1//===- CorrelatedValuePropagation.cpp - Propagate CFG-derived info --------===//
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//
9// This file implements the Correlated Value Propagation pass.
10//
11//===----------------------------------------------------------------------===//
12
13#include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h"
14#include "llvm/ADT/DepthFirstIterator.h"
15#include "llvm/ADT/SmallVector.h"
16#include "llvm/ADT/Statistic.h"
17#include "llvm/Analysis/DomTreeUpdater.h"
18#include "llvm/Analysis/GlobalsModRef.h"
19#include "llvm/Analysis/InstructionSimplify.h"
20#include "llvm/Analysis/LazyValueInfo.h"
21#include "llvm/Analysis/ValueTracking.h"
22#include "llvm/IR/Attributes.h"
23#include "llvm/IR/BasicBlock.h"
24#include "llvm/IR/CFG.h"
25#include "llvm/IR/Constant.h"
26#include "llvm/IR/ConstantRange.h"
27#include "llvm/IR/Constants.h"
28#include "llvm/IR/DerivedTypes.h"
29#include "llvm/IR/Function.h"
30#include "llvm/IR/IRBuilder.h"
31#include "llvm/IR/InstrTypes.h"
32#include "llvm/IR/Instruction.h"
33#include "llvm/IR/Instructions.h"
34#include "llvm/IR/IntrinsicInst.h"
35#include "llvm/IR/MDBuilder.h"
36#include "llvm/IR/Operator.h"
37#include "llvm/IR/PassManager.h"
38#include "llvm/IR/PatternMatch.h"
39#include "llvm/IR/ProfDataUtils.h"
40#include "llvm/IR/Type.h"
41#include "llvm/IR/Value.h"
42#include "llvm/Support/Casting.h"
43#include "llvm/Transforms/Utils/Local.h"
44#include <cassert>
45#include <optional>
46#include <utility>
47
48using namespace llvm;
49
50#define DEBUG_TYPE "correlated-value-propagation"
51
52STATISTIC(NumPhis, "Number of phis propagated");
53STATISTIC(NumPhiCommon, "Number of phis deleted via common incoming value");
54STATISTIC(NumSelects, "Number of selects propagated");
55STATISTIC(NumCmps, "Number of comparisons propagated");
56STATISTIC(NumReturns, "Number of return values propagated");
57STATISTIC(NumDeadCases, "Number of switch cases removed");
58STATISTIC(NumSDivSRemsNarrowed,
59 "Number of sdivs/srems whose width was decreased");
60STATISTIC(NumSDivs, "Number of sdiv converted to udiv");
61STATISTIC(NumUDivURemsNarrowed,
62 "Number of udivs/urems whose width was decreased");
63STATISTIC(NumAShrsConverted, "Number of ashr converted to lshr");
64STATISTIC(NumAShrsRemoved, "Number of ashr removed");
65STATISTIC(NumSRems, "Number of srem converted to urem");
66STATISTIC(NumSExt, "Number of sext converted to zext");
67STATISTIC(NumSIToFP, "Number of sitofp converted to uitofp");
68STATISTIC(NumSICmps, "Number of signed icmp preds simplified to unsigned");
69STATISTIC(NumAnd, "Number of ands removed");
70STATISTIC(NumNW, "Number of no-wrap deductions");
71STATISTIC(NumNSW, "Number of no-signed-wrap deductions");
72STATISTIC(NumNUW, "Number of no-unsigned-wrap deductions");
73STATISTIC(NumAddNW, "Number of no-wrap deductions for add");
74STATISTIC(NumAddNSW, "Number of no-signed-wrap deductions for add");
75STATISTIC(NumAddNUW, "Number of no-unsigned-wrap deductions for add");
76STATISTIC(NumSubNW, "Number of no-wrap deductions for sub");
77STATISTIC(NumSubNSW, "Number of no-signed-wrap deductions for sub");
78STATISTIC(NumSubNUW, "Number of no-unsigned-wrap deductions for sub");
79STATISTIC(NumMulNW, "Number of no-wrap deductions for mul");
80STATISTIC(NumMulNSW, "Number of no-signed-wrap deductions for mul");
81STATISTIC(NumMulNUW, "Number of no-unsigned-wrap deductions for mul");
82STATISTIC(NumShlNW, "Number of no-wrap deductions for shl");
83STATISTIC(NumShlNSW, "Number of no-signed-wrap deductions for shl");
84STATISTIC(NumShlNUW, "Number of no-unsigned-wrap deductions for shl");
85STATISTIC(NumAbs, "Number of llvm.abs intrinsics removed");
86STATISTIC(NumOverflows, "Number of overflow checks removed");
87STATISTIC(NumSaturating,
88 "Number of saturating arithmetics converted to normal arithmetics");
89STATISTIC(NumNonNull, "Number of function pointer arguments marked non-null");
90STATISTIC(NumCmpIntr, "Number of llvm.[us]cmp intrinsics removed");
91STATISTIC(NumMinMax, "Number of llvm.[us]{min,max} intrinsics removed");
92STATISTIC(NumSMinMax,
93 "Number of llvm.s{min,max} intrinsics simplified to unsigned");
94STATISTIC(NumUDivURemsNarrowedExpanded,
95 "Number of bound udiv's/urem's expanded");
96STATISTIC(NumNNeg, "Number of zext/uitofp non-negative deductions");
97
98static Constant *getConstantAt(Value *V, Instruction *At, LazyValueInfo *LVI) {
99 if (Constant *C = LVI->getConstant(V, CxtI: At))
100 return C;
101
102 // TODO: The following really should be sunk inside LVI's core algorithm, or
103 // at least the outer shims around such.
104 auto *C = dyn_cast<CmpInst>(Val: V);
105 if (!C)
106 return nullptr;
107
108 Value *Op0 = C->getOperand(i_nocapture: 0);
109 Constant *Op1 = dyn_cast<Constant>(Val: C->getOperand(i_nocapture: 1));
110 if (!Op1)
111 return nullptr;
112
113 return LVI->getPredicateAt(Pred: C->getPredicate(), V: Op0, C: Op1, CxtI: At,
114 /*UseBlockValue=*/false);
115}
116
117static bool processSelect(SelectInst *S, LazyValueInfo *LVI) {
118 if (S->getType()->isVectorTy() || isa<Constant>(Val: S->getCondition()))
119 return false;
120
121 bool Changed = false;
122 for (Use &U : make_early_inc_range(Range: S->uses())) {
123 auto *I = cast<Instruction>(Val: U.getUser());
124 Constant *C;
125 if (auto *PN = dyn_cast<PHINode>(Val: I))
126 C = LVI->getConstantOnEdge(V: S->getCondition(), FromBB: PN->getIncomingBlock(U),
127 ToBB: I->getParent(), CxtI: I);
128 else
129 C = getConstantAt(V: S->getCondition(), At: I, LVI);
130
131 auto *CI = dyn_cast_or_null<ConstantInt>(Val: C);
132 if (!CI)
133 continue;
134
135 U.set(CI->isOne() ? S->getTrueValue() : S->getFalseValue());
136 Changed = true;
137 ++NumSelects;
138 }
139
140 if (Changed && S->use_empty())
141 S->eraseFromParent();
142
143 return Changed;
144}
145
146/// Try to simplify a phi with constant incoming values that match the edge
147/// values of a non-constant value on all other edges:
148/// bb0:
149/// %isnull = icmp eq i8* %x, null
150/// br i1 %isnull, label %bb2, label %bb1
151/// bb1:
152/// br label %bb2
153/// bb2:
154/// %r = phi i8* [ %x, %bb1 ], [ null, %bb0 ]
155/// -->
156/// %r = %x
157static bool simplifyCommonValuePhi(PHINode *P, LazyValueInfo *LVI,
158 DominatorTree *DT) {
159 // Collect incoming constants and initialize possible common value.
160 SmallVector<std::pair<Constant *, unsigned>, 4> IncomingConstants;
161 Value *CommonValue = nullptr;
162 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
163 Value *Incoming = P->getIncomingValue(i);
164 if (auto *IncomingConstant = dyn_cast<Constant>(Val: Incoming)) {
165 IncomingConstants.push_back(Elt: std::make_pair(x&: IncomingConstant, y&: i));
166 } else if (!CommonValue) {
167 // The potential common value is initialized to the first non-constant.
