1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/ADT/APFloat.h"
17 #include "llvm/ADT/APInt.h"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/None.h"
20 #include "llvm/ADT/Optional.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/ADT/iterator_range.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/InstructionSimplify.h"
30 #include "llvm/Analysis/Loads.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
33 #include "llvm/Analysis/TargetLibraryInfo.h"
34 #include "llvm/IR/Argument.h"
35 #include "llvm/IR/Attributes.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CallSite.h"
38 #include "llvm/IR/Constant.h"
39 #include "llvm/IR/ConstantRange.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/DiagnosticInfo.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/GetElementPtrTypeIterator.h"
47 #include "llvm/IR/GlobalAlias.h"
48 #include "llvm/IR/GlobalValue.h"
49 #include "llvm/IR/GlobalVariable.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/Intrinsics.h"
55 #include "llvm/IR/LLVMContext.h"
56 #include "llvm/IR/Metadata.h"
57 #include "llvm/IR/Module.h"
58 #include "llvm/IR/Operator.h"
59 #include "llvm/IR/PatternMatch.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/IR/User.h"
62 #include "llvm/IR/Value.h"
63 #include "llvm/Support/Casting.h"
64 #include "llvm/Support/CommandLine.h"
65 #include "llvm/Support/Compiler.h"
66 #include "llvm/Support/ErrorHandling.h"
67 #include "llvm/Support/KnownBits.h"
68 #include "llvm/Support/MathExtras.h"
77 using namespace llvm::PatternMatch;
79 const unsigned MaxDepth = 6;
81 // Controls the number of uses of the value searched for possible
82 // dominating comparisons.
83 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
84 cl::Hidden, cl::init(20));
86 /// Returns the bitwidth of the given scalar or pointer type. For vector types,
87 /// returns the element type's bitwidth.
88 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
89 if (unsigned BitWidth = Ty->getScalarSizeInBits())
92 return DL.getIndexTypeSizeInBits(Ty);
97 // Simplifying using an assume can only be done in a particular control-flow
98 // context (the context instruction provides that context). If an assume and
99 // the context instruction are not in the same block then the DT helps in
100 // figuring out if we can use it.
102 const DataLayout &DL;
104 const Instruction *CxtI;
105 const DominatorTree *DT;
107 // Unlike the other analyses, this may be a nullptr because not all clients
108 // provide it currently.
109 OptimizationRemarkEmitter *ORE;
111 /// Set of assumptions that should be excluded from further queries.
112 /// This is because of the potential for mutual recursion to cause
113 /// computeKnownBits to repeatedly visit the same assume intrinsic. The
114 /// classic case of this is assume(x = y), which will attempt to determine
115 /// bits in x from bits in y, which will attempt to determine bits in y from
116 /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
117 /// isKnownNonZero, which calls computeKnownBits and isKnownToBeAPowerOfTwo
118 /// (all of which can call computeKnownBits), and so on.
119 std::array<const Value *, MaxDepth> Excluded;
121 unsigned NumExcluded = 0;
123 Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
124 const DominatorTree *DT, OptimizationRemarkEmitter *ORE = nullptr)
125 : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE) {}
127 Query(const Query &Q, const Value *NewExcl)
128 : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE),
129 NumExcluded(Q.NumExcluded) {
130 Excluded = Q.Excluded;
131 Excluded[NumExcluded++] = NewExcl;
132 assert(NumExcluded <= Excluded.size());
135 bool isExcluded(const Value *Value) const {
136 if (NumExcluded == 0)
138 auto End = Excluded.begin() + NumExcluded;
139 return std::find(Excluded.begin(), End, Value) != End;
143 } // end anonymous namespace
145 // Given the provided Value and, potentially, a context instruction, return
146 // the preferred context instruction (if any).
147 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
148 // If we've been provided with a context instruction, then use that (provided
149 // it has been inserted).
150 if (CxtI && CxtI->getParent())
153 // If the value is really an already-inserted instruction, then use that.
154 CxtI = dyn_cast<Instruction>(V);
155 if (CxtI && CxtI->getParent())
161 static void computeKnownBits(const Value *V, KnownBits &Known,
162 unsigned Depth, const Query &Q);
164 void llvm::computeKnownBits(const Value *V, KnownBits &Known,
165 const DataLayout &DL, unsigned Depth,
166 AssumptionCache *AC, const Instruction *CxtI,
167 const DominatorTree *DT,
168 OptimizationRemarkEmitter *ORE) {
169 ::computeKnownBits(V, Known, Depth,
170 Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
173 static KnownBits computeKnownBits(const Value *V, unsigned Depth,
176 KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL,
177 unsigned Depth, AssumptionCache *AC,
178 const Instruction *CxtI,
179 const DominatorTree *DT,
180 OptimizationRemarkEmitter *ORE) {
181 return ::computeKnownBits(V, Depth,
182 Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
185 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
186 const DataLayout &DL,
187 AssumptionCache *AC, const Instruction *CxtI,
188 const DominatorTree *DT) {
189 assert(LHS->getType() == RHS->getType() &&
190 "LHS and RHS should have the same type");
191 assert(LHS->getType()->isIntOrIntVectorTy() &&
192 "LHS and RHS should be integers");
193 // Look for an inverted mask: (X & ~M) op (Y & M).
195 if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
196 match(RHS, m_c_And(m_Specific(M), m_Value())))
198 if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
199 match(LHS, m_c_And(m_Specific(M), m_Value())))
201 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
202 KnownBits LHSKnown(IT->getBitWidth());
203 KnownBits RHSKnown(IT->getBitWidth());
204 computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT);
205 computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT);
206 return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
209 bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI) {
210 for (const User *U : CxtI->users()) {
211 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
212 if (IC->isEquality())
213 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
214 if (C->isNullValue())
221 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
224 bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
226 unsigned Depth, AssumptionCache *AC,
227 const Instruction *CxtI,
228 const DominatorTree *DT) {
229 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
230 Query(DL, AC, safeCxtI(V, CxtI), DT));
233 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
235 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
236 AssumptionCache *AC, const Instruction *CxtI,
237 const DominatorTree *DT) {
238 return ::isKnownNonZero(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
241 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
243 AssumptionCache *AC, const Instruction *CxtI,
244 const DominatorTree *DT) {
245 KnownBits Known = computeKnownBits(V, DL, Depth, AC, CxtI, DT);
246 return Known.isNonNegative();
249 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
250 AssumptionCache *AC, const Instruction *CxtI,
251 const DominatorTree *DT) {
252 if (auto *CI = dyn_cast<ConstantInt>(V))
253 return CI->getValue().isStrictlyPositive();
255 // TODO: We'd doing two recursive queries here. We should factor this such
256 // that only a single query is needed.
257 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT) &&
258 isKnownNonZero(V, DL, Depth, AC, CxtI, DT);
261 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
262 AssumptionCache *AC, const Instruction *CxtI,
263 const DominatorTree *DT) {
264 KnownBits Known = computeKnownBits(V, DL, Depth, AC, CxtI, DT);
265 return Known.isNegative();
268 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q);
270 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
271 const DataLayout &DL,
272 AssumptionCache *AC, const Instruction *CxtI,
273 const DominatorTree *DT) {
274 return ::isKnownNonEqual(V1, V2, Query(DL, AC,
275 safeCxtI(V1, safeCxtI(V2, CxtI)),
279 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
282 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
283 const DataLayout &DL,
284 unsigned Depth, AssumptionCache *AC,
285 const Instruction *CxtI, const DominatorTree *DT) {
286 return ::MaskedValueIsZero(V, Mask, Depth,
287 Query(DL, AC, safeCxtI(V, CxtI), DT));
290 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
293 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
294 unsigned Depth, AssumptionCache *AC,
295 const Instruction *CxtI,
296 const DominatorTree *DT) {
297 return ::ComputeNumSignBits(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
300 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
302 KnownBits &KnownOut, KnownBits &Known2,
303 unsigned Depth, const Query &Q) {
304 unsigned BitWidth = KnownOut.getBitWidth();
306 // If an initial sequence of bits in the result is not needed, the
307 // corresponding bits in the operands are not needed.
308 KnownBits LHSKnown(BitWidth);
309 computeKnownBits(Op0, LHSKnown, Depth + 1, Q);
310 computeKnownBits(Op1, Known2, Depth + 1, Q);
312 KnownOut = KnownBits::computeForAddSub(Add, NSW, LHSKnown, Known2);
315 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
316 KnownBits &Known, KnownBits &Known2,
317 unsigned Depth, const Query &Q) {
318 unsigned BitWidth = Known.getBitWidth();
319 computeKnownBits(Op1, Known, Depth + 1, Q);
320 computeKnownBits(Op0, Known2, Depth + 1, Q);
322 bool isKnownNegative = false;
323 bool isKnownNonNegative = false;
324 // If the multiplication is known not to overflow, compute the sign bit.
327 // The product of a number with itself is non-negative.
328 isKnownNonNegative = true;
330 bool isKnownNonNegativeOp1 = Known.isNonNegative();
331 bool isKnownNonNegativeOp0 = Known2.isNonNegative();
332 bool isKnownNegativeOp1 = Known.isNegative();
333 bool isKnownNegativeOp0 = Known2.isNegative();
334 // The product of two numbers with the same sign is non-negative.
335 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
336 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
337 // The product of a negative number and a non-negative number is either
339 if (!isKnownNonNegative)
340 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
341 isKnownNonZero(Op0, Depth, Q)) ||
342 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
343 isKnownNonZero(Op1, Depth, Q));
347 assert(!Known.hasConflict() && !Known2.hasConflict());
348 // Compute a conservative estimate for high known-0 bits.
349 unsigned LeadZ = std::max(Known.countMinLeadingZeros() +
350 Known2.countMinLeadingZeros(),
351 BitWidth) - BitWidth;
352 LeadZ = std::min(LeadZ, BitWidth);
354 // The result of the bottom bits of an integer multiply can be
355 // inferred by looking at the bottom bits of both operands and
356 // multiplying them together.
357 // We can infer at least the minimum number of known trailing bits
358 // of both operands. Depending on number of trailing zeros, we can
359 // infer more bits, because (a*b) <=> ((a/m) * (b/n)) * (m*n) assuming
360 // a and b are divisible by m and n respectively.
361 // We then calculate how many of those bits are inferrable and set
362 // the output. For example, the i8 mul:
365 // We know the bottom 3 bits are zero since the first can be divided by
366 // 4 and the second by 2, thus having ((12/4) * (14/2)) * (2*4).
367 // Applying the multiplication to the trimmed arguments gets:
377 // Which allows us to infer the 2 LSBs. Since we're multiplying the result
378 // by 8, the bottom 3 bits will be 0, so we can infer a total of 5 bits.
379 // The proof for this can be described as:
380 // Pre: (C1 >= 0) && (C1 < (1 << C5)) && (C2 >= 0) && (C2 < (1 << C6)) &&
381 // (C7 == (1 << (umin(countTrailingZeros(C1), C5) +
382 // umin(countTrailingZeros(C2), C6) +
383 // umin(C5 - umin(countTrailingZeros(C1), C5),
384 // C6 - umin(countTrailingZeros(C2), C6)))) - 1)
385 // %aa = shl i8 %a, C5
386 // %bb = shl i8 %b, C6
387 // %aaa = or i8 %aa, C1
388 // %bbb = or i8 %bb, C2
389 // %mul = mul i8 %aaa, %bbb
390 // %mask = and i8 %mul, C7
392 // %mask = i8 ((C1*C2)&C7)
393 // Where C5, C6 describe the known bits of %a, %b
394 // C1, C2 describe the known bottom bits of %a, %b.
395 // C7 describes the mask of the known bits of the result.
396 APInt Bottom0 = Known.One;
397 APInt Bottom1 = Known2.One;
399 // How many times we'd be able to divide each argument by 2 (shr by 1).
400 // This gives us the number of trailing zeros on the multiplication result.
401 unsigned TrailBitsKnown0 = (Known.Zero | Known.One).countTrailingOnes();
402 unsigned TrailBitsKnown1 = (Known2.Zero | Known2.One).countTrailingOnes();
403 unsigned TrailZero0 = Known.countMinTrailingZeros();
404 unsigned TrailZero1 = Known2.countMinTrailingZeros();
405 unsigned TrailZ = TrailZero0 + TrailZero1;
407 // Figure out the fewest known-bits operand.
408 unsigned SmallestOperand = std::min(TrailBitsKnown0 - TrailZero0,
409 TrailBitsKnown1 - TrailZero1);
410 unsigned ResultBitsKnown = std::min(SmallestOperand + TrailZ, BitWidth);
412 APInt BottomKnown = Bottom0.getLoBits(TrailBitsKnown0) *
413 Bottom1.getLoBits(TrailBitsKnown1);
416 Known.Zero.setHighBits(LeadZ);
417 Known.Zero |= (~BottomKnown).getLoBits(ResultBitsKnown);
418 Known.One |= BottomKnown.getLoBits(ResultBitsKnown);
420 // Only make use of no-wrap flags if we failed to compute the sign bit
421 // directly. This matters if the multiplication always overflows, in
422 // which case we prefer to follow the result of the direct computation,
423 // though as the program is invoking undefined behaviour we can choose
424 // whatever we like here.
425 if (isKnownNonNegative && !Known.isNegative())
426 Known.makeNonNegative();
427 else if (isKnownNegative && !Known.isNonNegative())
428 Known.makeNegative();
431 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
433 unsigned BitWidth = Known.getBitWidth();
434 unsigned NumRanges = Ranges.getNumOperands() / 2;
435 assert(NumRanges >= 1);
437 Known.Zero.setAllBits();
438 Known.One.setAllBits();
440 for (unsigned i = 0; i < NumRanges; ++i) {
442 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
444 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
445 ConstantRange Range(Lower->getValue(), Upper->getValue());
447 // The first CommonPrefixBits of all values in Range are equal.
448 unsigned CommonPrefixBits =
449 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
451 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
452 Known.One &= Range.getUnsignedMax() & Mask;
453 Known.Zero &= ~Range.getUnsignedMax() & Mask;
457 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
458 SmallVector<const Value *, 16> WorkSet(1, I);
459 SmallPtrSet<const Value *, 32> Visited;
460 SmallPtrSet<const Value *, 16> EphValues;
462 // The instruction defining an assumption's condition itself is always
463 // considered ephemeral to that assumption (even if it has other
464 // non-ephemeral users). See r246696's test case for an example.
465 if (is_contained(I->operands(), E))
468 while (!WorkSet.empty()) {
469 const Value *V = WorkSet.pop_back_val();
470 if (!Visited.insert(V).second)
473 // If all uses of this value are ephemeral, then so is this value.
474 if (llvm::all_of(V->users(), [&](const User *U) {
475 return EphValues.count(U);
480 if (V == I || isSafeToSpeculativelyExecute(V)) {
482 if (const User *U = dyn_cast<User>(V))
483 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
485 WorkSet.push_back(*J);
493 // Is this an intrinsic that cannot be speculated but also cannot trap?
494 bool llvm::isAssumeLikeIntrinsic(const Instruction *I) {
495 if (const CallInst *CI = dyn_cast<CallInst>(I))
496 if (Function *F = CI->getCalledFunction())
497 switch (F->getIntrinsicID()) {
499 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
500 case Intrinsic::assume:
501 case Intrinsic::sideeffect:
502 case Intrinsic::dbg_declare:
503 case Intrinsic::dbg_value:
504 case Intrinsic::dbg_label:
505 case Intrinsic::invariant_start:
506 case Intrinsic::invariant_end:
507 case Intrinsic::lifetime_start:
508 case Intrinsic::lifetime_end:
509 case Intrinsic::objectsize:
510 case Intrinsic::ptr_annotation:
511 case Intrinsic::var_annotation:
518 bool llvm::isValidAssumeForContext(const Instruction *Inv,
519 const Instruction *CxtI,
520 const DominatorTree *DT) {
521 // There are two restrictions on the use of an assume:
522 // 1. The assume must dominate the context (or the control flow must
523 // reach the assume whenever it reaches the context).
524 // 2. The context must not be in the assume's set of ephemeral values
525 // (otherwise we will use the assume to prove that the condition
526 // feeding the assume is trivially true, thus causing the removal of
530 if (DT->dominates(Inv, CxtI))
532 } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
533 // We don't have a DT, but this trivially dominates.
537 // With or without a DT, the only remaining case we will check is if the
538 // instructions are in the same BB. Give up if that is not the case.
539 if (Inv->getParent() != CxtI->getParent())
542 // If we have a dom tree, then we now know that the assume doesn't dominate
543 // the other instruction. If we don't have a dom tree then we can check if
544 // the assume is first in the BB.
546 // Search forward from the assume until we reach the context (or the end
547 // of the block); the common case is that the assume will come first.
548 for (auto I = std::next(BasicBlock::const_iterator(Inv)),
549 IE = Inv->getParent()->end(); I != IE; ++I)
554 // The context comes first, but they're both in the same block. Make sure
555 // there is nothing in between that might interrupt the control flow.
556 for (BasicBlock::const_iterator I =
557 std::next(BasicBlock::const_iterator(CxtI)), IE(Inv);
559 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
562 return !isEphemeralValueOf(Inv, CxtI);
565 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
566 unsigned Depth, const Query &Q) {
567 // Use of assumptions is context-sensitive. If we don't have a context, we
569 if (!Q.AC || !Q.CxtI)
572 unsigned BitWidth = Known.getBitWidth();
574 // Note that the patterns below need to be kept in sync with the code
575 // in AssumptionCache::updateAffectedValues.
577 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
580 CallInst *I = cast<CallInst>(AssumeVH);
581 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
582 "Got assumption for the wrong function!");
586 // Warning: This loop can end up being somewhat performance sensitive.
