bool isKnownNeverSNaN(Register Val, const MachineRegisterInfo &MRI)
Returns true if Val can be assumed to never be a signaling NaN.
Definition: Utils.h:352
LLVM_ABI std::optional< DefinitionAndSourceRegister > getDefSrcRegIgnoringCopies(Register Reg, const MachineRegisterInfo &MRI)
Find the def instruction for Reg, and underlying value Register folding away any copies.
Definition: Utils.cpp:467
bool is_contained(R &&Range, const E &Element)
Returns true if Element is found in Range.
Definition: STLExtras.h:1916
Align commonAlignment(Align A, uint64_t Offset)
Returns the alignment that satisfies both alignments.
Definition: Alignment.h:212
LLVM_ABI void eraseInstrs(ArrayRef< MachineInstr * > DeadInstrs, MachineRegisterInfo &MRI, LostDebugLocObserver *LocObserver=nullptr)
Definition: Utils.cpp:1703
void salvageDebugInfoForDbgValue(const MachineRegisterInfo &MRI, MachineInstr &MI, ArrayRef< MachineOperand * > DbgUsers)
Assuming the instruction MI is going to be deleted, attempt to salvage debug users of MI by writing t...
LLVM_ABI bool isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL, bool OrZero=false, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true, unsigned Depth=0)
Return true if the given value is known to have exactly one bit set when defined.
LLVM_ABI Register getSrcRegIgnoringCopies(Register Reg, const MachineRegisterInfo &MRI)
Find the source register for Reg, folding away any trivial copies.
Definition: Utils.cpp:499
LLVM_ABI LLVM_READNONE LLT getGCDType(LLT OrigTy, LLT TargetTy)
Return a type where the total size is the greatest common divisor of OrigTy and TargetTy.
Definition: Utils.cpp:1277
LLVM_ABI bool isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Returns true if V cannot be poison, but may be undef.
LLVM_READONLY APFloat minimum(const APFloat &A, const APFloat &B)
Implements IEEE 754-2019 minimum semantics.
Definition: APFloat.h:1616
LLVM_ABI std::optional< int64_t > getIConstantSplatSExtVal(const Register Reg, const MachineRegisterInfo &MRI)
Definition: Utils.cpp:1458
LLVM_ABI bool isAssertMI(const MachineInstr &MI)
Returns true if the instruction MI is one of the assert instructions.
Definition: Utils.cpp:2036
LLVM_ABI void extractVectorParts(Register Reg, unsigned NumElts, SmallVectorImpl< Register > &VRegs, MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI)
Version which handles irregular sub-vector splits.
Definition: Utils.cpp:609
LLVM_ABI int getSplatIndex(ArrayRef< int > Mask)
If all non-negative Mask elements are the same value, return that value.
LLVM_ABI bool isTriviallyDead(const MachineInstr &MI, const MachineRegisterInfo &MRI)
Check whether an instruction MI is dead: it only defines dead virtual registers, and doesn't have oth...
Definition: Utils.cpp:222
LLVM_ABI Align inferAlignFromPtrInfo(MachineFunction &MF, const MachinePointerInfo &MPO)
Definition: Utils.cpp:899
LLVM_ABI void reportGISelWarning(MachineFunction &MF, const TargetPassConfig &TPC, MachineOptimizationRemarkEmitter &MORE, MachineOptimizationRemarkMissed &R)
Report an ISel warning as a missed optimization remark to the LLVMContext's diagnostic stream.
Definition: Utils.cpp:253
LLVM_ABI void reportFatalUsageError(Error Err)
Report a fatal error that does not indicate a bug in LLVM.
Definition: Error.cpp:180
#define MORE()
Definition: regcomp.c:246
This struct is a compact representation of a valid (non-zero power of two) alignment.
Definition: Alignment.h:39
Simple struct used to hold a Register value and the instruction which defines it.
Definition: Utils.h:234
unsigned countMaxPopulation() const
Returns the maximum number of bits that could be one.
Definition: KnownBits.h:282
unsigned countMinPopulation() const
Returns the number of bits known to be one.
Definition: KnownBits.h:279
This class contains a discriminated union of information about pointers in memory operands,...
int64_t Offset
Offset - This is an offset from the base Value*.
PointerUnion< const Value *, const PseudoSourceValue * > V
This is the IR pointer value for the access, or it is null if unknown.
Simple struct used to hold a constant integer value and a virtual register.
Definition: Utils.h:193