168 CommonValue = Incoming;
169 } else if (Incoming != CommonValue) {
170 // There can be only one non-constant common value.
171 return false;
172 }
173 }
174
175 if (!CommonValue || IncomingConstants.empty())
176 return false;
177
178 // The common value must be valid in all incoming blocks.
179 BasicBlock *ToBB = P->getParent();
180 if (auto *CommonInst = dyn_cast<Instruction>(Val: CommonValue))
181 if (!DT->dominates(Def: CommonInst, BB: ToBB))
182 return false;
183
184 // We have a phi with exactly 1 variable incoming value and 1 or more constant
185 // incoming values. See if all constant incoming values can be mapped back to
186 // the same incoming variable value.
187 for (auto &IncomingConstant : IncomingConstants) {
188 Constant *C = IncomingConstant.first;
189 BasicBlock *IncomingBB = P->getIncomingBlock(i: IncomingConstant.second);
190 if (C != LVI->getConstantOnEdge(V: CommonValue, FromBB: IncomingBB, ToBB, CxtI: P))
191 return false;
192 }
193
194 // LVI only guarantees that the value matches a certain constant if the value
195 // is not poison. Make sure we don't replace a well-defined value with poison.
196 // This is usually satisfied due to a prior branch on the value.
197 if (!isGuaranteedNotToBePoison(V: CommonValue, AC: nullptr, CtxI: P, DT))
198 return false;
199
200 // All constant incoming values map to the same variable along the incoming
201 // edges of the phi. The phi is unnecessary.
202 P->replaceAllUsesWith(V: CommonValue);
203 P->eraseFromParent();
204 ++NumPhiCommon;
205 return true;
206}
207
208static Value *getValueOnEdge(LazyValueInfo *LVI, Value *Incoming,
209 BasicBlock *From, BasicBlock *To,
210 Instruction *CxtI) {
211 if (Constant *C = LVI->getConstantOnEdge(V: Incoming, FromBB: From, ToBB: To, CxtI))
212 return C;
213
214 // Look if the incoming value is a select with a scalar condition for which
215 // LVI can tells us the value. In that case replace the incoming value with
216 // the appropriate value of the select. This often allows us to remove the
217 // select later.
218 auto *SI = dyn_cast<SelectInst>(Val: Incoming);
219 if (!SI)
220 return nullptr;
221
222 // Once LVI learns to handle vector types, we could also add support
223 // for vector type constants that are not all zeroes or all ones.
224 Value *Condition = SI->getCondition();
225 if (!Condition->getType()->isVectorTy()) {
226 if (Constant *C = LVI->getConstantOnEdge(V: Condition, FromBB: From, ToBB: To, CxtI)) {
227 if (C->isOneValue())
228 return SI->getTrueValue();
229 if (C->isNullValue())
230 return SI->getFalseValue();
231 }
232 }
233
234 // Look if the select has a constant but LVI tells us that the incoming
235 // value can never be that constant. In that case replace the incoming
236 // value with the other value of the select. This often allows us to
237 // remove the select later.
238
239 // The "false" case
240 if (auto *C = dyn_cast<Constant>(Val: SI->getFalseValue()))
241 if (auto *Res = dyn_cast_or_null<ConstantInt>(
242 Val: LVI->getPredicateOnEdge(Pred: ICmpInst::ICMP_EQ, V: SI, C, FromBB: From, ToBB: To, CxtI));
243 Res && Res->isZero())
244 return SI->getTrueValue();
245
246 // The "true" case,
247 // similar to the select "false" case, but try the select "true" value
248 if (auto *C = dyn_cast<Constant>(Val: SI->getTrueValue()))
249 if (auto *Res = dyn_cast_or_null<ConstantInt>(
250 Val: LVI->getPredicateOnEdge(Pred: ICmpInst::ICMP_EQ, V: SI, C, FromBB: From, ToBB: To, CxtI));
251 Res && Res->isZero())
252 return SI->getFalseValue();
253
254 return nullptr;
255}
256
257static bool processPHI(PHINode *P, LazyValueInfo *LVI, DominatorTree *DT,
258 const SimplifyQuery &SQ) {
259 bool Changed = false;
260
261 BasicBlock *BB = P->getParent();
262 for (unsigned i = 0, e = P->getNumIncomingValues(); i < e; ++i) {
263 Value *Incoming = P->getIncomingValue(i);
264 if (isa<Constant>(Val: Incoming)) continue;
265
266 Value *V = getValueOnEdge(LVI, Incoming, From: P->getIncomingBlock(i), To: BB, CxtI: P);
267 if (V) {
268 P->setIncomingValue(i, V);
269 Changed = true;
270 }
271 }
272
273 if (Value *V = simplifyInstruction(I: P, Q: SQ)) {
274 P->replaceAllUsesWith(V);
275 P->eraseFromParent();
276 Changed = true;
277 }
278
279 if (!Changed)
280 Changed = simplifyCommonValuePhi(P, LVI, DT);
281
282 if (Changed)
283 ++NumPhis;
284
285 return Changed;
286}
287
288static bool processICmp(ICmpInst *Cmp, LazyValueInfo *LVI) {
289 // Only for signed relational comparisons of integers.
290 if (!Cmp->getOperand(i_nocapture: 0)->getType()->isIntOrIntVectorTy())
291 return false;
292
293 if (!Cmp->isSigned() && (!Cmp->isUnsigned() || Cmp->hasSameSign()))
294 return false;
295
296 bool Changed = false;
297
298 ConstantRange CR1 = LVI->getConstantRangeAtUse(U: Cmp->getOperandUse(i: 0),
299 /*UndefAllowed=*/false),
300 CR2 = LVI->getConstantRangeAtUse(U: Cmp->getOperandUse(i: 1),
301 /*UndefAllowed=*/false);
302
303 if (Cmp->isSigned()) {
304 ICmpInst::Predicate UnsignedPred =
305 ConstantRange::getEquivalentPredWithFlippedSignedness(
306 Pred: Cmp->getPredicate(), CR1, CR2);
307
308 if (UnsignedPred == ICmpInst::Predicate::BAD_ICMP_PREDICATE)
309 return false;
310
311 ++NumSICmps;
312 Cmp->setPredicate(UnsignedPred);
313 Changed = true;
314 }
315
316 if (ConstantRange::areInsensitiveToSignednessOfICmpPredicate(CR1, CR2)) {
317 Cmp->setSameSign();
318 Changed = true;
319 }
320
321 return Changed;
322}
323
324/// See if LazyValueInfo's ability to exploit edge conditions or range
325/// information is sufficient to prove this comparison. Even for local
326/// conditions, this can sometimes prove conditions instcombine can't by
327/// exploiting range information.
328static bool constantFoldCmp(CmpInst *Cmp, LazyValueInfo *LVI) {
329 Value *Op0 = Cmp->getOperand(i_nocapture: 0);
330 Value *Op1 = Cmp->getOperand(i_nocapture: 1);
331 Constant *Res = LVI->getPredicateAt(Pred: Cmp->getPredicate(), LHS: Op0, RHS: Op1, CxtI: Cmp,
332 /*UseBlockValue=*/true);
333 if (!Res)
334 return false;
335
336 bool Changed = Cmp->replaceUsesWithIf(
337 New: Res, ShouldReplace: [](Use &U) { return !isa<AssumeInst>(Val: U.getUser()); });
338 if (Cmp->use_empty()) {
339 Cmp->eraseFromParent();
340 Changed = true;
341 }
342
343 if (Changed)
344 ++NumCmps;
345
346 return Changed;
347}
348
349static bool processCmp(CmpInst *Cmp, LazyValueInfo *LVI) {
350 if (constantFoldCmp(Cmp, LVI))
351 return true;
352
353 if (auto *ICmp = dyn_cast<ICmpInst>(Val: Cmp))
354 if (processICmp(Cmp: ICmp, LVI))
355 return true;
356
357 return false;
358}
359
360/// Simplify a switch instruction by removing cases which can never fire. If the
361/// uselessness of a case could be determined locally then constant propagation
362/// would already have figured it out. Instead, walk the predecessors and
363/// statically evaluate cases based on information available on that edge. Cases
364/// that cannot fire no matter what the incoming edge can safely be removed. If
365/// a case fires on every incoming edge then the entire switch can be removed
366/// and replaced with a branch to the case destination.