587 // We're running this loop for once for each value queried resulting in a
588 // runtime of ~O(#assumes * #values).
590 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
591 "must be an assume intrinsic");
593 Value *Arg = I->getArgOperand(0);
595 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
596 assert(BitWidth == 1 && "assume operand is not i1?");
600 if (match(Arg, m_Not(m_Specific(V))) &&
601 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
602 assert(BitWidth == 1 && "assume operand is not i1?");
607 // The remaining tests are all recursive, so bail out if we hit the limit.
608 if (Depth == MaxDepth)
612 auto m_V = m_CombineOr(m_Specific(V),
613 m_CombineOr(m_PtrToInt(m_Specific(V)),
614 m_BitCast(m_Specific(V))));
616 CmpInst::Predicate Pred;
619 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
620 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
621 KnownBits RHSKnown(BitWidth);
622 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
623 Known.Zero |= RHSKnown.Zero;
624 Known.One |= RHSKnown.One;
626 } else if (match(Arg,
627 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
628 Pred == ICmpInst::ICMP_EQ &&
629 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
630 KnownBits RHSKnown(BitWidth);
631 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
632 KnownBits MaskKnown(BitWidth);
633 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
635 // For those bits in the mask that are known to be one, we can propagate
636 // known bits from the RHS to V.
637 Known.Zero |= RHSKnown.Zero & MaskKnown.One;
638 Known.One |= RHSKnown.One & MaskKnown.One;
639 // assume(~(v & b) = a)
640 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
642 Pred == ICmpInst::ICMP_EQ &&
643 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
644 KnownBits RHSKnown(BitWidth);
645 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
646 KnownBits MaskKnown(BitWidth);
647 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
649 // For those bits in the mask that are known to be one, we can propagate
650 // inverted known bits from the RHS to V.
651 Known.Zero |= RHSKnown.One & MaskKnown.One;
652 Known.One |= RHSKnown.Zero & MaskKnown.One;
654 } else if (match(Arg,
655 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
656 Pred == ICmpInst::ICMP_EQ &&
657 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
658 KnownBits RHSKnown(BitWidth);
659 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
660 KnownBits BKnown(BitWidth);
661 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
663 // For those bits in B that are known to be zero, we can propagate known
664 // bits from the RHS to V.
665 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
666 Known.One |= RHSKnown.One & BKnown.Zero;
667 // assume(~(v | b) = a)
668 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
670 Pred == ICmpInst::ICMP_EQ &&
671 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
672 KnownBits RHSKnown(BitWidth);
673 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
674 KnownBits BKnown(BitWidth);
675 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
677 // For those bits in B that are known to be zero, we can propagate
678 // inverted known bits from the RHS to V.
679 Known.Zero |= RHSKnown.One & BKnown.Zero;
680 Known.One |= RHSKnown.Zero & BKnown.Zero;
682 } else if (match(Arg,
683 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
684 Pred == ICmpInst::ICMP_EQ &&
685 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
686 KnownBits RHSKnown(BitWidth);
687 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
688 KnownBits BKnown(BitWidth);
689 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
691 // For those bits in B that are known to be zero, we can propagate known
692 // bits from the RHS to V. For those bits in B that are known to be one,
693 // we can propagate inverted known bits from the RHS to V.
694 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
695 Known.One |= RHSKnown.One & BKnown.Zero;
696 Known.Zero |= RHSKnown.One & BKnown.One;
697 Known.One |= RHSKnown.Zero & BKnown.One;
698 // assume(~(v ^ b) = a)
699 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
701 Pred == ICmpInst::ICMP_EQ &&
702 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
703 KnownBits RHSKnown(BitWidth);
704 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
705 KnownBits BKnown(BitWidth);
706 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
708 // For those bits in B that are known to be zero, we can propagate
709 // inverted known bits from the RHS to V. For those bits in B that are
710 // known to be one, we can propagate known bits from the RHS to V.
711 Known.Zero |= RHSKnown.One & BKnown.Zero;
712 Known.One |= RHSKnown.Zero & BKnown.Zero;
713 Known.Zero |= RHSKnown.Zero & BKnown.One;
714 Known.One |= RHSKnown.One & BKnown.One;
715 // assume(v << c = a)
716 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
718 Pred == ICmpInst::ICMP_EQ &&
719 isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
721 KnownBits RHSKnown(BitWidth);
722 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
723 // For those bits in RHS that are known, we can propagate them to known
724 // bits in V shifted to the right by C.
725 RHSKnown.Zero.lshrInPlace(C);
726 Known.Zero |= RHSKnown.Zero;
727 RHSKnown.One.lshrInPlace(C);
728 Known.One |= RHSKnown.One;
729 // assume(~(v << c) = a)
730 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
732 Pred == ICmpInst::ICMP_EQ &&
733 isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
735 KnownBits RHSKnown(BitWidth);
736 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
737 // For those bits in RHS that are known, we can propagate them inverted
738 // to known bits in V shifted to the right by C.
739 RHSKnown.One.lshrInPlace(C);
740 Known.Zero |= RHSKnown.One;
741 RHSKnown.Zero.lshrInPlace(C);
742 Known.One |= RHSKnown.Zero;
743 // assume(v >> c = a)
744 } else if (match(Arg,
745 m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)),
747 Pred == ICmpInst::ICMP_EQ &&
748 isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
750 KnownBits RHSKnown(BitWidth);
751 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
752 // For those bits in RHS that are known, we can propagate them to known
753 // bits in V shifted to the right by C.
754 Known.Zero |= RHSKnown.Zero << C;
755 Known.One |= RHSKnown.One << C;
756 // assume(~(v >> c) = a)
757 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))),
759 Pred == ICmpInst::ICMP_EQ &&
760 isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
762 KnownBits RHSKnown(BitWidth);
763 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
764 // For those bits in RHS that are known, we can propagate them inverted
765 // to known bits in V shifted to the right by C.
766 Known.Zero |= RHSKnown.One << C;
767 Known.One |= RHSKnown.Zero << C;
768 // assume(v >=_s c) where c is non-negative
769 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
770 Pred == ICmpInst::ICMP_SGE &&
771 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
772 KnownBits RHSKnown(BitWidth);
773 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
775 if (RHSKnown.isNonNegative()) {
776 // We know that the sign bit is zero.
777 Known.makeNonNegative();
779 // assume(v >_s c) where c is at least -1.
780 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
781 Pred == ICmpInst::ICMP_SGT &&
782 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
783 KnownBits RHSKnown(BitWidth);
784 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
786 if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
787 // We know that the sign bit is zero.
788 Known.makeNonNegative();
790 // assume(v <=_s c) where c is negative
791 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
792 Pred == ICmpInst::ICMP_SLE &&
793 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
794 KnownBits RHSKnown(BitWidth);
795 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
797 if (RHSKnown.isNegative()) {
798 // We know that the sign bit is one.
799 Known.makeNegative();
801 // assume(v <_s c) where c is non-positive
802 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
803 Pred == ICmpInst::ICMP_SLT &&
804 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
805 KnownBits RHSKnown(BitWidth);
806 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
808 if (RHSKnown.isZero() || RHSKnown.isNegative()) {
809 // We know that the sign bit is one.
810 Known.makeNegative();
813 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
814 Pred == ICmpInst::ICMP_ULE &&
815 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
816 KnownBits RHSKnown(BitWidth);
817 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
819 // Whatever high bits in c are zero are known to be zero.
820 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
822 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
823 Pred == ICmpInst::ICMP_ULT &&
824 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
825 KnownBits RHSKnown(BitWidth);
826 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
828 // If the RHS is known zero, then this assumption must be wrong (nothing
829 // is unsigned less than zero). Signal a conflict and get out of here.
830 if (RHSKnown.isZero()) {
831 Known.Zero.setAllBits();
832 Known.One.setAllBits();
836 // Whatever high bits in c are zero are known to be zero (if c is a power
837 // of 2, then one more).
838 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
839 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
841 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
845 // If assumptions conflict with each other or previous known bits, then we
846 // have a logical fallacy. It's possible that the assumption is not reachable,
847 // so this isn't a real bug. On the other hand, the program may have undefined
848 // behavior, or we might have a bug in the compiler. We can't assert/crash, so
849 // clear out the known bits, try to warn the user, and hope for the best.
850 if (Known.Zero.intersects(Known.One)) {
855 auto *CxtI = const_cast<Instruction *>(Q.CxtI);
856 return OptimizationRemarkAnalysis("value-tracking", "BadAssumption",
858 << "Detected conflicting code assumptions. Program may "
859 "have undefined behavior, or compiler may have "
865 /// Compute known bits from a shift operator, including those with a
866 /// non-constant shift amount. Known is the output of this function. Known2 is a
867 /// pre-allocated temporary with the same bit width as Known. KZF and KOF are
868 /// operator-specific functions that, given the known-zero or known-one bits
869 /// respectively, and a shift amount, compute the implied known-zero or
870 /// known-one bits of the shift operator's result respectively for that shift
871 /// amount. The results from calling KZF and KOF are conservatively combined for
872 /// all permitted shift amounts.
873 static void computeKnownBitsFromShiftOperator(
874 const Operator *I, KnownBits &Known, KnownBits &Known2,
875 unsigned Depth, const Query &Q,
876 function_ref<APInt(const APInt &, unsigned)> KZF,
877 function_ref<APInt(const APInt &, unsigned)> KOF) {
878 unsigned BitWidth = Known.getBitWidth();
880 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
881 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
883 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
884 Known.Zero = KZF(Known.Zero, ShiftAmt);
885 Known.One = KOF(Known.One, ShiftAmt);
886 // If the known bits conflict, this must be an overflowing left shift, so
887 // the shift result is poison. We can return anything we want. Choose 0 for
888 // the best folding opportunity.
889 if (Known.hasConflict())
895 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
897 // If the shift amount could be greater than or equal to the bit-width of the
898 // LHS, the value could be poison, but bail out because the check below is
899 // expensive. TODO: Should we just carry on?
900 if ((~Known.Zero).uge(BitWidth)) {
905 // Note: We cannot use Known.Zero.getLimitedValue() here, because if
906 // BitWidth > 64 and any upper bits are known, we'll end up returning the
907 // limit value (which implies all bits are known).
908 uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
909 uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
911 // It would be more-clearly correct to use the two temporaries for this
912 // calculation. Reusing the APInts here to prevent unnecessary allocations.
915 // If we know the shifter operand is nonzero, we can sometimes infer more
916 // known bits. However this is expensive to compute, so be lazy about it and
917 // only compute it when absolutely necessary.
918 Optional<bool> ShifterOperandIsNonZero;
920 // Early exit if we can't constrain any well-defined shift amount.
921 if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) &&
922 !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) {
923 ShifterOperandIsNonZero = isKnownNonZero(I->getOperand(1), Depth + 1, Q);
924 if (!*ShifterOperandIsNonZero)
928 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
930 Known.Zero.setAllBits();
931 Known.One.setAllBits();
932 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
933 // Combine the shifted known input bits only for those shift amounts
934 // compatible with its known constraints.
935 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
937 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
939 // If we know the shifter is nonzero, we may be able to infer more known
940 // bits. This check is sunk down as far as possible to avoid the expensive
941 // call to isKnownNonZero if the cheaper checks above fail.
943 if (!ShifterOperandIsNonZero.hasValue())
944 ShifterOperandIsNonZero =
945 isKnownNonZero(I->getOperand(1), Depth + 1, Q);
946 if (*ShifterOperandIsNonZero)
950 Known.Zero &= KZF(Known2.Zero, ShiftAmt);
951 Known.One &= KOF(Known2.One, ShiftAmt);
954 // If the known bits conflict, the result is poison. Return a 0 and hope the
955 // caller can further optimize that.
956 if (Known.hasConflict())
960 static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
961 unsigned Depth, const Query &Q) {
962 unsigned BitWidth = Known.getBitWidth();
964 KnownBits Known2(Known);
965 switch (I->getOpcode()) {
967 case Instruction::Load:
968 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
969 computeKnownBitsFromRangeMetadata(*MD, Known);
971 case Instruction::And: {
972 // If either the LHS or the RHS are Zero, the result is zero.
973 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
974 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
976 // Output known-1 bits are only known if set in both the LHS & RHS.
977 Known.One &= Known2.One;
978 // Output known-0 are known to be clear if zero in either the LHS | RHS.
979 Known.Zero |= Known2.Zero;
981 // and(x, add (x, -1)) is a common idiom that always clears the low bit;
982 // here we handle the more general case of adding any odd number by
983 // matching the form add(x, add(x, y)) where y is odd.
984 // TODO: This could be generalized to clearing any bit set in y where the
985 // following bit is known to be unset in y.
986 Value *X = nullptr, *Y = nullptr;
987 if (!Known.Zero[0] && !Known.One[0] &&
988 match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) {
990 computeKnownBits(Y, Known2, Depth + 1, Q);
991 if (Known2.countMinTrailingOnes() > 0)
992 Known.Zero.setBit(0);
996 case Instruction::Or:
997 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
998 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1000 // Output known-0 bits are only known if clear in both the LHS & RHS.
1001 Known.Zero &= Known2.Zero;
1002 // Output known-1 are known to be set if set in either the LHS | RHS.
1003 Known.One |= Known2.One;
1005 case Instruction::Xor: {
1006 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1007 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1009 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1010 APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
1011 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1012 Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
1013 Known.Zero = std::move(KnownZeroOut);
1016 case Instruction::Mul: {
1017 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1018 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
1022 case Instruction::UDiv: {
1023 // For the purposes of computing leading zeros we can conservatively
1024 // treat a udiv as a logical right shift by the power of 2 known to
1025 // be less than the denominator.
1026 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1027 unsigned LeadZ = Known2.countMinLeadingZeros();
1030 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1031 unsigned RHSMaxLeadingZeros = Known2.countMaxLeadingZeros();
1032 if (RHSMaxLeadingZeros != BitWidth)
1033 LeadZ = std::min(BitWidth, LeadZ + BitWidth - RHSMaxLeadingZeros - 1);
1035 Known.Zero.setHighBits(LeadZ);
1038 case Instruction::Select: {
1039 const Value *LHS, *RHS;
1040 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
1041 if (SelectPatternResult::isMinOrMax(SPF)) {
1042 computeKnownBits(RHS, Known, Depth + 1, Q);
1043 computeKnownBits(LHS, Known2, Depth + 1, Q);
1045 computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
1046 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1049 unsigned MaxHighOnes = 0;
1050 unsigned MaxHighZeros = 0;
1051 if (SPF == SPF_SMAX) {
1052 // If both sides are negative, the result is negative.
1053 if (Known.isNegative() && Known2.isNegative())
1054 // We can derive a lower bound on the result by taking the max of the
1055 // leading one bits.
1057 std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
1058 // If either side is non-negative, the result is non-negative.
1059 else if (Known.isNonNegative() || Known2.isNonNegative())
1061 } else if (SPF == SPF_SMIN) {
1062 // If both sides are non-negative, the result is non-negative.
1063 if (Known.isNonNegative() && Known2.isNonNegative())
1064 // We can derive an upper bound on the result by taking the max of the
1065 // leading zero bits.
1066 MaxHighZeros = std::max(Known.countMinLeadingZeros(),
1067 Known2.countMinLeadingZeros());
1068 // If either side is negative, the result is negative.
1069 else if (Known.isNegative() || Known2.isNegative())
1071 } else if (SPF == SPF_UMAX) {
1072 // We can derive a lower bound on the result by taking the max of the
1073 // leading one bits.
1075 std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
1076 } else if (SPF == SPF_UMIN) {
1077 // We can derive an upper bound on the result by taking the max of the
1078 // leading zero bits.
1080 std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
1081 } else if (SPF == SPF_ABS) {
1082 // RHS from matchSelectPattern returns the negation part of abs pattern.
1083 // If the negate has an NSW flag we can assume the sign bit of the result
1084 // will be 0 because that makes abs(INT_MIN) undefined.
1085 if (cast<Instruction>(RHS)->hasNoSignedWrap())
1089 // Only known if known in both the LHS and RHS.
1090 Known.One &= Known2.One;
1091 Known.Zero &= Known2.Zero;
1092 if (MaxHighOnes > 0)
1093 Known.One.setHighBits(MaxHighOnes);
1094 if (MaxHighZeros > 0)
1095 Known.Zero.setHighBits(MaxHighZeros);
1098 case Instruction::FPTrunc:
1099 case Instruction::FPExt:
1100 case Instruction::FPToUI:
1101 case Instruction::FPToSI:
1102 case Instruction::SIToFP:
1103 case Instruction::UIToFP:
1104 break; // Can't work with floating point.
1105 case Instruction::PtrToInt:
1106 case Instruction::IntToPtr:
1107 // Fall through and handle them the same as zext/trunc.
1109 case Instruction::ZExt:
1110 case Instruction::Trunc: {
1111 Type *SrcTy = I->getOperand(0)->getType();
1113 unsigned SrcBitWidth;
1114 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1115 // which fall through here.
1116 Type *ScalarTy = SrcTy->getScalarType();
1117 SrcBitWidth = ScalarTy->isPointerTy() ?
1118 Q.DL.getIndexTypeSizeInBits(ScalarTy) :
1119 Q.DL.getTypeSizeInBits(ScalarTy);
1121 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1122 Known = Known.zextOrTrunc(SrcBitWidth);
1123 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1124 Known = Known.zextOrTrunc(BitWidth);
1125 // Any top bits are known to be zero.
1126 if (BitWidth > SrcBitWidth)
1127 Known.Zero.setBitsFrom(SrcBitWidth);
1130 case Instruction::BitCast: {
1131 Type *SrcTy = I->getOperand(0)->getType();
1132 if (SrcTy->isIntOrPtrTy() &&
1133 // TODO: For now, not handling conversions like:
1134 // (bitcast i64 %x to <2 x i32>)
1135 !I->getType()->isVectorTy()) {
1136 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1141 case Instruction::SExt: {
1142 // Compute the bits in the result that are not present in the input.
1143 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1145 Known = Known.trunc(SrcBitWidth);
1146 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1147 // If the sign bit of the input is known set or clear, then we know the
1148 // top bits of the result.
1149 Known = Known.sext(BitWidth);
1152 case Instruction::Shl: {
1153 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1154 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1155 auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
1156 APInt KZResult = KnownZero << ShiftAmt;
1157 KZResult.setLowBits(ShiftAmt); // Low bits known 0.
1158 // If this shift has "nsw" keyword, then the result is either a poison
1159 // value or has the same sign bit as the first operand.
1160 if (NSW && KnownZero.isSignBitSet())
1161 KZResult.setSignBit();
1165 auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
1166 APInt KOResult = KnownOne << ShiftAmt;
1167 if (NSW && KnownOne.isSignBitSet())
1168 KOResult.setSignBit();
1172 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1175 case Instruction::LShr: {
1176 // (lshr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1177 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1178 APInt KZResult = KnownZero.lshr(ShiftAmt);
1179 // High bits known zero.
1180 KZResult.setHighBits(ShiftAmt);
1184 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1185 return KnownOne.lshr(ShiftAmt);
1188 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1191 case Instruction::AShr: {
1192 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1193 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1194 return KnownZero.ashr(ShiftAmt);
1197 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1198 return KnownOne.ashr(ShiftAmt);
1201 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1204 case Instruction::Sub: {
1205 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1206 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1207 Known, Known2, Depth, Q);
1210 case Instruction::Add: {
1211 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1212 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1213 Known, Known2, Depth, Q);
1216 case Instruction::SRem:
1217 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1218 APInt RA = Rem->getValue().abs();
1219 if (RA.isPowerOf2()) {
1220 APInt LowBits = RA - 1;
1221 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1223 // The low bits of the first operand are unchanged by the srem.
1224 Known.Zero = Known2.Zero & LowBits;
1225 Known.One = Known2.One & LowBits;
1227 // If the first operand is non-negative or has all low bits zero, then
1228 // the upper bits are all zero.
1229 if (Known2.isNonNegative() || LowBits.isSubsetOf(Known2.Zero))
1230 Known.Zero |= ~LowBits;
1232 // If the first operand is negative and not all low bits are zero, then
1233 // the upper bits are all one.
1234 if (Known2.isNegative() && LowBits.intersects(Known2.One))
1235 Known.One |= ~LowBits;
1237 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1242 // The sign bit is the LHS's sign bit, except when the result of the
1243 // remainder is zero.
1244 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1245 // If it's known zero, our sign bit is also zero.
1246 if (Known2.isNonNegative())
1247 Known.makeNonNegative();
1250 case Instruction::URem: {
1251 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1252 const APInt &RA = Rem->getValue();
1253 if (RA.isPowerOf2()) {
1254 APInt LowBits = (RA - 1);
1255 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1256 Known.Zero |= ~LowBits;
1257 Known.One &= LowBits;
1262 // Since the result is less than or equal to either operand, any leading
1263 // zero bits in either operand must also exist in the result.
1264 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1265 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1268 std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
1270 Known.Zero.setHighBits(Leaders);
1274 case Instruction::Alloca: {
1275 const AllocaInst *AI = cast<AllocaInst>(I);
1276 unsigned Align = AI->getAlignment();
1278 Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1281 Known.Zero.setLowBits(countTrailingZeros(Align));
1284 case Instruction::GetElementPtr: {
1285 // Analyze all of the subscripts of this getelementptr instruction
1286 // to determine if we can prove known low zero bits.
1287 KnownBits LocalKnown(BitWidth);
1288 computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
1289 unsigned TrailZ = LocalKnown.countMinTrailingZeros();
1291 gep_type_iterator GTI = gep_type_begin(I);
1292 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1293 Value *Index = I->getOperand(i);
1294 if (StructType *STy = GTI.getStructTypeOrNull()) {
1295 // Handle struct member offset arithmetic.
1297 // Handle case when index is vector zeroinitializer
1298 Constant *CIndex = cast<Constant>(Index);
1299 if (CIndex->isZeroValue())
1302 if (CIndex->getType()->isVectorTy())
1303 Index = CIndex->getSplatValue();
1305 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1306 const StructLayout *SL = Q.DL.getStructLayout(STy);
1307 uint64_t Offset = SL->getElementOffset(Idx);
1308 TrailZ = std::min<unsigned>(TrailZ,
1309 countTrailingZeros(Offset));
1311 // Handle array index arithmetic.
1312 Type *IndexedTy = GTI.getIndexedType();
1313 if (!IndexedTy->isSized()) {
1317 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1318 uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1319 LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
1320 computeKnownBits(Index, LocalKnown, Depth + 1, Q);
1321 TrailZ = std::min(TrailZ,
1322 unsigned(countTrailingZeros(TypeSize) +
1323 LocalKnown.countMinTrailingZeros()));
1327 Known.Zero.setLowBits(TrailZ);
1330 case Instruction::PHI: {
1331 const PHINode *P = cast<PHINode>(I);
1332 // Handle the case of a simple two-predecessor recurrence PHI.
1333 // There's a lot more that could theoretically be done here, but
1334 // this is sufficient to catch some interesting cases.
1335 if (P->getNumIncomingValues() == 2) {
1336 for (unsigned i = 0; i != 2; ++i) {
1337 Value *L = P->getIncomingValue(i);
1338 Value *R = P->getIncomingValue(!i);
1339 Operator *LU = dyn_cast<Operator>(L);
1342 unsigned Opcode = LU->getOpcode();
1343 // Check for operations that have the property that if
1344 // both their operands have low zero bits, the result
1345 // will have low zero bits.
1346 if (Opcode == Instruction::Add ||
1347 Opcode == Instruction::Sub ||
1348 Opcode == Instruction::And ||
1349 Opcode == Instruction::Or ||
1350 Opcode == Instruction::Mul) {
1351 Value *LL = LU->getOperand(0);
1352 Value *LR = LU->getOperand(1);
1353 // Find a recurrence.
1360 // Ok, we have a PHI of the form L op= R. Check for low
1362 computeKnownBits(R, Known2, Depth + 1, Q);
1364 // We need to take the minimum number of known bits
1365 KnownBits Known3(Known);
1366 computeKnownBits(L, Known3, Depth + 1, Q);
1368 Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(),
1369 Known3.countMinTrailingZeros()));
1371 auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
1372 if (OverflowOp && OverflowOp->hasNoSignedWrap()) {
1373 // If initial value of recurrence is nonnegative, and we are adding
1374 // a nonnegative number with nsw, the result can only be nonnegative
1375 // or poison value regardless of the number of times we execute the
1376 // add in phi recurrence. If initial value is negative and we are
1377 // adding a negative number with nsw, the result can only be
1378 // negative or poison value. Similar arguments apply to sub and mul.
1380 // (add non-negative, non-negative) --> non-negative
1381 // (add negative, negative) --> negative
1382 if (Opcode == Instruction::Add) {
1383 if (Known2.isNonNegative() && Known3.isNonNegative())
1384 Known.makeNonNegative();
1385 else if (Known2.isNegative() && Known3.isNegative())
1386 Known.makeNegative();
1389 // (sub nsw non-negative, negative) --> non-negative
1390 // (sub nsw negative, non-negative) --> negative
1391 else if (Opcode == Instruction::Sub && LL == I) {
1392 if (Known2.isNonNegative() && Known3.isNegative())
1393 Known.makeNonNegative();
1394 else if (Known2.isNegative() && Known3.isNonNegative())
1395 Known.makeNegative();
1398 // (mul nsw non-negative, non-negative) --> non-negative
1399 else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1400 Known3.isNonNegative())
1401 Known.makeNonNegative();
1409 // Unreachable blocks may have zero-operand PHI nodes.
1410 if (P->getNumIncomingValues() == 0)
1413 // Otherwise take the unions of the known bit sets of the operands,
1414 // taking conservative care to avoid excessive recursion.
1415 if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
1416 // Skip if every incoming value references to ourself.
1417 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1420 Known.Zero.setAllBits();
1421 Known.One.setAllBits();
1422 for (Value *IncValue : P->incoming_values()) {
1423 // Skip direct self references.
1424 if (IncValue == P) continue;
1426 Known2 = KnownBits(BitWidth);
1427 // Recurse, but cap the recursion to one level, because we don't
1428 // want to waste time spinning around in loops.
1429 computeKnownBits(IncValue, Known2, MaxDepth - 1, Q);
1430 Known.Zero &= Known2.Zero;
1431 Known.One &= Known2.One;
1432 // If all bits have been ruled out, there's no need to check
1434 if (!Known.Zero && !Known.One)
1440 case Instruction::Call:
1441 case Instruction::Invoke:
1442 // If range metadata is attached to this call, set known bits from that,
1443 // and then intersect with known bits based on other properties of the
1445 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1446 computeKnownBitsFromRangeMetadata(*MD, Known);
1447 if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
1448 computeKnownBits(RV, Known2, Depth + 1, Q);
1449 Known.Zero |= Known2.Zero;
1450 Known.One |= Known2.One;
1452 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1453 switch (II->getIntrinsicID()) {
1455 case Intrinsic::bitreverse:
1456 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1457 Known.Zero |= Known2.Zero.reverseBits();
1458 Known.One |= Known2.One.reverseBits();
1460 case Intrinsic::bswap:
1461 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1462 Known.Zero |= Known2.Zero.byteSwap();
1463 Known.One |= Known2.One.byteSwap();
1465 case Intrinsic::ctlz: {
1466 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1467 // If we have a known 1, its position is our upper bound.
1468 unsigned PossibleLZ = Known2.One.countLeadingZeros();
1469 // If this call is undefined for 0, the result will be less than 2^n.
1470 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1471 PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
1472 unsigned LowBits = Log2_32(PossibleLZ)+1;
1473 Known.Zero.setBitsFrom(LowBits);
1476 case Intrinsic::cttz: {
1477 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1478 // If we have a known 1, its position is our upper bound.
1479 unsigned PossibleTZ = Known2.One.countTrailingZeros();
1480 // If this call is undefined for 0, the result will be less than 2^n.
1481 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1482 PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
1483 unsigned LowBits = Log2_32(PossibleTZ)+1;
1484 Known.Zero.setBitsFrom(LowBits);
1487 case Intrinsic::ctpop: {
1488 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1489 // We can bound the space the count needs. Also, bits known to be zero
1490 // can't contribute to the population.
1491 unsigned BitsPossiblySet = Known2.countMaxPopulation();
1492 unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1493 Known.Zero.setBitsFrom(LowBits);
1494 // TODO: we could bound KnownOne using the lower bound on the number
1495 // of bits which might be set provided by popcnt KnownOne2.
1498 case Intrinsic::x86_sse42_crc32_64_64:
1499 Known.Zero.setBitsFrom(32);
1504 case Instruction::ExtractElement:
1505 // Look through extract element. At the moment we keep this simple and skip
1506 // tracking the specific element. But at least we might find information
1507 // valid for all elements of the vector (for example if vector is sign
1508 // extended, shifted, etc).
1509 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1511 case Instruction::ExtractValue:
1512 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1513 const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1514 if (EVI->getNumIndices() != 1) break;
1515 if (EVI->getIndices()[0] == 0) {
1516 switch (II->getIntrinsicID()) {
1518 case Intrinsic::uadd_with_overflow:
1519 case Intrinsic::sadd_with_overflow:
1520 computeKnownBitsAddSub(true, II->getArgOperand(0),
1521 II->getArgOperand(1), false, Known, Known2,
1524 case Intrinsic::usub_with_overflow:
1525 case Intrinsic::ssub_with_overflow:
1526 computeKnownBitsAddSub(false, II->getArgOperand(0),
1527 II->getArgOperand(1), false, Known, Known2,
1530 case Intrinsic::umul_with_overflow:
1531 case Intrinsic::smul_with_overflow:
1532 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1533 Known, Known2, Depth, Q);
1541 /// Determine which bits of V are known to be either zero or one and return
1543 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1544 KnownBits Known(getBitWidth(V->getType(), Q.DL));
1545 computeKnownBits(V, Known, Depth, Q);
1549 /// Determine which bits of V are known to be either zero or one and return
1550 /// them in the Known bit set.
1552 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1553 /// we cannot optimize based on the assumption that it is zero without changing
1554 /// it to be an explicit zero. If we don't change it to zero, other code could
1555 /// optimized based on the contradictory assumption that it is non-zero.
1556 /// Because instcombine aggressively folds operations with undef args anyway,
1557 /// this won't lose us code quality.
1559 /// This function is defined on values with integer type, values with pointer
1560 /// type, and vectors of integers. In the case
1561 /// where V is a vector, known zero, and known one values are the
1562 /// same width as the vector element, and the bit is set only if it is true
1563 /// for all of the elements in the vector.
1564 void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1566 assert(V && "No Value?");
1567 assert(Depth <= MaxDepth && "Limit Search Depth");
1568 unsigned BitWidth = Known.getBitWidth();
1570 assert((V->getType()->isIntOrIntVectorTy(BitWidth) ||
1571 V->getType()->isPtrOrPtrVectorTy()) &&
1572 "Not integer or pointer type!");
1574 Type *ScalarTy = V->getType()->getScalarType();
1575 unsigned ExpectedWidth = ScalarTy->isPointerTy() ?
1576 Q.DL.getIndexTypeSizeInBits(ScalarTy) : Q.DL.getTypeSizeInBits(ScalarTy);
1577 assert(ExpectedWidth == BitWidth && "V and Known should have same BitWidth");
1579 (void)ExpectedWidth;
1582 if (match(V, m_APInt(C))) {
1583 // We know all of the bits for a scalar constant or a splat vector constant!
1585 Known.Zero = ~Known.One;
1588 // Null and aggregate-zero are all-zeros.
1589 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1593 // Handle a constant vector by taking the intersection of the known bits of
1595 if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1596 // We know that CDS must be a vector of integers. Take the intersection of
1598 Known.Zero.setAllBits(); Known.One.setAllBits();
1599 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1600 APInt Elt = CDS->getElementAsAPInt(i);
1607 if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1608 // We know that CV must be a vector of integers. Take the intersection of
1610 Known.Zero.setAllBits(); Known.One.setAllBits();
1611 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1612 Constant *Element = CV->getAggregateElement(i);
1613 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1618 const APInt &Elt = ElementCI->getValue();
1625 // Start out not knowing anything.
1628 // We can't imply anything about undefs.
1629 if (isa<UndefValue>(V))
1632 // There's no point in looking through other users of ConstantData for
1633 // assumptions. Confirm that we've handled them all.
1634 assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1636 // Limit search depth.
1637 // All recursive calls that increase depth must come after this.
1638 if (Depth == MaxDepth)
1641 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1642 // the bits of its aliasee.
1643 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1644 if (!GA->isInterposable())
1645 computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1649 if (const Operator *I = dyn_cast<Operator>(V))
1650 computeKnownBitsFromOperator(I, Known, Depth, Q);
1652 // Aligned pointers have trailing zeros - refine Known.Zero set
1653 if (V->getType()->isPointerTy()) {
1654 unsigned Align = V->getPointerAlignment(Q.DL);
1656 Known.Zero.setLowBits(countTrailingZeros(Align));
1659 // computeKnownBitsFromAssume strictly refines Known.
1660 // Therefore, we run them after computeKnownBitsFromOperator.
1662 // Check whether a nearby assume intrinsic can determine some known bits.
1663 computeKnownBitsFromAssume(V, Known, Depth, Q);
1665 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1668 /// Return true if the given value is known to have exactly one
1669 /// bit set when defined. For vectors return true if every element is known to
1670 /// be a power of two when defined. Supports values with integer or pointer
1671 /// types and vectors of integers.
1672 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1674 assert(Depth <= MaxDepth && "Limit Search Depth");
1676 // Attempt to match against constants.
1677 if (OrZero && match(V, m_Power2OrZero()))
1679 if (match(V, m_Power2()))
1682 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1683 // it is shifted off the end then the result is undefined.
1684 if (match(V, m_Shl(m_One(), m_Value())))
1687 // (signmask) >>l X is clearly a power of two if the one is not shifted off
1688 // the bottom. If it is shifted off the bottom then the result is undefined.
1689 if (match(V, m_LShr(m_SignMask(), m_Value())))
1692 // The remaining tests are all recursive, so bail out if we hit the limit.
1693 if (Depth++ == MaxDepth)
1696 Value *X = nullptr, *Y = nullptr;
1697 // A shift left or a logical shift right of a power of two is a power of two
1699 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1700 match(V, m_LShr(m_Value(X), m_Value()))))
1701 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1703 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1704 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1706 if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1707 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1708 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1710 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1711 // A power of two and'd with anything is a power of two or zero.
1712 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1713 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1715 // X & (-X) is always a power of two or zero.
1716 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1721 // Adding a power-of-two or zero to the same power-of-two or zero yields
1722 // either the original power-of-two, a larger power-of-two or zero.
1723 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1724 const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1725 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1726 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1727 match(X, m_And(m_Value(), m_Specific(Y))))
1728 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1730 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1731 match(Y, m_And(m_Value(), m_Specific(X))))
1732 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1735 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1736 KnownBits LHSBits(BitWidth);
1737 computeKnownBits(X, LHSBits, Depth, Q);
1739 KnownBits RHSBits(BitWidth);
1740 computeKnownBits(Y, RHSBits, Depth, Q);
1741 // If i8 V is a power of two or zero:
1742 // ZeroBits: 1 1 1 0 1 1 1 1
1743 // ~ZeroBits: 0 0 0 1 0 0 0 0
1744 if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1745 // If OrZero isn't set, we cannot give back a zero result.