367static bool processSwitch(SwitchInst *I, LazyValueInfo *LVI,
368 DominatorTree *DT) {
369 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
370 Value *Cond = I->getCondition();
371 BasicBlock *BB = I->getParent();
372
373 // Analyse each switch case in turn.
374 bool Changed = false;
375 DenseMap<BasicBlock*, int> SuccessorsCount;
376 for (auto *Succ : successors(BB))
377 SuccessorsCount[Succ]++;
378
379 { // Scope for SwitchInstProfUpdateWrapper. It must not live during
380 // ConstantFoldTerminator() as the underlying SwitchInst can be changed.
381 SwitchInstProfUpdateWrapper SI(*I);
382 ConstantRange CR =
383 LVI->getConstantRangeAtUse(U: I->getOperandUse(i: 0), /*UndefAllowed=*/false);
384 unsigned ReachableCaseCount = 0;
385
386 for (auto CI = SI->case_begin(), CE = SI->case_end(); CI != CE;) {
387 ConstantInt *Case = CI->getCaseValue();
388 std::optional<bool> Predicate = std::nullopt;
389 if (!CR.contains(Val: Case->getValue()))
390 Predicate = false;
391 else if (CR.isSingleElement() &&
392 *CR.getSingleElement() == Case->getValue())
393 Predicate = true;
394 if (!Predicate) {
395 // Handle missing cases, e.g., the range has a hole.
396 auto *Res = dyn_cast_or_null<ConstantInt>(
397 Val: LVI->getPredicateAt(Pred: CmpInst::ICMP_EQ, V: Cond, C: Case, CxtI: I,
398 /* UseBlockValue=*/true));
399 if (Res && Res->isZero())
400 Predicate = false;
401 else if (Res && Res->isOne())
402 Predicate = true;
403 }
404
405 if (Predicate && !*Predicate) {
406 // This case never fires - remove it.
407 BasicBlock *Succ = CI->getCaseSuccessor();
408 Succ->removePredecessor(Pred: BB);
409 CI = SI.removeCase(I: CI);
410 CE = SI->case_end();
411
412 // The condition can be modified by removePredecessor's PHI simplification
413 // logic.
414 Cond = SI->getCondition();
415
416 ++NumDeadCases;
417 Changed = true;
418 if (--SuccessorsCount[Succ] == 0)
419 DTU.applyUpdatesPermissive(Updates: {{DominatorTree::Delete, BB, Succ}});
420 continue;
421 }
422 if (Predicate && *Predicate) {
423 // This case always fires. Arrange for the switch to be turned into an
424 // unconditional branch by replacing the switch condition with the case
425 // value.
426 SI->setCondition(Case);
427 NumDeadCases += SI->getNumCases();
428 Changed = true;
429 break;
430 }
431
432 // Increment the case iterator since we didn't delete it.
433 ++CI;
434 ++ReachableCaseCount;
435 }
436
437 // The default dest is unreachable if all cases are covered.
438 if (!SI->defaultDestUnreachable() &&
439 !CR.isSizeLargerThan(MaxSize: ReachableCaseCount)) {
440 BasicBlock *DefaultDest = SI->getDefaultDest();
441 BasicBlock *NewUnreachableBB =
442 BasicBlock::Create(Context&: BB->getContext(), Name: "default.unreachable",
443 Parent: BB->getParent(), InsertBefore: DefaultDest);
444 auto *UI = new UnreachableInst(BB->getContext(), NewUnreachableBB);
445 UI->setDebugLoc(DebugLoc::getTemporary());
446
447 DefaultDest->removePredecessor(Pred: BB);
448 SI->setDefaultDest(NewUnreachableBB);
449
450 if (SuccessorsCount[DefaultDest] == 1)
451 DTU.applyUpdates(Updates: {{DominatorTree::Delete, BB, DefaultDest}});
452 DTU.applyUpdates(Updates: {{DominatorTree::Insert, BB, NewUnreachableBB}});
453
454 ++NumDeadCases;
455 Changed = true;
456 }
457 }
458
459 if (Changed)
460 // If the switch has been simplified to the point where it can be replaced
461 // by a branch then do so now.
462 ConstantFoldTerminator(BB, /*DeleteDeadConditions = */ false,
463 /*TLI = */ nullptr, DTU: &DTU);
464 return Changed;
465}
466
467// See if we can prove that the given binary op intrinsic will not overflow.
468static bool willNotOverflow(BinaryOpIntrinsic *BO, LazyValueInfo *LVI) {
469 ConstantRange LRange =
470 LVI->getConstantRangeAtUse(U: BO->getOperandUse(i: 0), /*UndefAllowed*/ false);
471 ConstantRange RRange =
472 LVI->getConstantRangeAtUse(U: BO->getOperandUse(i: 1), /*UndefAllowed*/ false);
473 ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
474 BinOp: BO->getBinaryOp(), Other: RRange, NoWrapKind: BO->getNoWrapKind());
475 return NWRegion.contains(CR: LRange);
476}
477
478static void setDeducedOverflowingFlags(Value *V, Instruction::BinaryOps Opcode,
479 bool NewNSW, bool NewNUW) {
480 Statistic *OpcNW, *OpcNSW, *OpcNUW;
481 switch (Opcode) {
482 case Instruction::Add:
483 OpcNW = &NumAddNW;
484 OpcNSW = &NumAddNSW;
485 OpcNUW = &NumAddNUW;
486 break;
487 case Instruction::Sub:
488 OpcNW = &NumSubNW;
489 OpcNSW = &NumSubNSW;
490 OpcNUW = &NumSubNUW;
491 break;
492 case Instruction::Mul:
493 OpcNW = &NumMulNW;
494 OpcNSW = &NumMulNSW;
495 OpcNUW = &NumMulNUW;
496 break;
497 case Instruction::Shl:
498 OpcNW = &NumShlNW;
499 OpcNSW = &NumShlNSW;
500 OpcNUW = &NumShlNUW;
501 break;
502 default:
503 llvm_unreachable("Will not be called with other binops");
504 }
505
506 auto *Inst = dyn_cast<Instruction>(Val: V);
507 if (NewNSW) {
508 ++NumNW;
509 ++*OpcNW;
510 ++NumNSW;
511 ++*OpcNSW;
512 if (Inst)
513 Inst->setHasNoSignedWrap();
514 }
515 if (NewNUW) {
516 ++NumNW;
517 ++*OpcNW;
518 ++NumNUW;
519 ++*OpcNUW;
520 if (Inst)
521 Inst->setHasNoUnsignedWrap();
522 }
523}
524
525static bool processBinOp(BinaryOperator *BinOp, LazyValueInfo *LVI);
526
527// See if @llvm.abs argument is alays positive/negative, and simplify.
528// Notably, INT_MIN can belong to either range, regardless of the NSW,
529// because it is negation-invariant.
530static bool processAbsIntrinsic(IntrinsicInst *II, LazyValueInfo *LVI) {
531 Value *X = II->getArgOperand(i: 0);
532 bool IsIntMinPoison = cast<ConstantInt>(Val: II->getArgOperand(i: 1))->isOne();
533 APInt IntMin = APInt::getSignedMinValue(numBits: X->getType()->getScalarSizeInBits());
534 ConstantRange Range = LVI->getConstantRangeAtUse(
535 U: II->getOperandUse(i: 0), /*UndefAllowed*/ IsIntMinPoison);
536
537 // Is X in [0, IntMin]? NOTE: INT_MIN is fine!
538 if (Range.icmp(Pred: CmpInst::ICMP_ULE, Other: IntMin)) {
539 ++NumAbs;
540 II->replaceAllUsesWith(V: X);
541 II->eraseFromParent();
542 return true;
543 }
544
545 // Is X in [IntMin, 0]? NOTE: INT_MIN is fine!
546 if (Range.getSignedMax().isNonPositive()) {
547 IRBuilder<> B(II);
548 Value *NegX = B.CreateNeg(V: X, Name: II->getName(),
549 /*HasNSW=*/IsIntMinPoison);
550 ++NumAbs;
551 II->replaceAllUsesWith(V: NegX);
552 II->eraseFromParent();
553
554 // See if we can infer some no-wrap flags.