1746 // Make sure either the LHS or RHS has a bit set.
1747 if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1752 // An exact divide or right shift can only shift off zero bits, so the result
1753 // is a power of two only if the first operand is a power of two and not
1754 // copying a sign bit (sdiv int_min, 2).
1755 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1756 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1757 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1764 /// Test whether a GEP's result is known to be non-null.
1766 /// Uses properties inherent in a GEP to try to determine whether it is known
1769 /// Currently this routine does not support vector GEPs.
1770 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1772 const Function *F = nullptr;
1773 if (const Instruction *I = dyn_cast<Instruction>(GEP))
1774 F = I->getFunction();
1776 if (!GEP->isInBounds() ||
1777 NullPointerIsDefined(F, GEP->getPointerAddressSpace()))
1780 // FIXME: Support vector-GEPs.
1781 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1783 // If the base pointer is non-null, we cannot walk to a null address with an
1784 // inbounds GEP in address space zero.
1785 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1788 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1789 // If so, then the GEP cannot produce a null pointer, as doing so would
1790 // inherently violate the inbounds contract within address space zero.
1791 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1792 GTI != GTE; ++GTI) {
1793 // Struct types are easy -- they must always be indexed by a constant.
1794 if (StructType *STy = GTI.getStructTypeOrNull()) {
1795 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1796 unsigned ElementIdx = OpC->getZExtValue();
1797 const StructLayout *SL = Q.DL.getStructLayout(STy);
1798 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1799 if (ElementOffset > 0)
1804 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1805 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1808 // Fast path the constant operand case both for efficiency and so we don't
1809 // increment Depth when just zipping down an all-constant GEP.
1810 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1816 // We post-increment Depth here because while isKnownNonZero increments it
1817 // as well, when we pop back up that increment won't persist. We don't want
1818 // to recurse 10k times just because we have 10k GEP operands. We don't
1819 // bail completely out because we want to handle constant GEPs regardless
1821 if (Depth++ >= MaxDepth)
1824 if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1831 static bool isKnownNonNullFromDominatingCondition(const Value *V,
1832 const Instruction *CtxI,
1833 const DominatorTree *DT) {
1834 assert(V->getType()->isPointerTy() && "V must be pointer type");
1835 assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
1840 unsigned NumUsesExplored = 0;
1841 for (auto *U : V->users()) {
1842 // Avoid massive lists
1843 if (NumUsesExplored >= DomConditionsMaxUses)
1847 // If the value is used as an argument to a call or invoke, then argument
1848 // attributes may provide an answer about null-ness.
1849 if (auto CS = ImmutableCallSite(U))
1850 if (auto *CalledFunc = CS.getCalledFunction())
1851 for (const Argument &Arg : CalledFunc->args())
1852 if (CS.getArgOperand(Arg.getArgNo()) == V &&
1853 Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
1856 // Consider only compare instructions uniquely controlling a branch
1857 CmpInst::Predicate Pred;
1858 if (!match(const_cast<User *>(U),
1859 m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
1860 (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
1863 for (auto *CmpU : U->users()) {
1864 if (const BranchInst *BI = dyn_cast<BranchInst>(CmpU)) {
1865 assert(BI->isConditional() && "uses a comparison!");
1867 BasicBlock *NonNullSuccessor =
1868 BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
1869 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
1870 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
1872 } else if (Pred == ICmpInst::ICMP_NE &&
1873 match(CmpU, m_Intrinsic<Intrinsic::experimental_guard>()) &&
1874 DT->dominates(cast<Instruction>(CmpU), CtxI)) {
1883 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1884 /// ensure that the value it's attached to is never Value? 'RangeType' is
1885 /// is the type of the value described by the range.
1886 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
1887 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1888 assert(NumRanges >= 1);
1889 for (unsigned i = 0; i < NumRanges; ++i) {
1890 ConstantInt *Lower =
1891 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1892 ConstantInt *Upper =
1893 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1894 ConstantRange Range(Lower->getValue(), Upper->getValue());
1895 if (Range.contains(Value))
1901 /// Return true if the given value is known to be non-zero when defined. For
1902 /// vectors, return true if every element is known to be non-zero when
1903 /// defined. For pointers, if the context instruction and dominator tree are
1904 /// specified, perform context-sensitive analysis and return true if the
1905 /// pointer couldn't possibly be null at the specified instruction.
1906 /// Supports values with integer or pointer type and vectors of integers.
1907 bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
1908 if (auto *C = dyn_cast<Constant>(V)) {
1909 if (C->isNullValue())
1911 if (isa<ConstantInt>(C))
1912 // Must be non-zero due to null test above.
1915 // For constant vectors, check that all elements are undefined or known
1916 // non-zero to determine that the whole vector is known non-zero.
1917 if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1918 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1919 Constant *Elt = C->getAggregateElement(i);
1920 if (!Elt || Elt->isNullValue())
1922 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1928 // A global variable in address space 0 is non null unless extern weak
1929 // or an absolute symbol reference. Other address spaces may have null as a
1930 // valid address for a global, so we can't assume anything.
1931 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
1932 if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
1933 GV->getType()->getAddressSpace() == 0)
1939 if (auto *I = dyn_cast<Instruction>(V)) {
1940 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1941 // If the possible ranges don't contain zero, then the value is
1942 // definitely non-zero.
1943 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1944 const APInt ZeroValue(Ty->getBitWidth(), 0);
1945 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1951 // Some of the tests below are recursive, so bail out if we hit the limit.
1952 if (Depth++ >= MaxDepth)
1955 // Check for pointer simplifications.
1956 if (V->getType()->isPointerTy()) {
1957 // Alloca never returns null, malloc might.
1958 if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
1961 // A byval, inalloca, or nonnull argument is never null.
1962 if (const Argument *A = dyn_cast<Argument>(V))
1963 if (A->hasByValOrInAllocaAttr() || A->hasNonNullAttr())
1966 // A Load tagged with nonnull metadata is never null.
1967 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
1968 if (LI->getMetadata(LLVMContext::MD_nonnull))
1971 if (auto CS = ImmutableCallSite(V)) {
1972 if (CS.isReturnNonNull())
1974 if (const auto *RP = getArgumentAliasingToReturnedPointer(CS))
1975 return isKnownNonZero(RP, Depth, Q);
1980 // Check for recursive pointer simplifications.
1981 if (V->getType()->isPointerTy()) {
1982 if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
1985 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1986 if (isGEPKnownNonNull(GEP, Depth, Q))
1990 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
1992 // X | Y != 0 if X != 0 or Y != 0.
1993 Value *X = nullptr, *Y = nullptr;
1994 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1995 return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
1997 // ext X != 0 if X != 0.
1998 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1999 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
2001 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
2002 // if the lowest bit is shifted off the end.
2003 if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
2004 // shl nuw can't remove any non-zero bits.
2005 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2006 if (BO->hasNoUnsignedWrap())
2007 return isKnownNonZero(X, Depth, Q);
2009 KnownBits Known(BitWidth);
2010 computeKnownBits(X, Known, Depth, Q);
2014 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
2015 // defined if the sign bit is shifted off the end.
2016 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
2017 // shr exact can only shift out zero bits.
2018 const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
2020 return isKnownNonZero(X, Depth, Q);
2022 KnownBits Known = computeKnownBits(X, Depth, Q);
2023 if (Known.isNegative())
2026 // If the shifter operand is a constant, and all of the bits shifted
2027 // out are known to be zero, and X is known non-zero then at least one
2028 // non-zero bit must remain.
2029 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
2030 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
2031 // Is there a known one in the portion not shifted out?
2032 if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
2034 // Are all the bits to be shifted out known zero?
2035 if (Known.countMinTrailingZeros() >= ShiftVal)
2036 return isKnownNonZero(X, Depth, Q);
2039 // div exact can only produce a zero if the dividend is zero.
2040 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
2041 return isKnownNonZero(X, Depth, Q);
2044 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2045 KnownBits XKnown = computeKnownBits(X, Depth, Q);
2046 KnownBits YKnown = computeKnownBits(Y, Depth, Q);
2048 // If X and Y are both non-negative (as signed values) then their sum is not
2049 // zero unless both X and Y are zero.
2050 if (XKnown.isNonNegative() && YKnown.isNonNegative())
2051 if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
2054 // If X and Y are both negative (as signed values) then their sum is not
2055 // zero unless both X and Y equal INT_MIN.
2056 if (XKnown.isNegative() && YKnown.isNegative()) {
2057 APInt Mask = APInt::getSignedMaxValue(BitWidth);
2058 // The sign bit of X is set. If some other bit is set then X is not equal
2060 if (XKnown.One.intersects(Mask))
2062 // The sign bit of Y is set. If some other bit is set then Y is not equal
2064 if (YKnown.One.intersects(Mask))
2068 // The sum of a non-negative number and a power of two is not zero.
2069 if (XKnown.isNonNegative() &&
2070 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
2072 if (YKnown.isNonNegative() &&
2073 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
2077 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2078 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2079 // If X and Y are non-zero then so is X * Y as long as the multiplication
2080 // does not overflow.
2081 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
2082 isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
2085 // (C ? X : Y) != 0 if X != 0 and Y != 0.
2086 else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
2087 if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
2088 isKnownNonZero(SI->getFalseValue(), Depth, Q))
2092 else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
2093 // Try and detect a recurrence that monotonically increases from a
2094 // starting value, as these are common as induction variables.
2095 if (PN->getNumIncomingValues() == 2) {
2096 Value *Start = PN->getIncomingValue(0);
2097 Value *Induction = PN->getIncomingValue(1);
2098 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2099 std::swap(Start, Induction);
2100 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2101 if (!C->isZero() && !C->isNegative()) {
2103 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2104 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2110 // Check if all incoming values are non-zero constant.
2111 bool AllNonZeroConstants = llvm::all_of(PN->operands(), [](Value *V) {
2112 return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZero();
2114 if (AllNonZeroConstants)
2118 KnownBits Known(BitWidth);
2119 computeKnownBits(V, Known, Depth, Q);
2120 return Known.One != 0;
2123 /// Return true if V2 == V1 + X, where X is known non-zero.
2124 static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
2125 const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2126 if (!BO || BO->getOpcode() != Instruction::Add)
2128 Value *Op = nullptr;
2129 if (V2 == BO->getOperand(0))
2130 Op = BO->getOperand(1);
2131 else if (V2 == BO->getOperand(1))
2132 Op = BO->getOperand(0);
2135 return isKnownNonZero(Op, 0, Q);
2138 /// Return true if it is known that V1 != V2.
2139 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
2142 if (V1->getType() != V2->getType())
2143 // We can't look through casts yet.
2145 if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
2148 if (V1->getType()->isIntOrIntVectorTy()) {
2149 // Are any known bits in V1 contradictory to known bits in V2? If V1
2150 // has a known zero where V2 has a known one, they must not be equal.
2151 KnownBits Known1 = computeKnownBits(V1, 0, Q);
2152 KnownBits Known2 = computeKnownBits(V2, 0, Q);
2154 if (Known1.Zero.intersects(Known2.One) ||
2155 Known2.Zero.intersects(Known1.One))
2161 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2162 /// simplify operations downstream. Mask is known to be zero for bits that V
2165 /// This function is defined on values with integer type, values with pointer
2166 /// type, and vectors of integers. In the case
2167 /// where V is a vector, the mask, known zero, and known one values are the
2168 /// same width as the vector element, and the bit is set only if it is true
2169 /// for all of the elements in the vector.
2170 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2172 KnownBits Known(Mask.getBitWidth());
2173 computeKnownBits(V, Known, Depth, Q);
2174 return Mask.isSubsetOf(Known.Zero);
2177 /// For vector constants, loop over the elements and find the constant with the
2178 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2179 /// or if any element was not analyzed; otherwise, return the count for the
2180 /// element with the minimum number of sign bits.
2181 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2183 const auto *CV = dyn_cast<Constant>(V);
2184 if (!CV || !CV->getType()->isVectorTy())
2187 unsigned MinSignBits = TyBits;
2188 unsigned NumElts = CV->getType()->getVectorNumElements();
2189 for (unsigned i = 0; i != NumElts; ++i) {
2190 // If we find a non-ConstantInt, bail out.
2191 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2195 MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
2201 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2204 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2206 unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2207 assert(Result > 0 && "At least one sign bit needs to be present!");
2211 /// Return the number of times the sign bit of the register is replicated into
2212 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2213 /// (itself), but other cases can give us information. For example, immediately
2214 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2215 /// other, so we return 3. For vectors, return the number of sign bits for the
2216 /// vector element with the minimum number of known sign bits.
2217 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2219 assert(Depth <= MaxDepth && "Limit Search Depth");
2221 // We return the minimum number of sign bits that are guaranteed to be present
2222 // in V, so for undef we have to conservatively return 1. We don't have the
2223 // same behavior for poison though -- that's a FIXME today.
2225 Type *ScalarTy = V->getType()->getScalarType();
2226 unsigned TyBits = ScalarTy->isPointerTy() ?
2227 Q.DL.getIndexTypeSizeInBits(ScalarTy) :
2228 Q.DL.getTypeSizeInBits(ScalarTy);
2231 unsigned FirstAnswer = 1;
2233 // Note that ConstantInt is handled by the general computeKnownBits case
2236 if (Depth == MaxDepth)
2237 return 1; // Limit search depth.
2239 const Operator *U = dyn_cast<Operator>(V);
2240 switch (Operator::getOpcode(V)) {
2242 case Instruction::SExt:
2243 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2244 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2246 case Instruction::SDiv: {
2247 const APInt *Denominator;
2248 // sdiv X, C -> adds log(C) sign bits.
2249 if (match(U->getOperand(1), m_APInt(Denominator))) {
2251 // Ignore non-positive denominator.
2252 if (!Denominator->isStrictlyPositive())
2255 // Calculate the incoming numerator bits.
2256 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2258 // Add floor(log(C)) bits to the numerator bits.
2259 return std::min(TyBits, NumBits + Denominator->logBase2());
2264 case Instruction::SRem: {
2265 const APInt *Denominator;
2266 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2267 // positive constant. This let us put a lower bound on the number of sign
2269 if (match(U->getOperand(1), m_APInt(Denominator))) {
2271 // Ignore non-positive denominator.
2272 if (!Denominator->isStrictlyPositive())
2275 // Calculate the incoming numerator bits. SRem by a positive constant
2276 // can't lower the number of sign bits.
2278 ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2280 // Calculate the leading sign bit constraints by examining the
2281 // denominator. Given that the denominator is positive, there are two
2284 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2285 // (1 << ceilLogBase2(C)).
2287 // 2. the numerator is negative. Then the result range is (-C,0] and
2288 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2290 // Thus a lower bound on the number of sign bits is `TyBits -
2291 // ceilLogBase2(C)`.
2293 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2294 return std::max(NumrBits, ResBits);
2299 case Instruction::AShr: {
2300 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2301 // ashr X, C -> adds C sign bits. Vectors too.
2303 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2304 if (ShAmt->uge(TyBits))
2305 break; // Bad shift.
2306 unsigned ShAmtLimited = ShAmt->getZExtValue();
2307 Tmp += ShAmtLimited;
2308 if (Tmp > TyBits) Tmp = TyBits;
2312 case Instruction::Shl: {
2314 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2315 // shl destroys sign bits.
2316 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2317 if (ShAmt->uge(TyBits) || // Bad shift.
2318 ShAmt->uge(Tmp)) break; // Shifted all sign bits out.
2319 Tmp2 = ShAmt->getZExtValue();
2324 case Instruction::And:
2325 case Instruction::Or:
2326 case Instruction::Xor: // NOT is handled here.
2327 // Logical binary ops preserve the number of sign bits at the worst.
2328 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2330 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2331 FirstAnswer = std::min(Tmp, Tmp2);
2332 // We computed what we know about the sign bits as our first
2333 // answer. Now proceed to the generic code that uses
2334 // computeKnownBits, and pick whichever answer is better.
2338 case Instruction::Select:
2339 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2340 if (Tmp == 1) break;
2341 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2342 return std::min(Tmp, Tmp2);
2344 case Instruction::Add:
2345 // Add can have at most one carry bit. Thus we know that the output
2346 // is, at worst, one more bit than the inputs.
2347 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2348 if (Tmp == 1) break;
2350 // Special case decrementing a value (ADD X, -1):
2351 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2352 if (CRHS->isAllOnesValue()) {
2353 KnownBits Known(TyBits);
2354 computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2356 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2358 if ((Known.Zero | 1).isAllOnesValue())
2361 // If we are subtracting one from a positive number, there is no carry
2362 // out of the result.
2363 if (Known.isNonNegative())
2367 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2368 if (Tmp2 == 1) break;
2369 return std::min(Tmp, Tmp2)-1;
2371 case Instruction::Sub:
2372 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2373 if (Tmp2 == 1) break;
2376 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2377 if (CLHS->isNullValue()) {
2378 KnownBits Known(TyBits);
2379 computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2380 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2382 if ((Known.Zero | 1).isAllOnesValue())
2385 // If the input is known to be positive (the sign bit is known clear),
2386 // the output of the NEG has the same number of sign bits as the input.
2387 if (Known.isNonNegative())
2390 // Otherwise, we treat this like a SUB.
2393 // Sub can have at most one carry bit. Thus we know that the output
2394 // is, at worst, one more bit than the inputs.
2395 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2396 if (Tmp == 1) break;
2397 return std::min(Tmp, Tmp2)-1;
2399 case Instruction::Mul: {
2400 // The output of the Mul can be at most twice the valid bits in the inputs.
2401 unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2402 if (SignBitsOp0 == 1) break;
2403 unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2404 if (SignBitsOp1 == 1) break;
2405 unsigned OutValidBits =
2406 (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
2407 return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
2410 case Instruction::PHI: {
2411 const PHINode *PN = cast<PHINode>(U);
2412 unsigned NumIncomingValues = PN->getNumIncomingValues();
2413 // Don't analyze large in-degree PHIs.
2414 if (NumIncomingValues > 4) break;
2415 // Unreachable blocks may have zero-operand PHI nodes.
2416 if (NumIncomingValues == 0) break;
2418 // Take the minimum of all incoming values. This can't infinitely loop
2419 // because of our depth threshold.
2420 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2421 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2422 if (Tmp == 1) return Tmp;
2424 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2429 case Instruction::Trunc:
2430 // FIXME: it's tricky to do anything useful for this, but it is an important
2431 // case for targets like X86.
2434 case Instruction::ExtractElement:
2435 // Look through extract element. At the moment we keep this simple and skip
2436 // tracking the specific element. But at least we might find information
2437 // valid for all elements of the vector (for example if vector is sign
2438 // extended, shifted, etc).
2439 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2442 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2443 // use this information.
2445 // If we can examine all elements of a vector constant successfully, we're
2446 // done (we can't do any better than that). If not, keep trying.
2447 if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2450 KnownBits Known(TyBits);
2451 computeKnownBits(V, Known, Depth, Q);
2453 // If we know that the sign bit is either zero or one, determine the number of
2454 // identical bits in the top of the input value.
2455 return std::max(FirstAnswer, Known.countMinSignBits());
2458 /// This function computes the integer multiple of Base that equals V.
2459 /// If successful, it returns true and returns the multiple in
2460 /// Multiple. If unsuccessful, it returns false. It looks
2461 /// through SExt instructions only if LookThroughSExt is true.
2462 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2463 bool LookThroughSExt, unsigned Depth) {
2464 const unsigned MaxDepth = 6;
2466 assert(V && "No Value?");
2467 assert(Depth <= MaxDepth && "Limit Search Depth");
2468 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2470 Type *T = V->getType();
2472 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2482 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2483 Constant *BaseVal = ConstantInt::get(T, Base);
2484 if (CO && CO == BaseVal) {
2486 Multiple = ConstantInt::get(T, 1);
2490 if (CI && CI->getZExtValue() % Base == 0) {
2491 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2495 if (Depth == MaxDepth) return false; // Limit search depth.
2497 Operator *I = dyn_cast<Operator>(V);
2498 if (!I) return false;
2500 switch (I->getOpcode()) {
2502 case Instruction::SExt:
2503 if (!LookThroughSExt) return false;
2504 // otherwise fall through to ZExt
2506 case Instruction::ZExt:
2507 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2508 LookThroughSExt, Depth+1);
2509 case Instruction::Shl:
2510 case Instruction::Mul: {
2511 Value *Op0 = I->getOperand(0);
2512 Value *Op1 = I->getOperand(1);
2514 if (I->getOpcode() == Instruction::Shl) {
2515 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2516 if (!Op1CI) return false;
2517 // Turn Op0 << Op1 into Op0 * 2^Op1
2518 APInt Op1Int = Op1CI->getValue();
2519 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2520 APInt API(Op1Int.getBitWidth(), 0);
2521 API.setBit(BitToSet);
2522 Op1 = ConstantInt::get(V->getContext(), API);
2525 Value *Mul0 = nullptr;
2526 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2527 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2528 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2529 if (Op1C->getType()->getPrimitiveSizeInBits() <
2530 MulC->getType()->getPrimitiveSizeInBits())
2531 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2532 if (Op1C->getType()->getPrimitiveSizeInBits() >
2533 MulC->getType()->getPrimitiveSizeInBits())
2534 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2536 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2537 Multiple = ConstantExpr::getMul(MulC, Op1C);
2541 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2542 if (Mul0CI->getValue() == 1) {
2543 // V == Base * Op1, so return Op1
2549 Value *Mul1 = nullptr;
2550 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2551 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2552 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2553 if (Op0C->getType()->getPrimitiveSizeInBits() <
2554 MulC->getType()->getPrimitiveSizeInBits())
2555 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2556 if (Op0C->getType()->getPrimitiveSizeInBits() >
2557 MulC->getType()->getPrimitiveSizeInBits())
2558 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2560 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2561 Multiple = ConstantExpr::getMul(MulC, Op0C);
2565 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2566 if (Mul1CI->getValue() == 1) {
2567 // V == Base * Op0, so return Op0
2575 // We could not determine if V is a multiple of Base.
2579 Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS,
2580 const TargetLibraryInfo *TLI) {
2581 const Function *F = ICS.getCalledFunction();
2583 return Intrinsic::not_intrinsic;
2585 if (F->isIntrinsic())
2586 return F->getIntrinsicID();
2589 return Intrinsic::not_intrinsic;
2592 // We're going to make assumptions on the semantics of the functions, check
2593 // that the target knows that it's available in this environment and it does
2594 // not have local linkage.
2595 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2596 return Intrinsic::not_intrinsic;
2598 if (!ICS.onlyReadsMemory())
2599 return Intrinsic::not_intrinsic;
2601 // Otherwise check if we have a call to a function that can be turned into a
2602 // vector intrinsic.
2609 return Intrinsic::sin;
2613 return Intrinsic::cos;
2617 return Intrinsic::exp;
2621 return Intrinsic::exp2;
2625 return Intrinsic::log;
2627 case LibFunc_log10f:
2628 case LibFunc_log10l:
2629 return Intrinsic::log10;
2633 return Intrinsic::log2;
2637 return Intrinsic::fabs;
2641 return Intrinsic::minnum;
2645 return Intrinsic::maxnum;
2646 case LibFunc_copysign:
2647 case LibFunc_copysignf:
2648 case LibFunc_copysignl:
2649 return Intrinsic::copysign;
2651 case LibFunc_floorf:
2652 case LibFunc_floorl:
2653 return Intrinsic::floor;
2657 return Intrinsic::ceil;
2659 case LibFunc_truncf:
2660 case LibFunc_truncl:
2661 return Intrinsic::trunc;
2665 return Intrinsic::rint;
2666 case LibFunc_nearbyint:
2667 case LibFunc_nearbyintf:
2668 case LibFunc_nearbyintl:
2669 return Intrinsic::nearbyint;
2671 case LibFunc_roundf:
2672 case LibFunc_roundl:
2673 return Intrinsic::round;
2677 return Intrinsic::pow;
2681 return Intrinsic::sqrt;
2684 return Intrinsic::not_intrinsic;
2687 /// Return true if we can prove that the specified FP value is never equal to
2690 /// NOTE: this function will need to be revisited when we support non-default
2692 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
2694 if (auto *CFP = dyn_cast<ConstantFP>(V))
2695 return !CFP->getValueAPF().isNegZero();
2697 // Limit search depth.
2698 if (Depth == MaxDepth)
2701 auto *Op = dyn_cast<Operator>(V);
2705 // Check if the nsz fast-math flag is set.
2706 if (auto *FPO = dyn_cast<FPMathOperator>(Op))
2707 if (FPO->hasNoSignedZeros())
2710 // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
2711 if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
2714 // sitofp and uitofp turn into +0.0 for zero.
2715 if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
2718 if (auto *Call = dyn_cast<CallInst>(Op)) {
2719 Intrinsic::ID IID = getIntrinsicForCallSite(Call, TLI);
2723 // sqrt(-0.0) = -0.0, no other negative results are possible.
2724 case Intrinsic::sqrt:
2725 return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
2727 case Intrinsic::fabs:
2735 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
2736 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
2737 /// bit despite comparing equal.
2738 static bool cannotBeOrderedLessThanZeroImpl(const Value *V,
2739 const TargetLibraryInfo *TLI,
2742 // TODO: This function does not do the right thing when SignBitOnly is true
2743 // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
2744 // which flips the sign bits of NaNs. See
2745 // https://llvm.org/bugs/show_bug.cgi?id=31702.
2747 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2748 return !CFP->getValueAPF().isNegative() ||
2749 (!SignBitOnly && CFP->getValueAPF().isZero());
2752 // Handle vector of constants.
2753 if (auto *CV = dyn_cast<Constant>(V)) {
2754 if (CV->getType()->isVectorTy()) {
2755 unsigned NumElts = CV->getType()->getVectorNumElements();
2756 for (unsigned i = 0; i != NumElts; ++i) {
2757 auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
2760 if (CFP->getValueAPF().isNegative() &&
2761 (SignBitOnly || !CFP->getValueAPF().isZero()))
2765 // All non-negative ConstantFPs.
2770 if (Depth == MaxDepth)
2771 return false; // Limit search depth.
2773 const Operator *I = dyn_cast<Operator>(V);
2777 switch (I->getOpcode()) {
2780 // Unsigned integers are always nonnegative.
2781 case Instruction::UIToFP:
2783 case Instruction::FMul:
2784 // x*x is always non-negative or a NaN.
2785 if (I->getOperand(0) == I->getOperand(1) &&
2786 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
2790 case Instruction::FAdd:
2791 case Instruction::FDiv:
2792 case Instruction::FRem:
2793 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2795 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2797 case Instruction::Select:
2798 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2800 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2802 case Instruction::FPExt:
2803 case Instruction::FPTrunc:
2804 // Widening/narrowing never change sign.
2805 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2807 case Instruction::ExtractElement:
2808 // Look through extract element. At the moment we keep this simple and skip
2809 // tracking the specific element. But at least we might find information
2810 // valid for all elements of the vector.
2811 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2813 case Instruction::Call:
2814 const auto *CI = cast<CallInst>(I);
2815 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2819 case Intrinsic::maxnum:
2820 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2822 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2824 case Intrinsic::minnum:
2825 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2827 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2829 case Intrinsic::exp:
2830 case Intrinsic::exp2:
2831 case Intrinsic::fabs:
2834 case Intrinsic::sqrt:
2835 // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
2838 return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
2839 CannotBeNegativeZero(CI->getOperand(0), TLI));
2841 case Intrinsic::powi:
2842 if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
2843 // powi(x,n) is non-negative if n is even.
2844 if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
2847 // TODO: This is not correct. Given that exp is an integer, here are the
2848 // ways that pow can return a negative value:
2850 // pow(x, exp) --> negative if exp is odd and x is negative.
2851 // pow(-0, exp) --> -inf if exp is negative odd.
2852 // pow(-0, exp) --> -0 if exp is positive odd.
2853 // pow(-inf, exp) --> -0 if exp is negative odd.
2854 // pow(-inf, exp) --> -inf if exp is positive odd.
2856 // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
2857 // but we must return false if x == -0. Unfortunately we do not currently
2858 // have a way of expressing this constraint. See details in
2859 // https://llvm.org/bugs/show_bug.cgi?id=31702.
2860 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2863 case Intrinsic::fma:
2864 case Intrinsic::fmuladd:
2865 // x*x+y is non-negative if y is non-negative.
2866 return I->getOperand(0) == I->getOperand(1) &&
2867 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
2868 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2876 bool llvm::CannotBeOrderedLessThanZero(const Value *V,
2877 const TargetLibraryInfo *TLI) {
2878 return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
2881 bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) {
2882 return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
2885 bool llvm::isKnownNeverNaN(const Value *V) {
2886 assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
2888 // If we're told that NaNs won't happen, assume they won't.
2889 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
2890 if (FPMathOp->hasNoNaNs())
2893 // TODO: Handle instructions and potentially recurse like other 'isKnown'
2894 // functions. For example, the result of sitofp is never NaN.
2896 // Handle scalar constants.
2897 if (auto *CFP = dyn_cast<ConstantFP>(V))
2898 return !CFP->isNaN();
2900 // Bail out for constant expressions, but try to handle vector constants.
2901 if (!V->getType()->isVectorTy() || !isa<Constant>(V))
2904 // For vectors, verify that each element is not NaN.
2905 unsigned NumElts = V->getType()->getVectorNumElements();
2906 for (unsigned i = 0; i != NumElts; ++i) {
2907 Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
2910 if (isa<UndefValue>(Elt))
2912 auto *CElt = dyn_cast<ConstantFP>(Elt);
2913 if (!CElt || CElt->isNaN())
2916 // All elements were confirmed not-NaN or undefined.
2920 /// If the specified value can be set by repeating the same byte in memory,
2921 /// return the i8 value that it is represented with. This is
2922 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2923 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2924 /// byte store (e.g. i16 0x1234), return null.
2925 Value *llvm::isBytewiseValue(Value *V) {
2926 // All byte-wide stores are splatable, even of arbitrary variables.
2927 if (V->getType()->isIntegerTy(8)) return V;
2929 // Handle 'null' ConstantArrayZero etc.
2930 if (Constant *C = dyn_cast<Constant>(V))
2931 if (C->isNullValue())
2932 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2934 // Constant float and double values can be handled as integer values if the
2935 // corresponding integer value is "byteable". An important case is 0.0.
2936 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2937 if (CFP->getType()->isFloatTy())
2938 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2939 if (CFP->getType()->isDoubleTy())
2940 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2941 // Don't handle long double formats, which have strange constraints.
2944 // We can handle constant integers that are multiple of 8 bits.
2945 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2946 if (CI->getBitWidth() % 8 == 0) {
2947 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2949 if (!CI->getValue().isSplat(8))
2951 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2955 // A ConstantDataArray/Vector is splatable if all its members are equal and
2957 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2958 Value *Elt = CA->getElementAsConstant(0);
2959 Value *Val = isBytewiseValue(Elt);
2963 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2964 if (CA->getElementAsConstant(I) != Elt)
2970 // Conceptually, we could handle things like:
2971 // %a = zext i8 %X to i16
2972 // %b = shl i16 %a, 8
2973 // %c = or i16 %a, %b
2974 // but until there is an example that actually needs this, it doesn't seem
2975 // worth worrying about.
2979 // This is the recursive version of BuildSubAggregate. It takes a few different
2980 // arguments. Idxs is the index within the nested struct From that we are
2981 // looking at now (which is of type IndexedType). IdxSkip is the number of
2982 // indices from Idxs that should be left out when inserting into the resulting
2983 // struct. To is the result struct built so far, new insertvalue instructions
2985 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2986 SmallVectorImpl<unsigned> &Idxs,
2988 Instruction *InsertBefore) {
2989 StructType *STy = dyn_cast<StructType>(IndexedType);
2991 // Save the original To argument so we can modify it
2993 // General case, the type indexed by Idxs is a struct
2994 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2995 // Process each struct element recursively
2998 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
3002 // Couldn't find any inserted value for this index? Cleanup
3003 while (PrevTo != OrigTo) {
3004 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
3005 PrevTo = Del->getAggregateOperand();
3006 Del->eraseFromParent();
3008 // Stop processing elements
3012 // If we successfully found a value for each of our subaggregates
3016 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
3017 // the struct's elements had a value that was inserted directly. In the latter
3018 // case, perhaps we can't determine each of the subelements individually, but
3019 // we might be able to find the complete struct somewhere.
3021 // Find the value that is at that particular spot
3022 Value *V = FindInsertedValue(From, Idxs);
3027 // Insert the value in the new (sub) aggregate
3028 return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
3029 "tmp", InsertBefore);
3032 // This helper takes a nested struct and extracts a part of it (which is again a
3033 // struct) into a new value. For example, given the struct:
3034 // { a, { b, { c, d }, e } }
3035 // and the indices "1, 1" this returns
3038 // It does this by inserting an insertvalue for each element in the resulting
3039 // struct, as opposed to just inserting a single struct. This will only work if
3040 // each of the elements of the substruct are known (ie, inserted into From by an
3041 // insertvalue instruction somewhere).
3043 // All inserted insertvalue instructions are inserted before InsertBefore
3044 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
3045 Instruction *InsertBefore) {
3046 assert(InsertBefore && "Must have someplace to insert!");
3047 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
3049 Value *To = UndefValue::get(IndexedType);
3050 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
3051 unsigned IdxSkip = Idxs.size();
3053 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
3056 /// Given an aggregate and a sequence of indices, see if the scalar value
3057 /// indexed is already around as a register, for example if it was inserted
3058 /// directly into the aggregate.
3060 /// If InsertBefore is not null, this function will duplicate (modified)
3061 /// insertvalues when a part of a nested struct is extracted.
3062 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
3063 Instruction *InsertBefore) {
3064 // Nothing to index? Just return V then (this is useful at the end of our
3066 if (idx_range.empty())
3068 // We have indices, so V should have an indexable type.
3069 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
3070 "Not looking at a struct or array?");
3071 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
3072 "Invalid indices for type?");
3074 if (Constant *C = dyn_cast<Constant>(V)) {
3075 C = C->getAggregateElement(idx_range[0]);
3076 if (!C) return nullptr;
3077 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
3080 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
3081 // Loop the indices for the insertvalue instruction in parallel with the
3082 // requested indices
3083 const unsigned *req_idx = idx_range.begin();
3084 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
3085 i != e; ++i, ++req_idx) {
3086 if (req_idx == idx_range.end()) {
3087 // We can't handle this without inserting insertvalues
3091 // The requested index identifies a part of a nested aggregate. Handle
3092 // this specially. For example,
3093 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
3094 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
3095 // %C = extractvalue {i32, { i32, i32 } } %B, 1
3096 // This can be changed into
3097 // %A = insertvalue {i32, i32 } undef, i32 10, 0
3098 // %C = insertvalue {i32, i32 } %A, i32 11, 1
3099 // which allows the unused 0,0 element from the nested struct to be
3101 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
3105 // This insert value inserts something else than what we are looking for.
3106 // See if the (aggregate) value inserted into has the value we are
3107 // looking for, then.
3109 return FindInsertedValue(I->getAggregateOperand(), idx_range,
3112 // If we end up here, the indices of the insertvalue match with those
3113 // requested (though possibly only partially). Now we recursively look at
3114 // the inserted value, passing any remaining indices.