555 if (auto *BO = dyn_cast<BinaryOperator>(Val: NegX))
556 processBinOp(BinOp: BO, LVI);
557
558 return true;
559 }
560
561 // Argument's range crosses zero.
562 // Can we at least tell that the argument is never INT_MIN?
563 if (!IsIntMinPoison && !Range.contains(Val: IntMin)) {
564 ++NumNSW;
565 ++NumSubNSW;
566 II->setArgOperand(i: 1, v: ConstantInt::getTrue(Context&: II->getContext()));
567 return true;
568 }
569 return false;
570}
571
572static bool processCmpIntrinsic(CmpIntrinsic *CI, LazyValueInfo *LVI) {
573 ConstantRange LHS_CR =
574 LVI->getConstantRangeAtUse(U: CI->getOperandUse(i: 0), /*UndefAllowed*/ false);
575 ConstantRange RHS_CR =
576 LVI->getConstantRangeAtUse(U: CI->getOperandUse(i: 1), /*UndefAllowed*/ false);
577
578 if (LHS_CR.icmp(Pred: CI->getGTPredicate(), Other: RHS_CR)) {
579 ++NumCmpIntr;
580 CI->replaceAllUsesWith(V: ConstantInt::get(Ty: CI->getType(), V: 1));
581 CI->eraseFromParent();
582 return true;
583 }
584 if (LHS_CR.icmp(Pred: CI->getLTPredicate(), Other: RHS_CR)) {
585 ++NumCmpIntr;
586 CI->replaceAllUsesWith(V: ConstantInt::getSigned(Ty: CI->getType(), V: -1));
587 CI->eraseFromParent();
588 return true;
589 }
590 if (LHS_CR.icmp(Pred: ICmpInst::ICMP_EQ, Other: RHS_CR)) {
591 ++NumCmpIntr;
592 CI->replaceAllUsesWith(V: ConstantInt::get(Ty: CI->getType(), V: 0));
593 CI->eraseFromParent();
594 return true;
595 }
596
597 return false;
598}
599
600// See if this min/max intrinsic always picks it's one specific operand.
601// If not, check whether we can canonicalize signed minmax into unsigned version
602static bool processMinMaxIntrinsic(MinMaxIntrinsic *MM, LazyValueInfo *LVI) {
603 CmpInst::Predicate Pred = CmpInst::getNonStrictPredicate(pred: MM->getPredicate());
604 ConstantRange LHS_CR = LVI->getConstantRangeAtUse(U: MM->getOperandUse(i: 0),
605 /*UndefAllowed*/ false);
606 ConstantRange RHS_CR = LVI->getConstantRangeAtUse(U: MM->getOperandUse(i: 1),
607 /*UndefAllowed*/ false);
608 if (LHS_CR.icmp(Pred, Other: RHS_CR)) {
609 ++NumMinMax;
610 MM->replaceAllUsesWith(V: MM->getLHS());
611 MM->eraseFromParent();
612 return true;
613 }
614 if (RHS_CR.icmp(Pred, Other: LHS_CR)) {
615 ++NumMinMax;
616 MM->replaceAllUsesWith(V: MM->getRHS());
617 MM->eraseFromParent();
618 return true;
619 }
620
621 if (MM->isSigned() &&
622 ConstantRange::areInsensitiveToSignednessOfICmpPredicate(CR1: LHS_CR,
623 CR2: RHS_CR)) {
624 ++NumSMinMax;
625 IRBuilder<> B(MM);
626 MM->replaceAllUsesWith(V: B.CreateBinaryIntrinsic(
627 ID: MM->getIntrinsicID() == Intrinsic::smin ? Intrinsic::umin
628 : Intrinsic::umax,
629 LHS: MM->getLHS(), RHS: MM->getRHS()));
630 MM->eraseFromParent();
631 return true;
632 }
633
634 return false;
635}
636
637// Rewrite this with.overflow intrinsic as non-overflowing.
638static bool processOverflowIntrinsic(WithOverflowInst *WO, LazyValueInfo *LVI) {
639 IRBuilder<> B(WO);
640 Instruction::BinaryOps Opcode = WO->getBinaryOp();
641 bool NSW = WO->isSigned();
642 bool NUW = !WO->isSigned();
643
644 Value *NewOp =
645 B.CreateBinOp(Opc: Opcode, LHS: WO->getLHS(), RHS: WO->getRHS(), Name: WO->getName());
646 setDeducedOverflowingFlags(V: NewOp, Opcode, NewNSW: NSW, NewNUW: NUW);
647
648 StructType *ST = cast<StructType>(Val: WO->getType());
649 Constant *Struct = ConstantStruct::get(T: ST,
650 V: { PoisonValue::get(T: ST->getElementType(N: 0)),
651 ConstantInt::getFalse(Ty: ST->getElementType(N: 1)) });
652 Value *NewI = B.CreateInsertValue(Agg: Struct, Val: NewOp, Idxs: 0);
653 WO->replaceAllUsesWith(V: NewI);
654 WO->eraseFromParent();
655 ++NumOverflows;
656
657 // See if we can infer the other no-wrap too.
658 if (auto *BO = dyn_cast<BinaryOperator>(Val: NewOp))
659 processBinOp(BinOp: BO, LVI);
660
661 return true;
662}
663
664static bool processSaturatingInst(SaturatingInst *SI, LazyValueInfo *LVI) {
665 Instruction::BinaryOps Opcode = SI->getBinaryOp();
666 bool NSW = SI->isSigned();
667 bool NUW = !SI->isSigned();
668 BinaryOperator *BinOp = BinaryOperator::Create(
669 Op: Opcode, S1: SI->getLHS(), S2: SI->getRHS(), Name: SI->getName(), InsertBefore: SI->getIterator());
670 BinOp->setDebugLoc(SI->getDebugLoc());
671 setDeducedOverflowingFlags(V: BinOp, Opcode, NewNSW: NSW, NewNUW: NUW);
672
673 SI->replaceAllUsesWith(V: BinOp);
674 SI->eraseFromParent();
675 ++NumSaturating;
676
677 // See if we can infer the other no-wrap too.
678 processBinOp(BinOp, LVI);
679
680 return true;
681}
682
683/// Infer nonnull attributes for the arguments at the specified callsite.
684static bool processCallSite(CallBase &CB, LazyValueInfo *LVI) {
685
686 if (CB.getIntrinsicID() == Intrinsic::abs) {
687 return processAbsIntrinsic(II: &cast<IntrinsicInst>(Val&: CB), LVI);
688 }
689
690 if (auto *CI = dyn_cast<CmpIntrinsic>(Val: &CB)) {
691 return processCmpIntrinsic(CI, LVI);
692 }
693
694 if (auto *MM = dyn_cast<MinMaxIntrinsic>(Val: &CB)) {
695 return processMinMaxIntrinsic(MM, LVI);
696 }
697
698 if (auto *WO = dyn_cast<WithOverflowInst>(Val: &CB)) {
699 if (willNotOverflow(BO: WO, LVI))
700 return processOverflowIntrinsic(WO, LVI);
701 }
702
703 if (auto *SI = dyn_cast<SaturatingInst>(Val: &CB)) {
704 if (willNotOverflow(BO: SI, LVI))
705 return processSaturatingInst(SI, LVI);
706 }
707
708 bool Changed = false;
709
710 // Deopt bundle operands are intended to capture state with minimal
711 // perturbance of the code otherwise. If we can find a constant value for
712 // any such operand and remove a use of the original value, that's
713 // desireable since it may allow further optimization of that value (e.g. via
714 // single use rules in instcombine). Since deopt uses tend to,
715 // idiomatically, appear along rare conditional paths, it's reasonable likely
716 // we may have a conditional fact with which LVI can fold.
717 if (auto DeoptBundle = CB.getOperandBundle(ID: LLVMContext::OB_deopt)) {
718 for (const Use &ConstU : DeoptBundle->Inputs) {
719 Use &U = const_cast<Use&>(ConstU);
720 Value *V = U.get();
721 if (V->getType()->isVectorTy()) continue;
722 if (isa<Constant>(Val: V)) continue;
723
724 Constant *C = LVI->getConstant(V, CxtI: &CB);
725 if (!C) continue;
726 U.set(C);
727 Changed = true;
728 }
729 }
730
731 SmallVector<unsigned, 4> ArgNos;
732 unsigned ArgNo = 0;
733
734 for (Value *V : CB.args()) {
735 PointerType *Type = dyn_cast<PointerType>(Val: V->getType());
736 // Try to mark pointer typed parameters as non-null. We skip the
737 // relatively expensive analysis for constants which are obviously either
738 // null or non-null to start with.