3115 return FindInsertedValue(I->getInsertedValueOperand(),
3116 makeArrayRef(req_idx, idx_range.end()),
3120 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
3121 // If we're extracting a value from an aggregate that was extracted from
3122 // something else, we can extract from that something else directly instead.
3123 // However, we will need to chain I's indices with the requested indices.
3125 // Calculate the number of indices required
3126 unsigned size = I->getNumIndices() + idx_range.size();
3127 // Allocate some space to put the new indices in
3128 SmallVector<unsigned, 5> Idxs;
3130 // Add indices from the extract value instruction
3131 Idxs.append(I->idx_begin(), I->idx_end());
3133 // Add requested indices
3134 Idxs.append(idx_range.begin(), idx_range.end());
3136 assert(Idxs.size() == size
3137 && "Number of indices added not correct?");
3139 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
3141 // Otherwise, we don't know (such as, extracting from a function return value
3142 // or load instruction)
3146 /// Analyze the specified pointer to see if it can be expressed as a base
3147 /// pointer plus a constant offset. Return the base and offset to the caller.
3148 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
3149 const DataLayout &DL) {
3150 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ptr->getType());
3151 APInt ByteOffset(BitWidth, 0);
3153 // We walk up the defs but use a visited set to handle unreachable code. In
3154 // that case, we stop after accumulating the cycle once (not that it
3156 SmallPtrSet<Value *, 16> Visited;
3157 while (Visited.insert(Ptr).second) {
3158 if (Ptr->getType()->isVectorTy())
3161 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
3162 // If one of the values we have visited is an addrspacecast, then
3163 // the pointer type of this GEP may be different from the type
3164 // of the Ptr parameter which was passed to this function. This
3165 // means when we construct GEPOffset, we need to use the size
3166 // of GEP's pointer type rather than the size of the original
3168 APInt GEPOffset(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
3169 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
3172 ByteOffset += GEPOffset.getSExtValue();
3174 Ptr = GEP->getPointerOperand();
3175 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
3176 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
3177 Ptr = cast<Operator>(Ptr)->getOperand(0);
3178 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
3179 if (GA->isInterposable())
3181 Ptr = GA->getAliasee();
3186 Offset = ByteOffset.getSExtValue();
3190 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP,
3191 unsigned CharSize) {
3192 // Make sure the GEP has exactly three arguments.
3193 if (GEP->getNumOperands() != 3)
3196 // Make sure the index-ee is a pointer to array of \p CharSize integers.
3198 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
3199 if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
3202 // Check to make sure that the first operand of the GEP is an integer and
3203 // has value 0 so that we are sure we're indexing into the initializer.
3204 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
3205 if (!FirstIdx || !FirstIdx->isZero())
3211 bool llvm::getConstantDataArrayInfo(const Value *V,
3212 ConstantDataArraySlice &Slice,
3213 unsigned ElementSize, uint64_t Offset) {
3216 // Look through bitcast instructions and geps.
3217 V = V->stripPointerCasts();
3219 // If the value is a GEP instruction or constant expression, treat it as an
3221 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3222 // The GEP operator should be based on a pointer to string constant, and is
3223 // indexing into the string constant.
3224 if (!isGEPBasedOnPointerToString(GEP, ElementSize))
3227 // If the second index isn't a ConstantInt, then this is a variable index
3228 // into the array. If this occurs, we can't say anything meaningful about
3230 uint64_t StartIdx = 0;
3231 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3232 StartIdx = CI->getZExtValue();
3235 return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
3239 // The GEP instruction, constant or instruction, must reference a global
3240 // variable that is a constant and is initialized. The referenced constant
3241 // initializer is the array that we'll use for optimization.
3242 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3243 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3246 const ConstantDataArray *Array;
3248 if (GV->getInitializer()->isNullValue()) {
3249 Type *GVTy = GV->getValueType();
3250 if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
3251 // A zeroinitializer for the array; there is no ConstantDataArray.
3254 const DataLayout &DL = GV->getParent()->getDataLayout();
3255 uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy);
3256 uint64_t Length = SizeInBytes / (ElementSize / 8);
3257 if (Length <= Offset)
3260 Slice.Array = nullptr;
3262 Slice.Length = Length - Offset;
3266 // This must be a ConstantDataArray.
3267 Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3270 ArrayTy = Array->getType();
3272 if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
3275 uint64_t NumElts = ArrayTy->getArrayNumElements();
3276 if (Offset > NumElts)
3279 Slice.Array = Array;
3280 Slice.Offset = Offset;
3281 Slice.Length = NumElts - Offset;
3285 /// This function computes the length of a null-terminated C string pointed to
3286 /// by V. If successful, it returns true and returns the string in Str.
3287 /// If unsuccessful, it returns false.
3288 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
3289 uint64_t Offset, bool TrimAtNul) {
3290 ConstantDataArraySlice Slice;
3291 if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
3294 if (Slice.Array == nullptr) {
3299 if (Slice.Length == 1) {
3300 Str = StringRef("", 1);
3303 // We cannot instantiate a StringRef as we do not have an appropriate string
3308 // Start out with the entire array in the StringRef.
3309 Str = Slice.Array->getAsString();
3310 // Skip over 'offset' bytes.
3311 Str = Str.substr(Slice.Offset);
3314 // Trim off the \0 and anything after it. If the array is not nul
3315 // terminated, we just return the whole end of string. The client may know
3316 // some other way that the string is length-bound.
3317 Str = Str.substr(0, Str.find('\0'));
3322 // These next two are very similar to the above, but also look through PHI
3324 // TODO: See if we can integrate these two together.
3326 /// If we can compute the length of the string pointed to by
3327 /// the specified pointer, return 'len+1'. If we can't, return 0.
3328 static uint64_t GetStringLengthH(const Value *V,
3329 SmallPtrSetImpl<const PHINode*> &PHIs,
3330 unsigned CharSize) {
3331 // Look through noop bitcast instructions.
3332 V = V->stripPointerCasts();
3334 // If this is a PHI node, there are two cases: either we have already seen it
3336 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3337 if (!PHIs.insert(PN).second)
3338 return ~0ULL; // already in the set.
3340 // If it was new, see if all the input strings are the same length.
3341 uint64_t LenSoFar = ~0ULL;
3342 for (Value *IncValue : PN->incoming_values()) {
3343 uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
3344 if (Len == 0) return 0; // Unknown length -> unknown.
3346 if (Len == ~0ULL) continue;
3348 if (Len != LenSoFar && LenSoFar != ~0ULL)
3349 return 0; // Disagree -> unknown.
3353 // Success, all agree.
3357 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3358 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3359 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
3360 if (Len1 == 0) return 0;
3361 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
3362 if (Len2 == 0) return 0;
3363 if (Len1 == ~0ULL) return Len2;
3364 if (Len2 == ~0ULL) return Len1;
3365 if (Len1 != Len2) return 0;
3369 // Otherwise, see if we can read the string.
3370 ConstantDataArraySlice Slice;
3371 if (!getConstantDataArrayInfo(V, Slice, CharSize))
3374 if (Slice.Array == nullptr)
3377 // Search for nul characters
3378 unsigned NullIndex = 0;
3379 for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
3380 if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
3384 return NullIndex + 1;
3387 /// If we can compute the length of the string pointed to by
3388 /// the specified pointer, return 'len+1'. If we can't, return 0.
3389 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
3390 if (!V->getType()->isPointerTy())
3393 SmallPtrSet<const PHINode*, 32> PHIs;
3394 uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
3395 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3396 // an empty string as a length.
3397 return Len == ~0ULL ? 1 : Len;
3400 const Value *llvm::getArgumentAliasingToReturnedPointer(ImmutableCallSite CS) {
3402 "getArgumentAliasingToReturnedPointer only works on nonnull CallSite");
3403 if (const Value *RV = CS.getReturnedArgOperand())
3405 // This can be used only as a aliasing property.
3406 if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(CS))
3407 return CS.getArgOperand(0);
3411 bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(
3412 ImmutableCallSite CS) {
3413 return CS.getIntrinsicID() == Intrinsic::launder_invariant_group ||
3414 CS.getIntrinsicID() == Intrinsic::strip_invariant_group;
3417 /// \p PN defines a loop-variant pointer to an object. Check if the
3418 /// previous iteration of the loop was referring to the same object as \p PN.
3419 static bool isSameUnderlyingObjectInLoop(const PHINode *PN,
3420 const LoopInfo *LI) {
3421 // Find the loop-defined value.
3422 Loop *L = LI->getLoopFor(PN->getParent());
3423 if (PN->getNumIncomingValues() != 2)
3426 // Find the value from previous iteration.
3427 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3428 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3429 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3430 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3433 // If a new pointer is loaded in the loop, the pointer references a different
3434 // object in every iteration. E.g.:
3438 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3439 if (!L->isLoopInvariant(Load->getPointerOperand()))
3444 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3445 unsigned MaxLookup) {
3446 if (!V->getType()->isPointerTy())
3448 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3449 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3450 V = GEP->getPointerOperand();
3451 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3452 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3453 V = cast<Operator>(V)->getOperand(0);
3454 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3455 if (GA->isInterposable())
3457 V = GA->getAliasee();
3458 } else if (isa<AllocaInst>(V)) {
3459 // An alloca can't be further simplified.
3462 if (auto CS = CallSite(V)) {
3463 // CaptureTracking can know about special capturing properties of some
3464 // intrinsics like launder.invariant.group, that can't be expressed with
3465 // the attributes, but have properties like returning aliasing pointer.
3466 // Because some analysis may assume that nocaptured pointer is not
3467 // returned from some special intrinsic (because function would have to
3468 // be marked with returns attribute), it is crucial to use this function
3469 // because it should be in sync with CaptureTracking. Not using it may
3470 // cause weird miscompilations where 2 aliasing pointers are assumed to
3472 if (auto *RP = getArgumentAliasingToReturnedPointer(CS)) {
3478 // See if InstructionSimplify knows any relevant tricks.
3479 if (Instruction *I = dyn_cast<Instruction>(V))
3480 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3481 if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
3488 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3493 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3494 const DataLayout &DL, LoopInfo *LI,
3495 unsigned MaxLookup) {
3496 SmallPtrSet<Value *, 4> Visited;
3497 SmallVector<Value *, 4> Worklist;
3498 Worklist.push_back(V);
3500 Value *P = Worklist.pop_back_val();
3501 P = GetUnderlyingObject(P, DL, MaxLookup);
3503 if (!Visited.insert(P).second)
3506 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3507 Worklist.push_back(SI->getTrueValue());
3508 Worklist.push_back(SI->getFalseValue());
3512 if (PHINode *PN = dyn_cast<PHINode>(P)) {
3513 // If this PHI changes the underlying object in every iteration of the
3514 // loop, don't look through it. Consider:
3517 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3521 // Prev is tracking Curr one iteration behind so they refer to different
3522 // underlying objects.
3523 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3524 isSameUnderlyingObjectInLoop(PN, LI))
3525 for (Value *IncValue : PN->incoming_values())
3526 Worklist.push_back(IncValue);
3530 Objects.push_back(P);
3531 } while (!Worklist.empty());
3534 /// This is the function that does the work of looking through basic
3535 /// ptrtoint+arithmetic+inttoptr sequences.
3536 static const Value *getUnderlyingObjectFromInt(const Value *V) {
3538 if (const Operator *U = dyn_cast<Operator>(V)) {
3539 // If we find a ptrtoint, we can transfer control back to the
3540 // regular getUnderlyingObjectFromInt.
3541 if (U->getOpcode() == Instruction::PtrToInt)
3542 return U->getOperand(0);
3543 // If we find an add of a constant, a multiplied value, or a phi, it's
3544 // likely that the other operand will lead us to the base
3545 // object. We don't have to worry about the case where the
3546 // object address is somehow being computed by the multiply,
3547 // because our callers only care when the result is an
3548 // identifiable object.
3549 if (U->getOpcode() != Instruction::Add ||
3550 (!isa<ConstantInt>(U->getOperand(1)) &&
3551 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
3552 !isa<PHINode>(U->getOperand(1))))
3554 V = U->getOperand(0);
3558 assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
3562 /// This is a wrapper around GetUnderlyingObjects and adds support for basic
3563 /// ptrtoint+arithmetic+inttoptr sequences.
3564 /// It returns false if unidentified object is found in GetUnderlyingObjects.
3565 bool llvm::getUnderlyingObjectsForCodeGen(const Value *V,
3566 SmallVectorImpl<Value *> &Objects,
3567 const DataLayout &DL) {
3568 SmallPtrSet<const Value *, 16> Visited;
3569 SmallVector<const Value *, 4> Working(1, V);
3571 V = Working.pop_back_val();
3573 SmallVector<Value *, 4> Objs;
3574 GetUnderlyingObjects(const_cast<Value *>(V), Objs, DL);
3576 for (Value *V : Objs) {
3577 if (!Visited.insert(V).second)
3579 if (Operator::getOpcode(V) == Instruction::IntToPtr) {
3581 getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
3582 if (O->getType()->isPointerTy()) {
3583 Working.push_back(O);
3587 // If GetUnderlyingObjects fails to find an identifiable object,
3588 // getUnderlyingObjectsForCodeGen also fails for safety.
3589 if (!isIdentifiedObject(V)) {
3593 Objects.push_back(const_cast<Value *>(V));
3595 } while (!Working.empty());
3599 /// Return true if the only users of this pointer are lifetime markers.
3600 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3601 for (const User *U : V->users()) {
3602 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3603 if (!II) return false;
3605 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3606 II->getIntrinsicID() != Intrinsic::lifetime_end)
3612 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3613 const Instruction *CtxI,
3614 const DominatorTree *DT) {
3615 const Operator *Inst = dyn_cast<Operator>(V);
3619 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3620 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3624 switch (Inst->getOpcode()) {
3627 case Instruction::UDiv:
3628 case Instruction::URem: {
3629 // x / y is undefined if y == 0.
3631 if (match(Inst->getOperand(1), m_APInt(V)))
3635 case Instruction::SDiv:
3636 case Instruction::SRem: {
3637 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3638 const APInt *Numerator, *Denominator;
3639 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3641 // We cannot hoist this division if the denominator is 0.
3642 if (*Denominator == 0)
3644 // It's safe to hoist if the denominator is not 0 or -1.
3645 if (*Denominator != -1)
3647 // At this point we know that the denominator is -1. It is safe to hoist as
3648 // long we know that the numerator is not INT_MIN.
3649 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3650 return !Numerator->isMinSignedValue();
3651 // The numerator *might* be MinSignedValue.
3654 case Instruction::Load: {
3655 const LoadInst *LI = cast<LoadInst>(Inst);
3656 if (!LI->isUnordered() ||
3657 // Speculative load may create a race that did not exist in the source.
3658 LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3659 // Speculative load may load data from dirty regions.
3660 LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
3661 LI->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
3663 const DataLayout &DL = LI->getModule()->getDataLayout();
3664 return isDereferenceableAndAlignedPointer(LI->getPointerOperand(),
3665 LI->getAlignment(), DL, CtxI, DT);
3667 case Instruction::Call: {
3668 auto *CI = cast<const CallInst>(Inst);
3669 const Function *Callee = CI->getCalledFunction();
3671 // The called function could have undefined behavior or side-effects, even
3672 // if marked readnone nounwind.
3673 return Callee && Callee->isSpeculatable();
3675 case Instruction::VAArg:
3676 case Instruction::Alloca:
3677 case Instruction::Invoke:
3678 case Instruction::PHI:
3679 case Instruction::Store:
3680 case Instruction::Ret:
3681 case Instruction::Br:
3682 case Instruction::IndirectBr:
3683 case Instruction::Switch:
3684 case Instruction::Unreachable:
3685 case Instruction::Fence:
3686 case Instruction::AtomicRMW:
3687 case Instruction::AtomicCmpXchg:
3688 case Instruction::LandingPad:
3689 case Instruction::Resume:
3690 case Instruction::CatchSwitch:
3691 case Instruction::CatchPad:
3692 case Instruction::CatchRet:
3693 case Instruction::CleanupPad:
3694 case Instruction::CleanupRet:
3695 return false; // Misc instructions which have effects
3699 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3700 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3703 OverflowResult llvm::computeOverflowForUnsignedMul(const Value *LHS,
3705 const DataLayout &DL,
3706 AssumptionCache *AC,
3707 const Instruction *CxtI,
3708 const DominatorTree *DT) {
3709 // Multiplying n * m significant bits yields a result of n + m significant
3710 // bits. If the total number of significant bits does not exceed the
3711 // result bit width (minus 1), there is no overflow.
3712 // This means if we have enough leading zero bits in the operands
3713 // we can guarantee that the result does not overflow.
3714 // Ref: "Hacker's Delight" by Henry Warren
3715 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3716 KnownBits LHSKnown(BitWidth);
3717 KnownBits RHSKnown(BitWidth);
3718 computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3719 computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3720 // Note that underestimating the number of zero bits gives a more
3721 // conservative answer.
3722 unsigned ZeroBits = LHSKnown.countMinLeadingZeros() +
3723 RHSKnown.countMinLeadingZeros();
3724 // First handle the easy case: if we have enough zero bits there's
3725 // definitely no overflow.
3726 if (ZeroBits >= BitWidth)
3727 return OverflowResult::NeverOverflows;
3729 // Get the largest possible values for each operand.
3730 APInt LHSMax = ~LHSKnown.Zero;
3731 APInt RHSMax = ~RHSKnown.Zero;
3733 // We know the multiply operation doesn't overflow if the maximum values for
3734 // each operand will not overflow after we multiply them together.
3736 (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
3738 return OverflowResult::NeverOverflows;
3740 // We know it always overflows if multiplying the smallest possible values for
3741 // the operands also results in overflow.
3743 (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
3745 return OverflowResult::AlwaysOverflows;
3747 return OverflowResult::MayOverflow;
3750 OverflowResult llvm::computeOverflowForSignedMul(const Value *LHS,
3752 const DataLayout &DL,
3753 AssumptionCache *AC,
3754 const Instruction *CxtI,
3755 const DominatorTree *DT) {
3756 // Multiplying n * m significant bits yields a result of n + m significant
3757 // bits. If the total number of significant bits does not exceed the
3758 // result bit width (minus 1), there is no overflow.
3759 // This means if we have enough leading sign bits in the operands
3760 // we can guarantee that the result does not overflow.
3761 // Ref: "Hacker's Delight" by Henry Warren
3762 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3764 // Note that underestimating the number of sign bits gives a more
3765 // conservative answer.
3766 unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
3767 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
3769 // First handle the easy case: if we have enough sign bits there's
3770 // definitely no overflow.
3771 if (SignBits > BitWidth + 1)
3772 return OverflowResult::NeverOverflows;
3774 // There are two ambiguous cases where there can be no overflow:
3775 // SignBits == BitWidth + 1 and
3776 // SignBits == BitWidth
3777 // The second case is difficult to check, therefore we only handle the
3779 if (SignBits == BitWidth + 1) {
3780 // It overflows only when both arguments are negative and the true
3781 // product is exactly the minimum negative number.
3782 // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
3783 // For simplicity we just check if at least one side is not negative.
3784 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3785 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3786 if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
3787 return OverflowResult::NeverOverflows;
3789 return OverflowResult::MayOverflow;
3792 OverflowResult llvm::computeOverflowForUnsignedAdd(const Value *LHS,
3794 const DataLayout &DL,
3795 AssumptionCache *AC,
3796 const Instruction *CxtI,
3797 const DominatorTree *DT) {
3798 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3799 if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
3800 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3802 if (LHSKnown.isNegative() && RHSKnown.isNegative()) {
3803 // The sign bit is set in both cases: this MUST overflow.
3804 // Create a simple add instruction, and insert it into the struct.
3805 return OverflowResult::AlwaysOverflows;
3808 if (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()) {
3809 // The sign bit is clear in both cases: this CANNOT overflow.
3810 // Create a simple add instruction, and insert it into the struct.
3811 return OverflowResult::NeverOverflows;
3815 return OverflowResult::MayOverflow;
3818 /// Return true if we can prove that adding the two values of the
3819 /// knownbits will not overflow.
3820 /// Otherwise return false.
3821 static bool checkRippleForSignedAdd(const KnownBits &LHSKnown,
3822 const KnownBits &RHSKnown) {
3823 // Addition of two 2's complement numbers having opposite signs will never
3825 if ((LHSKnown.isNegative() && RHSKnown.isNonNegative()) ||
3826 (LHSKnown.isNonNegative() && RHSKnown.isNegative()))
3829 // If either of the values is known to be non-negative, adding them can only
3830 // overflow if the second is also non-negative, so we can assume that.
3831 // Two non-negative numbers will only overflow if there is a carry to the
3832 // sign bit, so we can check if even when the values are as big as possible
3833 // there is no overflow to the sign bit.
3834 if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) {
3835 APInt MaxLHS = ~LHSKnown.Zero;
3836 MaxLHS.clearSignBit();
3837 APInt MaxRHS = ~RHSKnown.Zero;
3838 MaxRHS.clearSignBit();
3839 APInt Result = std::move(MaxLHS) + std::move(MaxRHS);
3840 return Result.isSignBitClear();
3843 // If either of the values is known to be negative, adding them can only
3844 // overflow if the second is also negative, so we can assume that.
3845 // Two negative number will only overflow if there is no carry to the sign
3846 // bit, so we can check if even when the values are as small as possible
3847 // there is overflow to the sign bit.
3848 if (LHSKnown.isNegative() || RHSKnown.isNegative()) {
3849 APInt MinLHS = LHSKnown.One;
3850 MinLHS.clearSignBit();
3851 APInt MinRHS = RHSKnown.One;
3852 MinRHS.clearSignBit();
3853 APInt Result = std::move(MinLHS) + std::move(MinRHS);
3854 return Result.isSignBitSet();
3857 // If we reached here it means that we know nothing about the sign bits.
3858 // In this case we can't know if there will be an overflow, since by
3859 // changing the sign bits any two values can be made to overflow.
3863 static OverflowResult computeOverflowForSignedAdd(const Value *LHS,
3865 const AddOperator *Add,
3866 const DataLayout &DL,
3867 AssumptionCache *AC,
3868 const Instruction *CxtI,
3869 const DominatorTree *DT) {
3870 if (Add && Add->hasNoSignedWrap()) {
3871 return OverflowResult::NeverOverflows;
3874 // If LHS and RHS each have at least two sign bits, the addition will look
3880 // If the carry into the most significant position is 0, X and Y can't both
3881 // be 1 and therefore the carry out of the addition is also 0.
3883 // If the carry into the most significant position is 1, X and Y can't both
3884 // be 0 and therefore the carry out of the addition is also 1.
3886 // Since the carry into the most significant position is always equal to
3887 // the carry out of the addition, there is no signed overflow.
3888 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
3889 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
3890 return OverflowResult::NeverOverflows;
3892 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3893 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3895 if (checkRippleForSignedAdd(LHSKnown, RHSKnown))
3896 return OverflowResult::NeverOverflows;
3898 // The remaining code needs Add to be available. Early returns if not so.
3900 return OverflowResult::MayOverflow;
3902 // If the sign of Add is the same as at least one of the operands, this add
3903 // CANNOT overflow. This is particularly useful when the sum is
3904 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3906 bool LHSOrRHSKnownNonNegative =
3907 (LHSKnown.isNonNegative() || RHSKnown.isNonNegative());
3908 bool LHSOrRHSKnownNegative =
3909 (LHSKnown.isNegative() || RHSKnown.isNegative());
3910 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3911 KnownBits AddKnown = computeKnownBits(Add, DL, /*Depth=*/0, AC, CxtI, DT);
3912 if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
3913 (AddKnown.isNegative() && LHSOrRHSKnownNegative)) {
3914 return OverflowResult::NeverOverflows;
3918 return OverflowResult::MayOverflow;
3921 OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS,
3923 const DataLayout &DL,
3924 AssumptionCache *AC,
3925 const Instruction *CxtI,
3926 const DominatorTree *DT) {
3927 // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
3928 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3929 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3930 if (LHSKnown.isNegative() && RHSKnown.isNonNegative())
3931 return OverflowResult::NeverOverflows;
3933 return OverflowResult::MayOverflow;
3936 OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS,
3938 const DataLayout &DL,
3939 AssumptionCache *AC,
3940 const Instruction *CxtI,
3941 const DominatorTree *DT) {
3942 // If LHS and RHS each have at least two sign bits, the subtraction
3944 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
3945 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
3946 return OverflowResult::NeverOverflows;
3948 KnownBits LHSKnown = computeKnownBits(LHS, DL, 0, AC, CxtI, DT);
3950 KnownBits RHSKnown = computeKnownBits(RHS, DL, 0, AC, CxtI, DT);
3952 // Subtraction of two 2's complement numbers having identical signs will
3954 if ((LHSKnown.isNegative() && RHSKnown.isNegative()) ||
3955 (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()))
3956 return OverflowResult::NeverOverflows;
3958 // TODO: implement logic similar to checkRippleForAdd
3959 return OverflowResult::MayOverflow;
3962 bool llvm::isOverflowIntrinsicNoWrap(const IntrinsicInst *II,
3963 const DominatorTree &DT) {
3965 auto IID = II->getIntrinsicID();
3966 assert((IID == Intrinsic::sadd_with_overflow ||
3967 IID == Intrinsic::uadd_with_overflow ||
3968 IID == Intrinsic::ssub_with_overflow ||
3969 IID == Intrinsic::usub_with_overflow ||
3970 IID == Intrinsic::smul_with_overflow ||
3971 IID == Intrinsic::umul_with_overflow) &&
3972 "Not an overflow intrinsic!");
3975 SmallVector<const BranchInst *, 2> GuardingBranches;
3976 SmallVector<const ExtractValueInst *, 2> Results;
3978 for (const User *U : II->users()) {
3979 if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
3980 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
3982 if (EVI->getIndices()[0] == 0)
3983 Results.push_back(EVI);
3985 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
3987 for (const auto *U : EVI->users())
3988 if (const auto *B = dyn_cast<BranchInst>(U)) {
3989 assert(B->isConditional() && "How else is it using an i1?");
3990 GuardingBranches.push_back(B);
3994 // We are using the aggregate directly in a way we don't want to analyze
3995 // here (storing it to a global, say).
4000 auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
4001 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
4002 if (!NoWrapEdge.isSingleEdge())
4005 // Check if all users of the add are provably no-wrap.
4006 for (const auto *Result : Results) {
4007 // If the extractvalue itself is not executed on overflow, the we don't
4008 // need to check each use separately, since domination is transitive.
4009 if (DT.dominates(NoWrapEdge, Result->getParent()))
4012 for (auto &RU : Result->uses())
4013 if (!DT.dominates(NoWrapEdge, RU))
4020 return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
4024 OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add,
4025 const DataLayout &DL,
4026 AssumptionCache *AC,
4027 const Instruction *CxtI,
4028 const DominatorTree *DT) {
4029 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
4030 Add, DL, AC, CxtI, DT);
4033 OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS,
4035 const DataLayout &DL,
4036 AssumptionCache *AC,
4037 const Instruction *CxtI,
4038 const DominatorTree *DT) {
4039 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
4042 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
4043 // A memory operation returns normally if it isn't volatile. A volatile
4044 // operation is allowed to trap.
4046 // An atomic operation isn't guaranteed to return in a reasonable amount of
4047 // time because it's possible for another thread to interfere with it for an
4048 // arbitrary length of time, but programs aren't allowed to rely on that.
4049 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
4050 return !LI->isVolatile();
4051 if (const StoreInst *SI = dyn_cast<StoreInst>(I))
4052 return !SI->isVolatile();
4053 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
4054 return !CXI->isVolatile();
4055 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
4056 return !RMWI->isVolatile();
4057 if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
4058 return !MII->isVolatile();
4060 // If there is no successor, then execution can't transfer to it.
4061 if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
4062 return !CRI->unwindsToCaller();
4063 if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
4064 return !CatchSwitch->unwindsToCaller();
4065 if (isa<ResumeInst>(I))
4067 if (isa<ReturnInst>(I))
4069 if (isa<UnreachableInst>(I))
4072 // Calls can throw, or contain an infinite loop, or kill the process.
4073 if (auto CS = ImmutableCallSite(I)) {
4074 // Call sites that throw have implicit non-local control flow.
4075 if (!CS.doesNotThrow())
4078 // Non-throwing call sites can loop infinitely, call exit/pthread_exit
4079 // etc. and thus not return. However, LLVM already assumes that
4081 // - Thread exiting actions are modeled as writes to memory invisible to
4084 // - Loops that don't have side effects (side effects are volatile/atomic
4085 // stores and IO) always terminate (see http://llvm.org/PR965).
4086 // Furthermore IO itself is also modeled as writes to memory invisible to
4089 // We rely on those assumptions here, and use the memory effects of the call
4090 // target as a proxy for checking that it always returns.
4092 // FIXME: This isn't aggressive enough; a call which only writes to a global
4093 // is guaranteed to return.
4094 return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
4095 match(I, m_Intrinsic<Intrinsic::assume>()) ||
4096 match(I, m_Intrinsic<Intrinsic::sideeffect>());
4099 // Other instructions return normally.
4103 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) {
4104 // TODO: This is slightly consdervative for invoke instruction since exiting
4105 // via an exception *is* normal control for them.
4106 for (auto I = BB->begin(), E = BB->end(); I != E; ++I)
4107 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
4112 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
4114 // The loop header is guaranteed to be executed for every iteration.
4116 // FIXME: Relax this constraint to cover all basic blocks that are
4117 // guaranteed to be executed at every iteration.
4118 if (I->getParent() != L->getHeader()) return false;
4120 for (const Instruction &LI : *L->getHeader()) {
4121 if (&LI == I) return true;
4122 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
4124 llvm_unreachable("Instruction not contained in its own parent basic block.");
4127 bool llvm::propagatesFullPoison(const Instruction *I) {
4128 switch (I->getOpcode()) {
4129 case Instruction::Add:
4130 case Instruction::Sub:
4131 case Instruction::Xor:
4132 case Instruction::Trunc:
4133 case Instruction::BitCast:
4134 case Instruction::AddrSpaceCast:
4135 case Instruction::Mul:
4136 case Instruction::Shl:
4137 case Instruction::GetElementPtr:
4138 // These operations all propagate poison unconditionally. Note that poison
4139 // is not any particular value, so xor or subtraction of poison with
4140 // itself still yields poison, not zero.
4143 case Instruction::AShr:
4144 case Instruction::SExt:
4145 // For these operations, one bit of the input is replicated across
4146 // multiple output bits. A replicated poison bit is still poison.
4149 case Instruction::ICmp:
4150 // Comparing poison with any value yields poison. This is why, for
4151 // instance, x s< (x +nsw 1) can be folded to true.
4159 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
4160 switch (I->getOpcode()) {
4161 case Instruction::Store:
4162 return cast<StoreInst>(I)->getPointerOperand();
4164 case Instruction::Load:
4165 return cast<LoadInst>(I)->getPointerOperand();
4167 case Instruction::AtomicCmpXchg:
4168 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
4170 case Instruction::AtomicRMW:
4171 return cast<AtomicRMWInst>(I)->getPointerOperand();
4173 case Instruction::UDiv:
4174 case Instruction::SDiv:
4175 case Instruction::URem:
4176 case Instruction::SRem:
4177 return I->getOperand(1);
4184 bool llvm::programUndefinedIfFullPoison(const Instruction *PoisonI) {
4185 // We currently only look for uses of poison values within the same basic
4186 // block, as that makes it easier to guarantee that the uses will be
4187 // executed given that PoisonI is executed.
4189 // FIXME: Expand this to consider uses beyond the same basic block. To do
4190 // this, look out for the distinction between post-dominance and strong
4192 const BasicBlock *BB = PoisonI->getParent();
4194 // Set of instructions that we have proved will yield poison if PoisonI
4196 SmallSet<const Value *, 16> YieldsPoison;
4197 SmallSet<const BasicBlock *, 4> Visited;
4198 YieldsPoison.insert(PoisonI);
4199 Visited.insert(PoisonI->getParent());
4201 BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
4204 while (Iter++ < MaxDepth) {
4205 for (auto &I : make_range(Begin, End)) {
4206 if (&I != PoisonI) {
4207 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
4208 if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
4210 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
4214 // Mark poison that propagates from I through uses of I.
4215 if (YieldsPoison.count(&I)) {
4216 for (const User *User : I.users()) {
4217 const Instruction *UserI = cast<Instruction>(User);
4218 if (propagatesFullPoison(UserI))
4219 YieldsPoison.insert(User);
4224 if (auto *NextBB = BB->getSingleSuccessor()) {
4225 if (Visited.insert(NextBB).second) {
4227 Begin = BB->getFirstNonPHI()->getIterator();
4238 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
4242 if (auto *C = dyn_cast<ConstantFP>(V))
4247 static bool isKnownNonZero(const Value *V) {
4248 if (auto *C = dyn_cast<ConstantFP>(V))
4249 return !C->isZero();
4253 /// Match clamp pattern for float types without care about NaNs or signed zeros.
4254 /// Given non-min/max outer cmp/select from the clamp pattern this
4255 /// function recognizes if it can be substitued by a "canonical" min/max
4257 static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred,
4258 Value *CmpLHS, Value *CmpRHS,
4259 Value *TrueVal, Value *FalseVal,
4260 Value *&LHS, Value *&RHS) {
4262 // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2))
4263 // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2))
4264 // and return description of the outer Max/Min.
4266 // First, check if select has inverse order:
4267 if (CmpRHS == FalseVal) {
4268 std::swap(TrueVal, FalseVal);
4269 Pred = CmpInst::getInversePredicate(Pred);
4272 // Assume success now. If there's no match, callers should not use these anyway.
4277 if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite())
4278 return {SPF_UNKNOWN, SPNB_NA, false};
4282 case CmpInst::FCMP_OLT:
4283 case CmpInst::FCMP_OLE:
4284 case CmpInst::FCMP_ULT:
4285 case CmpInst::FCMP_ULE:
4287 m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)),
4288 m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4289 FC1->compare(*FC2) == APFloat::cmpResult::cmpLessThan)
4290 return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false};
4292 case CmpInst::FCMP_OGT:
4293 case CmpInst::FCMP_OGE:
4294 case CmpInst::FCMP_UGT:
4295 case CmpInst::FCMP_UGE:
4297 m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)),
4298 m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4299 FC1->compare(*FC2) == APFloat::cmpResult::cmpGreaterThan)
4300 return {SPF_FMINNUM, SPNB_RETURNS_ANY, false};
4306 return {SPF_UNKNOWN, SPNB_NA, false};
4309 /// Recognize variations of:
4310 /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
4311 static SelectPatternResult matchClamp(CmpInst::Predicate Pred,
4312 Value *CmpLHS, Value *CmpRHS,
4313 Value *TrueVal, Value *FalseVal) {
4314 // Swap the select operands and predicate to match the patterns below.