739 if (Type && !CB.paramHasAttr(ArgNo, Kind: Attribute::NonNull) &&
740 !isa<Constant>(Val: V))
741 if (auto *Res = dyn_cast_or_null<ConstantInt>(Val: LVI->getPredicateAt(
742 Pred: ICmpInst::ICMP_EQ, V, C: ConstantPointerNull::get(T: Type), CxtI: &CB,
743 /*UseBlockValue=*/false));
744 Res && Res->isZero())
745 ArgNos.push_back(Elt: ArgNo);
746 ArgNo++;
747 }
748
749 assert(ArgNo == CB.arg_size() && "Call arguments not processed correctly.");
750
751 if (ArgNos.empty())
752 return Changed;
753
754 NumNonNull += ArgNos.size();
755 AttributeList AS = CB.getAttributes();
756 LLVMContext &Ctx = CB.getContext();
757 AS = AS.addParamAttribute(C&: Ctx, ArgNos,
758 A: Attribute::get(Context&: Ctx, Kind: Attribute::NonNull));
759 CB.setAttributes(AS);
760
761 return true;
762}
763
764enum class Domain { NonNegative, NonPositive, Unknown };
765
766static Domain getDomain(const ConstantRange &CR) {
767 if (CR.isAllNonNegative())
768 return Domain::NonNegative;
769 if (CR.icmp(Pred: ICmpInst::ICMP_SLE, Other: APInt::getZero(numBits: CR.getBitWidth())))
770 return Domain::NonPositive;
771 return Domain::Unknown;
772}
773
774/// Try to shrink a sdiv/srem's width down to the smallest power of two that's
775/// sufficient to contain its operands.
776static bool narrowSDivOrSRem(BinaryOperator *Instr, const ConstantRange &LCR,
777 const ConstantRange &RCR) {
778 assert(Instr->getOpcode() == Instruction::SDiv ||
779 Instr->getOpcode() == Instruction::SRem);
780
781 // Find the smallest power of two bitwidth that's sufficient to hold Instr's
782 // operands.
783 unsigned OrigWidth = Instr->getType()->getScalarSizeInBits();
784
785 // What is the smallest bit width that can accommodate the entire value ranges
786 // of both of the operands?
787 unsigned MinSignedBits =
788 std::max(a: LCR.getMinSignedBits(), b: RCR.getMinSignedBits());
789
790 // sdiv/srem is UB if divisor is -1 and divident is INT_MIN, so unless we can
791 // prove that such a combination is impossible, we need to bump the bitwidth.
792 if (RCR.contains(Val: APInt::getAllOnes(numBits: OrigWidth)) &&
793 LCR.contains(Val: APInt::getSignedMinValue(numBits: MinSignedBits).sext(width: OrigWidth)))
794 ++MinSignedBits;
795
796 // Don't shrink below 8 bits wide.
797 unsigned NewWidth = std::max<unsigned>(a: PowerOf2Ceil(A: MinSignedBits), b: 8);
798
799 // NewWidth might be greater than OrigWidth if OrigWidth is not a power of
800 // two.
801 if (NewWidth >= OrigWidth)
802 return false;
803
804 ++NumSDivSRemsNarrowed;
805 IRBuilder<> B{Instr};
806 auto *TruncTy = Instr->getType()->getWithNewBitWidth(NewBitWidth: NewWidth);
807 auto *LHS = B.CreateTruncOrBitCast(V: Instr->getOperand(i_nocapture: 0), DestTy: TruncTy,
808 Name: Instr->getName() + ".lhs.trunc");
809 auto *RHS = B.CreateTruncOrBitCast(V: Instr->getOperand(i_nocapture: 1), DestTy: TruncTy,
810 Name: Instr->getName() + ".rhs.trunc");
811 auto *BO = B.CreateBinOp(Opc: Instr->getOpcode(), LHS, RHS, Name: Instr->getName());
812 auto *Sext = B.CreateSExt(V: BO, DestTy: Instr->getType(), Name: Instr->getName() + ".sext");
813 if (auto *BinOp = dyn_cast<BinaryOperator>(Val: BO))
814 if (BinOp->getOpcode() == Instruction::SDiv)
815 BinOp->setIsExact(Instr->isExact());
816
817 Instr->replaceAllUsesWith(V: Sext);
818 Instr->eraseFromParent();
819 return true;
820}
821
822static bool expandUDivOrURem(BinaryOperator *Instr, const ConstantRange &XCR,
823 const ConstantRange &YCR) {
824 Type *Ty = Instr->getType();
825 assert(Instr->getOpcode() == Instruction::UDiv ||
826 Instr->getOpcode() == Instruction::URem);
827 bool IsRem = Instr->getOpcode() == Instruction::URem;
828
829 Value *X = Instr->getOperand(i_nocapture: 0);
830 Value *Y = Instr->getOperand(i_nocapture: 1);
831
832 // X u/ Y -> 0 iff X u< Y
833 // X u% Y -> X iff X u< Y
834 if (XCR.icmp(Pred: ICmpInst::ICMP_ULT, Other: YCR)) {
835 Instr->replaceAllUsesWith(V: IsRem ? X : Constant::getNullValue(Ty));
836 Instr->eraseFromParent();
837 ++NumUDivURemsNarrowedExpanded;
838 return true;
839 }
840
841 // Given
842 // R = X u% Y
843 // We can represent the modulo operation as a loop/self-recursion:
844 // urem_rec(X, Y):
845 // Z = X - Y
846 // if X u< Y
847 // ret X
848 // else
849 // ret urem_rec(Z, Y)
850 // which isn't better, but if we only need a single iteration
851 // to compute the answer, this becomes quite good:
852 // R = X < Y ? X : X - Y iff X u< 2*Y (w/ unsigned saturation)
853 // Now, we do not care about all full multiples of Y in X, they do not change
854 // the answer, thus we could rewrite the expression as:
855 // X* = X - (Y * |_ X / Y _|)
856 // R = X* % Y
857 // so we don't need the *first* iteration to return, we just need to
858 // know *which* iteration will always return, so we could also rewrite it as:
859 // X* = X - (Y * |_ X / Y _|)
860 // R = X* % Y iff X* u< 2*Y (w/ unsigned saturation)
861 // but that does not seem profitable here.
862
863 // Even if we don't know X's range, the divisor may be so large, X can't ever
864 // be 2x larger than that. I.e. if divisor is always negative.
865 if (!XCR.icmp(Pred: ICmpInst::ICMP_ULT, Other: YCR.uadd_sat(Other: YCR)) && !YCR.isAllNegative())
866 return false;
867
868 IRBuilder<> B(Instr);
869 Value *ExpandedOp;
870 if (XCR.icmp(Pred: ICmpInst::ICMP_UGE, Other: YCR)) {
871 // If X is between Y and 2*Y the result is known.
872 if (IsRem)
873 ExpandedOp = B.CreateNUWSub(LHS: X, RHS: Y);
874 else
875 ExpandedOp = ConstantInt::get(Ty: Instr->getType(), V: 1);
876 } else if (IsRem) {
877 // NOTE: this transformation introduces two uses of X,
878 // but it may be undef so we must freeze it first.
879 Value *FrozenX = X;
880 if (!isGuaranteedNotToBeUndef(V: X))
881 FrozenX = B.CreateFreeze(V: X, Name: X->getName() + ".frozen");
882 Value *FrozenY = Y;
883 if (!isGuaranteedNotToBeUndef(V: Y))
884 FrozenY = B.CreateFreeze(V: Y, Name: Y->getName() + ".frozen");
885 auto *AdjX = B.CreateNUWSub(LHS: FrozenX, RHS: FrozenY, Name: Instr->getName() + ".urem");
886 auto *Cmp = B.CreateICmp(P: ICmpInst::ICMP_ULT, LHS: FrozenX, RHS: FrozenY,
887 Name: Instr->getName() + ".cmp");
888 ExpandedOp =
889 B.CreateSelectWithUnknownProfile(C: Cmp, True: FrozenX, False: AdjX, DEBUG_TYPE);
890 } else {
891 auto *Cmp =
892 B.CreateICmp(P: ICmpInst::ICMP_UGE, LHS: X, RHS: Y, Name: Instr->getName() + ".cmp");
893 ExpandedOp = B.CreateZExt(V: Cmp, DestTy: Ty, Name: Instr->getName() + ".udiv");
894 }
895 ExpandedOp->takeName(V: Instr);
896 Instr->replaceAllUsesWith(V: ExpandedOp);
897 Instr->eraseFromParent();
898 ++NumUDivURemsNarrowedExpanded;
899 return true;
900}
901
902/// Try to shrink a udiv/urem's width down to the smallest power of two that's
903/// sufficient to contain its operands.
904static bool narrowUDivOrURem(BinaryOperator *Instr, const ConstantRange &XCR,
905 const ConstantRange &YCR) {
906 assert(Instr->getOpcode() == Instruction::UDiv ||
907 Instr->getOpcode() == Instruction::URem);
908
909 // Find the smallest power of two bitwidth that's sufficient to hold Instr's
910 // operands.
911
912 // What is the smallest bit width that can accommodate the entire value ranges
913 // of both of the operands?
914 unsigned MaxActiveBits = std::max(a: XCR.getActiveBits(), b: YCR.getActiveBits());
915 // Don't shrink below 8 bits wide.
916 unsigned NewWidth = std::max<unsigned>(a: PowerOf2Ceil(A: MaxActiveBits), b: 8);
917
918 // NewWidth might be greater than OrigWidth if OrigWidth is not a power of
919 // two.
920 if (NewWidth >= Instr->getType()->getScalarSizeInBits())
921 return false;
922
923 ++NumUDivURemsNarrowed;
924 IRBuilder<> B{Instr};
925 auto *TruncTy = Instr->getType()->getWithNewBitWidth(NewBitWidth: NewWidth);
926 auto *LHS = B.CreateTruncOrBitCast(V: Instr->getOperand(i_nocapture: 0), DestTy: TruncTy,
927 Name: Instr->getName() + ".lhs.trunc");
928 auto *RHS = B.CreateTruncOrBitCast(V: Instr->getOperand(i_nocapture: 1), DestTy: TruncTy,
929 Name: Instr->getName() + ".rhs.trunc");
930 auto *BO = B.CreateBinOp(Opc: Instr->getOpcode(), LHS, RHS, Name: Instr->getName());
931 auto *Zext = B.CreateZExt(V: BO, DestTy: Instr->getType(), Name: Instr->getName() + ".zext");
932 if (auto *BinOp = dyn_cast<BinaryOperator>(Val: BO))
933 if (BinOp->getOpcode() == Instruction::UDiv)
934 BinOp->setIsExact(Instr->isExact());
935
936 Instr->replaceAllUsesWith(V: Zext);
937 Instr->eraseFromParent();
938 return true;
939}
940
941static bool processUDivOrURem(BinaryOperator *Instr, LazyValueInfo *LVI) {
942 assert(Instr->getOpcode() == Instruction::UDiv ||
943 Instr->getOpcode() == Instruction::URem);
944 ConstantRange XCR = LVI->getConstantRangeAtUse(U: Instr->getOperandUse(i: 0),
945 /*UndefAllowed*/ false);
946 // Allow undef for RHS, as we can assume it is division by zero UB.
947 ConstantRange YCR = LVI->getConstantRangeAtUse(U: Instr->getOperandUse(i: 1),
948 /*UndefAllowed*/ true);
949 if (expandUDivOrURem(Instr, XCR, YCR))
950 return true;
951
952 return narrowUDivOrURem(Instr, XCR, YCR);
953}
954
955static bool processSRem(BinaryOperator *SDI, const ConstantRange &LCR,
956 const ConstantRange &RCR, LazyValueInfo *LVI) {
957 assert(SDI->getOpcode() == Instruction::SRem);
958
959 if (LCR.abs().icmp(Pred: CmpInst::ICMP_ULT, Other: RCR.abs())) {
960 SDI->replaceAllUsesWith(V: SDI->getOperand(i_nocapture: 0));
961 SDI->eraseFromParent();
962 return true;
963 }
964
965 struct Operand {
966 Value *V;
967 Domain D;
968 };
969 std::array<Operand, 2> Ops = {._M_elems: {{.V: SDI->getOperand(i_nocapture: 0), .D: getDomain(CR: LCR)},
970 {.V: SDI->getOperand(i_nocapture: 1), .D: getDomain(CR: RCR)}}};
971 if (Ops[0].D == Domain::Unknown || Ops[1].D == Domain::Unknown)
972 return false;
973
974 // We know domains of both of the operands!
975 ++NumSRems;
976
977 // We need operands to be non-negative, so negate each one that isn't.
978 for (Operand &Op : Ops) {
979 if (Op.D == Domain::NonNegative)
980 continue;
981 auto *BO = BinaryOperator::CreateNeg(Op: Op.V, Name: Op.V->getName() + ".nonneg",
982 InsertBefore: SDI->getIterator());
983 BO->setDebugLoc(SDI->getDebugLoc());
984 Op.V = BO;
985 }
986
987 auto *URem = BinaryOperator::CreateURem(V1: Ops[0].V, V2: Ops[1].V, Name: SDI->getName(),
988 InsertBefore: SDI->getIterator());
989 URem->setDebugLoc(SDI->getDebugLoc());
990
991 auto *Res = URem;
992
993 // If the divident was non-positive, we need to negate the result.
994 if (Ops[0].D == Domain::NonPositive) {
995 Res = BinaryOperator::CreateNeg(Op: Res, Name: Res->getName() + ".neg",
996 InsertBefore: SDI->getIterator());
997 Res->setDebugLoc(SDI->getDebugLoc());
998 }
999
1000 SDI->replaceAllUsesWith(V: Res);
1001 SDI->eraseFromParent();
1002
1003 // Try to simplify our new urem.
1004 processUDivOrURem(Instr: URem, LVI);
1005
1006 return true;
1007}
1008
1009/// See if LazyValueInfo's ability to exploit edge conditions or range
1010/// information is sufficient to prove the signs of both operands of this SDiv.
1011/// If this is the case, replace the SDiv with a UDiv. Even for local
1012/// conditions, this can sometimes prove conditions instcombine can't by
1013/// exploiting range information.
1014static bool processSDiv(BinaryOperator *SDI, const ConstantRange &LCR,
1015 const ConstantRange &RCR, LazyValueInfo *LVI) {
1016 assert(SDI->getOpcode() == Instruction::SDiv);
1017
1018 // Check whether the division folds to a constant.
1019 ConstantRange DivCR = LCR.sdiv(Other: RCR);
1020 if (const APInt *Elem = DivCR.getSingleElement()) {
1021 SDI->replaceAllUsesWith(V: ConstantInt::get(Ty: SDI->getType(), V: *Elem));
1022 SDI->eraseFromParent();
1023 return true;
1024 }
1025
1026 struct Operand {
1027 Value *V;
1028 Domain D;
1029 };
1030 std::array<Operand, 2> Ops = {._M_elems: {{.V: SDI->getOperand(i_nocapture: 0), .D: getDomain(CR: LCR)},
1031 {.V: SDI->getOperand(i_nocapture: 1), .D: getDomain(CR: RCR)}}};
1032 if (Ops[0].D == Domain::Unknown || Ops[1].D == Domain::Unknown)
1033 return false;
1034
1035 // We know domains of both of the operands!
1036 ++NumSDivs;
1037
1038 // We need operands to be non-negative, so negate each one that isn't.
1039 for (Operand &Op : Ops) {
1040 if (Op.D == Domain::NonNegative)
1041 continue;
1042 auto *BO = BinaryOperator::CreateNeg(Op: Op.V, Name: Op.V->getName() + ".nonneg",
1043 InsertBefore: SDI->getIterator());
1044 BO->setDebugLoc(SDI->getDebugLoc());
1045 Op.V = BO;
1046 }
1047
1048 auto *UDiv = BinaryOperator::CreateUDiv(V1: Ops[0].V, V2: Ops[1].V, Name: SDI->getName(),
1049 InsertBefore: SDI->getIterator());
1050 UDiv->setDebugLoc(SDI->getDebugLoc());
1051 UDiv->setIsExact(SDI->isExact());
1052
1053 auto *Res = UDiv;
1054
1055 // If the operands had two different domains, we need to negate the result.
1056 if (Ops[0].D != Ops[1].D) {
1057 Res = BinaryOperator::CreateNeg(Op: Res, Name: Res->getName() + ".neg",
1058 InsertBefore: SDI->getIterator());
1059 Res->setDebugLoc(SDI->getDebugLoc());
1060 }
1061
1062 SDI->replaceAllUsesWith(V: Res);
1063 SDI->eraseFromParent();
1064
1065 // Try to simplify our new udiv.
1066 processUDivOrURem(Instr: UDiv, LVI);
1067
1068 return true;
1069}
1070
1071static bool processSDivOrSRem(BinaryOperator *Instr, LazyValueInfo *LVI) {
1072 assert(Instr->getOpcode() == Instruction::SDiv ||
1073 Instr->getOpcode() == Instruction::SRem);
1074 ConstantRange LCR =
1075 LVI->getConstantRangeAtUse(U: Instr->getOperandUse(i: 0), /*AllowUndef*/ UndefAllowed: false);
1076 // Allow undef for RHS, as we can assume it is division by zero UB.
1077 ConstantRange RCR =
1078 LVI->getConstantRangeAtUse(U: Instr->getOperandUse(i: 1), /*AlloweUndef*/ UndefAllowed: true);
1079 if (Instr->getOpcode() == Instruction::SDiv)
1080 if (processSDiv(SDI: Instr, LCR, RCR, LVI))
1081 return true;
1082
1083 if (Instr->getOpcode() == Instruction::SRem) {
1084 if (processSRem(SDI: Instr, LCR, RCR, LVI))
1085 return true;
1086 }
1087
1088 return narrowSDivOrSRem(Instr, LCR, RCR);
1089}
1090
1091static bool processAShr(BinaryOperator *SDI, LazyValueInfo *LVI) {
1092 ConstantRange LRange =
1093 LVI->getConstantRangeAtUse(U: SDI->getOperandUse(i: 0), /*UndefAllowed*/ false);
1094 unsigned OrigWidth = SDI->getType()->getScalarSizeInBits();
1095 ConstantRange NegOneOrZero =
1096 ConstantRange(APInt(OrigWidth, (uint64_t)-1, true), APInt(OrigWidth, 1));
1097 if (NegOneOrZero.contains(CR: LRange)) {
1098 // ashr of -1 or 0 never changes the value, so drop the whole instruction
1099 ++NumAShrsRemoved;
1100 SDI->replaceAllUsesWith(V: SDI->getOperand(i_nocapture: 0));
1101 SDI->eraseFromParent();
1102 return true;
1103 }
1104
1105 if (!LRange.isAllNonNegative())
1106 return false;
1107
1108 ++NumAShrsConverted;
1109 auto *BO = BinaryOperator::CreateLShr(V1: SDI->getOperand(i_nocapture: 0), V2: SDI->getOperand(i_nocapture: 1),
1110 Name: "", InsertBefore: SDI->getIterator());
1111 BO->takeName(V: SDI);
1112 BO->setDebugLoc(SDI->getDebugLoc());
1113 BO->setIsExact(SDI->isExact());
1114 SDI->replaceAllUsesWith(V: BO);
1115 SDI->eraseFromParent();
1116
1117 return true;
1118}
1119
1120static bool processSExt(SExtInst *SDI, LazyValueInfo *LVI) {
1121 const Use &Base = SDI->getOperandUse(i: 0);
1122 if (!LVI->getConstantRangeAtUse(U: Base, /*UndefAllowed*/ false)
1123 .isAllNonNegative())
1124 return false;
1125
1126 ++NumSExt;
1127 auto *ZExt = CastInst::CreateZExtOrBitCast(S: Base, Ty: SDI->getType(), Name: "",
1128 InsertBefore: SDI->getIterator());
1129 ZExt->takeName(V: SDI);
1130 ZExt->setDebugLoc(SDI->getDebugLoc());
1131 ZExt->setNonNeg();
1132 SDI->replaceAllUsesWith(V: ZExt);
1133 SDI->eraseFromParent();
1134
1135 return true;
1136}
1137
1138static bool processPossibleNonNeg(PossiblyNonNegInst *I, LazyValueInfo *LVI) {
1139 if (I->hasNonNeg())
1140 return false;
1141
1142 const Use &Base = I->getOperandUse(i: 0);
1143 if (!LVI->getConstantRangeAtUse(U: Base, /*UndefAllowed*/ false)
1144 .isAllNonNegative())
1145 return false;
1146
1147 ++NumNNeg;
1148 I->setNonNeg();
1149
1150 return true;
1151}
1152
1153static bool processZExt(ZExtInst *ZExt, LazyValueInfo *LVI) {
1154 return processPossibleNonNeg(I: cast<PossiblyNonNegInst>(Val: ZExt), LVI);
1155}
1156
1157static bool processUIToFP(UIToFPInst *UIToFP, LazyValueInfo *LVI) {
1158 return processPossibleNonNeg(I: cast<PossiblyNonNegInst>(Val: UIToFP), LVI);
1159}
1160
1161static bool processSIToFP(SIToFPInst *SIToFP, LazyValueInfo *LVI) {
1162 const Use &Base = SIToFP->getOperandUse(i: 0);
1163 if (!LVI->getConstantRangeAtUse(U: Base, /*UndefAllowed*/ false)
1164 .isAllNonNegative())
1165 return false;
1166
1167 ++NumSIToFP;
1168 auto *UIToFP = CastInst::Create(Instruction::UIToFP, S: Base, Ty: SIToFP->getType(),
1169 Name: "", InsertBefore: SIToFP->getIterator());
1170 UIToFP->takeName(V: SIToFP);
1171 UIToFP->setDebugLoc(SIToFP->getDebugLoc());
1172 UIToFP->setNonNeg();
1173 SIToFP->replaceAllUsesWith(V: UIToFP);
1174 SIToFP->eraseFromParent();
1175
1176 return true;
1177}
1178
1179static bool processBinOp(BinaryOperator *BinOp, LazyValueInfo *LVI) {
1180 using OBO = OverflowingBinaryOperator;
1181
1182 bool NSW = BinOp->hasNoSignedWrap();
1183 bool NUW = BinOp->hasNoUnsignedWrap();
1184 if (NSW && NUW)
1185 return false;
1186
1187 Instruction::BinaryOps Opcode = BinOp->getOpcode();
1188 ConstantRange LRange = LVI->getConstantRangeAtUse(U: BinOp->getOperandUse(i: 0),
1189 /*UndefAllowed=*/false);
1190 ConstantRange RRange = LVI->getConstantRangeAtUse(U: BinOp->getOperandUse(i: 1),
1191 /*UndefAllowed=*/false);
1192
1193 bool Changed = false;
1194 bool NewNUW = false, NewNSW = false;
1195 if (!NUW) {
1196 ConstantRange NUWRange = ConstantRange::makeGuaranteedNoWrapRegion(
1197 BinOp: Opcode, Other: RRange, NoWrapKind: OBO::NoUnsignedWrap);
1198 NewNUW = NUWRange.contains(CR: LRange);
1199 Changed |= NewNUW;
1200 }
1201 if (!NSW) {
1202 ConstantRange NSWRange = ConstantRange::makeGuaranteedNoWrapRegion(
1203 BinOp: Opcode, Other: RRange, NoWrapKind: OBO::NoSignedWrap);
1204 NewNSW = NSWRange.contains(CR: LRange);
1205 Changed |= NewNSW;
1206 }
1207
1208 setDeducedOverflowingFlags(V: BinOp, Opcode, NewNSW, NewNUW);
1209
1210 return Changed;
1211}
1212
1213static bool processAnd(BinaryOperator *BinOp, LazyValueInfo *LVI) {
1214 using namespace llvm::PatternMatch;
1215
1216 // Pattern match (and lhs, C) where C includes a superset of bits which might
1217 // be set in lhs. This is a common truncation idiom created by instcombine.
1218 const Use &LHS = BinOp->getOperandUse(i: 0);
1219 const APInt *RHS;
1220 if (!match(V: BinOp->getOperand(i_nocapture: 1), P: m_LowBitMask(V&: RHS)))
1221 return false;
1222
1223 // We can only replace the AND with LHS based on range info if the range does
1224 // not include undef.
1225 ConstantRange LRange =
1226 LVI->getConstantRangeAtUse(U: LHS, /*UndefAllowed=*/false);
1227 if (!LRange.getUnsignedMax().ule(RHS: *RHS))
1228 return false;
1229
1230 BinOp->replaceAllUsesWith(V: LHS);
1231 BinOp->eraseFromParent();
1232 NumAnd++;
1233 return true;
1234}
1235
1236static bool processTrunc(TruncInst *TI, LazyValueInfo *LVI) {
1237 if (TI->hasNoSignedWrap() && TI->hasNoUnsignedWrap())
1238 return false;
1239
1240 ConstantRange Range =
1241 LVI->getConstantRangeAtUse(U: TI->getOperandUse(i: 0), /*UndefAllowed=*/false);
1242 uint64_t DestWidth = TI->getDestTy()->getScalarSizeInBits();
1243 bool Changed = false;
1244
1245 if (!TI->hasNoUnsignedWrap()) {
1246 if (Range.getActiveBits() <= DestWidth) {
1247 TI->setHasNoUnsignedWrap(true);
1248 ++NumNUW;
1249 Changed = true;
1250 }
1251 }
1252
1253 if (!TI->hasNoSignedWrap()) {
1254 if (Range.getMinSignedBits() <= DestWidth) {
1255 TI->setHasNoSignedWrap(true);
1256 ++NumNSW;
1257 Changed = true;
1258 }
1259 }
1260
1261 return Changed;
1262}
1263
1264static bool runImpl(Function &F, LazyValueInfo *LVI, DominatorTree *DT,
1265 const SimplifyQuery &SQ) {
1266 bool FnChanged = false;
1267 std::optional<ConstantRange> RetRange;
1268 if (F.hasExactDefinition() && F.getReturnType()->isIntOrIntVectorTy())
1269 RetRange =
1270 ConstantRange::getEmpty(BitWidth: F.getReturnType()->getScalarSizeInBits());
1271
1272 // Visiting in a pre-order depth-first traversal causes us to simplify early
1273 // blocks before querying later blocks (which require us to analyze early
1274 // blocks). Eagerly simplifying shallow blocks means there is strictly less
1275 // work to do for deep blocks. This also means we don't visit unreachable
1276 // blocks.
1277 for (BasicBlock *BB : depth_first(G: &F.getEntryBlock())) {
1278 bool BBChanged = false;
1279 for (Instruction &II : llvm::make_early_inc_range(Range&: *BB)) {
1280 switch (II.getOpcode()) {
1281 case Instruction::Select:
1282 BBChanged |= processSelect(S: cast<SelectInst>(Val: &II), LVI);
1283 break;
1284 case Instruction::PHI:
1285 BBChanged |= processPHI(P: cast<PHINode>(Val: &II), LVI, DT, SQ);
1286 break;
1287 case Instruction::ICmp:
1288 case Instruction::FCmp:
1289 BBChanged |= processCmp(Cmp: cast<CmpInst>(Val: &II), LVI);
1290 break;
1291 case Instruction::Call:
1292 case Instruction::Invoke:
1293 BBChanged |= processCallSite(CB&: cast<CallBase>(Val&: II), LVI);
1294 break;
1295 case Instruction::SRem:
1296 case Instruction::SDiv:
1297 BBChanged |= processSDivOrSRem(Instr: cast<BinaryOperator>(Val: &II), LVI);
1298 break;
1299 case Instruction::UDiv:
1300 case Instruction::URem:
1301 BBChanged |= processUDivOrURem(Instr: cast<BinaryOperator>(Val: &II), LVI);
1302 break;
1303 case Instruction::AShr:
1304 BBChanged |= processAShr(SDI: cast<BinaryOperator>(Val: &II), LVI);
1305 break;
1306 case Instruction::SExt:
1307 BBChanged |= processSExt(SDI: cast<SExtInst>(Val: &II), LVI);
1308 break;
1309 case Instruction::ZExt:
1310 BBChanged |= processZExt(ZExt: cast<ZExtInst>(Val: &II), LVI);
1311 break;
1312 case Instruction::UIToFP:
1313 BBChanged |= processUIToFP(UIToFP: cast<UIToFPInst>(Val: &II), LVI);
1314 break;
1315 case Instruction::SIToFP:
1316 BBChanged |= processSIToFP(SIToFP: cast<SIToFPInst>(Val: &II), LVI);
1317 break;
1318 case Instruction::Add:
1319 case Instruction::Sub:
1320 case Instruction::Mul:
1321 case Instruction::Shl:
1322 BBChanged |= processBinOp(BinOp: cast<BinaryOperator>(Val: &II), LVI);
1323 break;
1324 case Instruction::And:
1325 BBChanged |= processAnd(BinOp: cast<BinaryOperator>(Val: &II), LVI);
1326 break;
1327 case Instruction::Trunc:
1328 BBChanged |= processTrunc(TI: cast<TruncInst>(Val: &II), LVI);
1329 break;
1330 }
1331 }
1332
1333 Instruction *Term = BB->getTerminator();
1334 switch (Term->getOpcode()) {
1335 case Instruction::Switch:
1336 BBChanged |= processSwitch(I: cast<SwitchInst>(Val: Term), LVI, DT);
1337 break;
1338 case Instruction::Ret: {
1339 auto *RI = cast<ReturnInst>(Val: Term);
1340 // Try to determine the return value if we can. This is mainly here to
1341 // simplify the writing of unit tests, but also helps to enable IPO by
1342 // constant folding the return values of callees.
1343 auto *RetVal = RI->getReturnValue();
1344 if (!RetVal) break; // handle "ret void"
1345 if (RetRange && !RetRange->isFullSet())
1346 RetRange =
1347 RetRange->unionWith(CR: LVI->getConstantRange(V: RetVal, CxtI: RI,
1348 /*UndefAllowed=*/false));
1349
1350 if (isa<Constant>(Val: RetVal)) break; // nothing to do
1351 if (auto *C = getConstantAt(V: RetVal, At: RI, LVI)) {
1352 ++NumReturns;
1353 RI->replaceUsesOfWith(From: RetVal, To: C);
1354 BBChanged = true;
1355 }
1356 }
1357 }
1358
1359 FnChanged |= BBChanged;
1360 }
1361
1362 // Infer range attribute on return value.
1363 if (RetRange && !RetRange->isFullSet()) {
1364 Attribute RangeAttr = F.getRetAttribute(Kind: Attribute::Range);
1365 if (RangeAttr.isValid())
1366 RetRange = RetRange->intersectWith(CR: RangeAttr.getRange());
1367 // Don't add attribute for constant integer returns to reduce noise. These
1368 // are propagated across functions by IPSCCP.
1369 if (!RetRange->isEmptySet() && !RetRange->isSingleElement()) {
1370 F.addRangeRetAttr(CR: *RetRange);
1371 FnChanged = true;
1372 }
1373 }
1374 return FnChanged;
1375}
1376
1377PreservedAnalyses
1378CorrelatedValuePropagationPass::run(Function &F, FunctionAnalysisManager &AM) {
1379 LazyValueInfo *LVI = &AM.getResult<LazyValueAnalysis>(IR&: F);
1380 DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(IR&: F);
1381
1382 bool Changed = runImpl(F, LVI, DT, SQ: getBestSimplifyQuery(AM, F));
1383
1384 PreservedAnalyses PA;
1385 if (!Changed) {
1386 PA = PreservedAnalyses::all();
1387 } else {
1388#if defined(EXPENSIVE_CHECKS)
1389 assert(DT->verify(DominatorTree::VerificationLevel::Full));
1390#else
1391 assert(DT->verify(DominatorTree::VerificationLevel::Fast));
1392#endif // EXPENSIVE_CHECKS
1393
1394 PA.preserve<DominatorTreeAnalysis>();
1395 PA.preserve<LazyValueAnalysis>();
1396 }
1397
1398 // Keeping LVI alive is expensive, both because it uses a lot of memory, and
1399 // because invalidating values in LVI is expensive. While CVP does preserve
1400 // LVI, we know that passes after JumpThreading+CVP will not need the result
1401 // of this analysis, so we forcefully discard it early.
1402 PA.abandon<LazyValueAnalysis>();
1403 return PA;
1404}
1405