4315 if (CmpRHS != TrueVal) {
4316 Pred = ICmpInst::getSwappedPredicate(Pred);
4317 std::swap(TrueVal, FalseVal);
4320 if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
4322 // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
4323 if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4324 C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
4325 return {SPF_SMAX, SPNB_NA, false};
4327 // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
4328 if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4329 C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
4330 return {SPF_SMIN, SPNB_NA, false};
4332 // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
4333 if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4334 C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
4335 return {SPF_UMAX, SPNB_NA, false};
4337 // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
4338 if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4339 C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
4340 return {SPF_UMIN, SPNB_NA, false};
4342 return {SPF_UNKNOWN, SPNB_NA, false};
4345 /// Recognize variations of:
4346 /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c))
4347 static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred,
4348 Value *CmpLHS, Value *CmpRHS,
4349 Value *TVal, Value *FVal,
4351 // TODO: Allow FP min/max with nnan/nsz.
4352 assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison");
4355 SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1);
4356 if (!SelectPatternResult::isMinOrMax(L.Flavor))
4357 return {SPF_UNKNOWN, SPNB_NA, false};
4360 SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1);
4361 if (L.Flavor != R.Flavor)
4362 return {SPF_UNKNOWN, SPNB_NA, false};
4364 // We have something like: x Pred y ? min(a, b) : min(c, d).
4365 // Try to match the compare to the min/max operations of the select operands.
4366 // First, make sure we have the right compare predicate.
4369 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {
4370 Pred = ICmpInst::getSwappedPredicate(Pred);
4371 std::swap(CmpLHS, CmpRHS);
4373 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
4375 return {SPF_UNKNOWN, SPNB_NA, false};
4377 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
4378 Pred = ICmpInst::getSwappedPredicate(Pred);
4379 std::swap(CmpLHS, CmpRHS);
4381 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
4383 return {SPF_UNKNOWN, SPNB_NA, false};
4385 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
4386 Pred = ICmpInst::getSwappedPredicate(Pred);
4387 std::swap(CmpLHS, CmpRHS);
4389 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)
4391 return {SPF_UNKNOWN, SPNB_NA, false};
4393 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
4394 Pred = ICmpInst::getSwappedPredicate(Pred);
4395 std::swap(CmpLHS, CmpRHS);
4397 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
4399 return {SPF_UNKNOWN, SPNB_NA, false};
4401 return {SPF_UNKNOWN, SPNB_NA, false};
4404 // If there is a common operand in the already matched min/max and the other
4405 // min/max operands match the compare operands (either directly or inverted),
4406 // then this is min/max of the same flavor.
4408 // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4409 // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4411 if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4412 match(A, m_Not(m_Specific(CmpRHS)))))
4413 return {L.Flavor, SPNB_NA, false};
4415 // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4416 // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4418 if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4419 match(A, m_Not(m_Specific(CmpRHS)))))
4420 return {L.Flavor, SPNB_NA, false};
4422 // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4423 // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4425 if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4426 match(B, m_Not(m_Specific(CmpRHS)))))
4427 return {L.Flavor, SPNB_NA, false};
4429 // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4430 // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4432 if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4433 match(B, m_Not(m_Specific(CmpRHS)))))
4434 return {L.Flavor, SPNB_NA, false};
4437 return {SPF_UNKNOWN, SPNB_NA, false};
4440 /// Match non-obvious integer minimum and maximum sequences.
4441 static SelectPatternResult matchMinMax(CmpInst::Predicate Pred,
4442 Value *CmpLHS, Value *CmpRHS,
4443 Value *TrueVal, Value *FalseVal,
4444 Value *&LHS, Value *&RHS,
4446 // Assume success. If there's no match, callers should not use these anyway.
4450 SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal);
4451 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
4454 SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth);
4455 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
4458 if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
4459 return {SPF_UNKNOWN, SPNB_NA, false};
4462 // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
4463 // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
4464 if (match(TrueVal, m_Zero()) &&
4465 match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4466 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4469 // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
4470 // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
4471 if (match(FalseVal, m_Zero()) &&
4472 match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4473 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4476 if (!match(CmpRHS, m_APInt(C1)))
4477 return {SPF_UNKNOWN, SPNB_NA, false};
4479 // An unsigned min/max can be written with a signed compare.
4481 if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
4482 (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
4483 // Is the sign bit set?
4484 // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
4485 // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
4486 if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() &&
4487 C2->isMaxSignedValue())
4488 return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4490 // Is the sign bit clear?
4491 // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
4492 // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
4493 if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
4494 C2->isMinSignedValue())
4495 return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4498 // Look through 'not' ops to find disguised signed min/max.
4499 // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
4500 // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
4501 if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
4502 match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
4503 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4505 // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
4506 // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
4507 if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
4508 match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
4509 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4511 return {SPF_UNKNOWN, SPNB_NA, false};
4514 bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) {
4515 assert(X && Y && "Invalid operand");
4517 // X = sub (0, Y) || X = sub nsw (0, Y)
4518 if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) ||
4519 (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y)))))
4522 // Y = sub (0, X) || Y = sub nsw (0, X)
4523 if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) ||
4524 (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X)))))
4527 // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A)
4529 return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) &&
4530 match(Y, m_Sub(m_Specific(B), m_Specific(A))))) ||
4531 (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) &&
4532 match(Y, m_NSWSub(m_Specific(B), m_Specific(A)))));
4535 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
4537 Value *CmpLHS, Value *CmpRHS,
4538 Value *TrueVal, Value *FalseVal,
4539 Value *&LHS, Value *&RHS,
4544 // Signed zero may return inconsistent results between implementations.
4545 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
4546 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
4547 // Therefore, we behave conservatively and only proceed if at least one of the
4548 // operands is known to not be zero or if we don't care about signed zero.
4551 // FIXME: Include OGT/OLT/UGT/ULT.
4552 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
4553 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
4554 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4555 !isKnownNonZero(CmpRHS))
4556 return {SPF_UNKNOWN, SPNB_NA, false};
4559 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
4560 bool Ordered = false;
4562 // When given one NaN and one non-NaN input:
4563 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
4564 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
4565 // ordered comparison fails), which could be NaN or non-NaN.
4566 // so here we discover exactly what NaN behavior is required/accepted.
4567 if (CmpInst::isFPPredicate(Pred)) {
4568 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
4569 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
4571 if (LHSSafe && RHSSafe) {
4572 // Both operands are known non-NaN.
4573 NaNBehavior = SPNB_RETURNS_ANY;
4574 } else if (CmpInst::isOrdered(Pred)) {
4575 // An ordered comparison will return false when given a NaN, so it
4579 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
4580 NaNBehavior = SPNB_RETURNS_NAN;
4582 NaNBehavior = SPNB_RETURNS_OTHER;
4584 // Completely unsafe.
4585 return {SPF_UNKNOWN, SPNB_NA, false};
4588 // An unordered comparison will return true when given a NaN, so it
4591 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
4592 NaNBehavior = SPNB_RETURNS_OTHER;
4594 NaNBehavior = SPNB_RETURNS_NAN;
4596 // Completely unsafe.
4597 return {SPF_UNKNOWN, SPNB_NA, false};
4601 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
4602 std::swap(CmpLHS, CmpRHS);
4603 Pred = CmpInst::getSwappedPredicate(Pred);
4604 if (NaNBehavior == SPNB_RETURNS_NAN)
4605 NaNBehavior = SPNB_RETURNS_OTHER;
4606 else if (NaNBehavior == SPNB_RETURNS_OTHER)
4607 NaNBehavior = SPNB_RETURNS_NAN;
4611 // ([if]cmp X, Y) ? X : Y
4612 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
4614 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
4615 case ICmpInst::ICMP_UGT:
4616 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
4617 case ICmpInst::ICMP_SGT:
4618 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
4619 case ICmpInst::ICMP_ULT:
4620 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
4621 case ICmpInst::ICMP_SLT:
4622 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
4623 case FCmpInst::FCMP_UGT:
4624 case FCmpInst::FCMP_UGE:
4625 case FCmpInst::FCMP_OGT:
4626 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
4627 case FCmpInst::FCMP_ULT:
4628 case FCmpInst::FCMP_ULE:
4629 case FCmpInst::FCMP_OLT:
4630 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
4634 if (isKnownNegation(TrueVal, FalseVal)) {
4635 // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can
4636 // match against either LHS or sext(LHS).
4637 auto MaybeSExtCmpLHS =
4638 m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS)));
4639 auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes());
4640 auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One());
4641 if (match(TrueVal, MaybeSExtCmpLHS)) {
4642 // Set the return values. If the compare uses the negated value (-X >s 0),
4643 // swap the return values because the negated value is always 'RHS'.
4646 if (match(CmpLHS, m_Neg(m_Specific(FalseVal))))
4647 std::swap(LHS, RHS);
4649 // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X)
4650 // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X)
4651 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4652 return {SPF_ABS, SPNB_NA, false};
4654 // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X)
4655 // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X)
4656 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4657 return {SPF_NABS, SPNB_NA, false};
4659 else if (match(FalseVal, MaybeSExtCmpLHS)) {
4660 // Set the return values. If the compare uses the negated value (-X >s 0),
4661 // swap the return values because the negated value is always 'RHS'.
4664 if (match(CmpLHS, m_Neg(m_Specific(TrueVal))))
4665 std::swap(LHS, RHS);
4667 // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X)
4668 // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X)
4669 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4670 return {SPF_NABS, SPNB_NA, false};
4672 // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X)
4673 // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X)
4674 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4675 return {SPF_ABS, SPNB_NA, false};
4679 if (CmpInst::isIntPredicate(Pred))
4680 return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth);
4682 // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar
4683 // may return either -0.0 or 0.0, so fcmp/select pair has stricter
4684 // semantics than minNum. Be conservative in such case.
4685 if (NaNBehavior != SPNB_RETURNS_ANY ||
4686 (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4687 !isKnownNonZero(CmpRHS)))
4688 return {SPF_UNKNOWN, SPNB_NA, false};
4690 return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
4693 /// Helps to match a select pattern in case of a type mismatch.
4695 /// The function processes the case when type of true and false values of a
4696 /// select instruction differs from type of the cmp instruction operands because
4697 /// of a cast instruction. The function checks if it is legal to move the cast
4698 /// operation after "select". If yes, it returns the new second value of
4699 /// "select" (with the assumption that cast is moved):
4700 /// 1. As operand of cast instruction when both values of "select" are same cast
4702 /// 2. As restored constant (by applying reverse cast operation) when the first
4703 /// value of the "select" is a cast operation and the second value is a
4705 /// NOTE: We return only the new second value because the first value could be
4706 /// accessed as operand of cast instruction.
4707 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4708 Instruction::CastOps *CastOp) {
4709 auto *Cast1 = dyn_cast<CastInst>(V1);
4713 *CastOp = Cast1->getOpcode();
4714 Type *SrcTy = Cast1->getSrcTy();
4715 if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
4716 // If V1 and V2 are both the same cast from the same type, look through V1.
4717 if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
4718 return Cast2->getOperand(0);
4722 auto *C = dyn_cast<Constant>(V2);
4726 Constant *CastedTo = nullptr;
4728 case Instruction::ZExt:
4729 if (CmpI->isUnsigned())
4730 CastedTo = ConstantExpr::getTrunc(C, SrcTy);
4732 case Instruction::SExt:
4733 if (CmpI->isSigned())
4734 CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
4736 case Instruction::Trunc:
4738 if (match(CmpI->getOperand(1), m_Constant(CmpConst)) &&
4739 CmpConst->getType() == SrcTy) {
4740 // Here we have the following case:
4742 // %cond = cmp iN %x, CmpConst
4743 // %tr = trunc iN %x to iK
4744 // %narrowsel = select i1 %cond, iK %t, iK C
4746 // We can always move trunc after select operation:
4748 // %cond = cmp iN %x, CmpConst
4749 // %widesel = select i1 %cond, iN %x, iN CmpConst
4750 // %tr = trunc iN %widesel to iK
4752 // Note that C could be extended in any way because we don't care about
4753 // upper bits after truncation. It can't be abs pattern, because it would
4756 // select i1 %cond, x, -x.
4758 // So only min/max pattern could be matched. Such match requires widened C
4759 // == CmpConst. That is why set widened C = CmpConst, condition trunc
4760 // CmpConst == C is checked below.
4761 CastedTo = CmpConst;
4763 CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
4766 case Instruction::FPTrunc:
4767 CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
4769 case Instruction::FPExt:
4770 CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
4772 case Instruction::FPToUI:
4773 CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
4775 case Instruction::FPToSI:
4776 CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
4778 case Instruction::UIToFP:
4779 CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
4781 case Instruction::SIToFP:
4782 CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
4791 // Make sure the cast doesn't lose any information.
4792 Constant *CastedBack =
4793 ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
4794 if (CastedBack != C)
4800 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
4801 Instruction::CastOps *CastOp,
4803 if (Depth >= MaxDepth)
4804 return {SPF_UNKNOWN, SPNB_NA, false};
4806 SelectInst *SI = dyn_cast<SelectInst>(V);
4807 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4809 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4810 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4812 CmpInst::Predicate Pred = CmpI->getPredicate();
4813 Value *CmpLHS = CmpI->getOperand(0);
4814 Value *CmpRHS = CmpI->getOperand(1);
4815 Value *TrueVal = SI->getTrueValue();
4816 Value *FalseVal = SI->getFalseValue();
4818 if (isa<FPMathOperator>(CmpI))
4819 FMF = CmpI->getFastMathFlags();
4822 if (CmpI->isEquality())
4823 return {SPF_UNKNOWN, SPNB_NA, false};
4825 // Deal with type mismatches.
4826 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4827 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) {
4828 // If this is a potential fmin/fmax with a cast to integer, then ignore
4829 // -0.0 because there is no corresponding integer value.
4830 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
4831 FMF.setNoSignedZeros();
4832 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4833 cast<CastInst>(TrueVal)->getOperand(0), C,
4836 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) {
4837 // If this is a potential fmin/fmax with a cast to integer, then ignore
4838 // -0.0 because there is no corresponding integer value.
4839 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
4840 FMF.setNoSignedZeros();
4841 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4842 C, cast<CastInst>(FalseVal)->getOperand(0),
4846 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4850 CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) {
4851 if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT;
4852 if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT;
4853 if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT;
4854 if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT;
4855 if (SPF == SPF_FMINNUM)
4856 return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
4857 if (SPF == SPF_FMAXNUM)
4858 return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
4859 llvm_unreachable("unhandled!");
4862 SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) {
4863 if (SPF == SPF_SMIN) return SPF_SMAX;
4864 if (SPF == SPF_UMIN) return SPF_UMAX;
4865 if (SPF == SPF_SMAX) return SPF_SMIN;
4866 if (SPF == SPF_UMAX) return SPF_UMIN;
4867 llvm_unreachable("unhandled!");
4870 CmpInst::Predicate llvm::getInverseMinMaxPred(SelectPatternFlavor SPF) {
4871 return getMinMaxPred(getInverseMinMaxFlavor(SPF));
4874 /// Return true if "icmp Pred LHS RHS" is always true.
4875 static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS,
4876 const Value *RHS, const DataLayout &DL,
4878 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
4879 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
4886 case CmpInst::ICMP_SLE: {
4889 // LHS s<= LHS +_{nsw} C if C >= 0
4890 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
4891 return !C->isNegative();
4895 case CmpInst::ICMP_ULE: {
4898 // LHS u<= LHS +_{nuw} C for any C
4899 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
4902 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
4903 auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B,
4905 const APInt *&CA, const APInt *&CB) {
4906 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
4907 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
4910 // If X & C == 0 then (X | C) == X +_{nuw} C
4911 if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
4912 match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
4913 KnownBits Known(CA->getBitWidth());
4914 computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr,
4915 /*CxtI*/ nullptr, /*DT*/ nullptr);
4916 if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero))
4924 const APInt *CLHS, *CRHS;
4925 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
4926 return CLHS->ule(*CRHS);
4933 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
4934 /// ALHS ARHS" is true. Otherwise, return None.
4935 static Optional<bool>
4936 isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS,
4937 const Value *ARHS, const Value *BLHS, const Value *BRHS,
4938 const DataLayout &DL, unsigned Depth) {
4943 case CmpInst::ICMP_SLT:
4944 case CmpInst::ICMP_SLE:
4945 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) &&
4946 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth))
4950 case CmpInst::ICMP_ULT:
4951 case CmpInst::ICMP_ULE:
4952 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) &&
4953 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth))
4959 /// Return true if the operands of the two compares match. IsSwappedOps is true
4960 /// when the operands match, but are swapped.
4961 static bool isMatchingOps(const Value *ALHS, const Value *ARHS,
4962 const Value *BLHS, const Value *BRHS,
4963 bool &IsSwappedOps) {
4965 bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS);
4966 IsSwappedOps = (ALHS == BRHS && ARHS == BLHS);
4967 return IsMatchingOps || IsSwappedOps;
4970 /// Return true if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS BRHS" is
4971 /// true. Return false if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS
4972 /// BRHS" is false. Otherwise, return None if we can't infer anything.
4973 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred,
4976 CmpInst::Predicate BPred,
4979 bool IsSwappedOps) {
4980 // Canonicalize the operands so they're matching.
4982 std::swap(BLHS, BRHS);
4983 BPred = ICmpInst::getSwappedPredicate(BPred);
4985 if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred))
4987 if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred))
4993 /// Return true if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS C2" is
4994 /// true. Return false if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS
4995 /// C2" is false. Otherwise, return None if we can't infer anything.
4996 static Optional<bool>
4997 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, const Value *ALHS,
4998 const ConstantInt *C1,
4999 CmpInst::Predicate BPred,
5000 const Value *BLHS, const ConstantInt *C2) {
5001 assert(ALHS == BLHS && "LHS operands must match.");
5002 ConstantRange DomCR =
5003 ConstantRange::makeExactICmpRegion(APred, C1->getValue());
5005 ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue());
5006 ConstantRange Intersection = DomCR.intersectWith(CR);
5007 ConstantRange Difference = DomCR.difference(CR);
5008 if (Intersection.isEmptySet())
5010 if (Difference.isEmptySet())
5015 /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is
5016 /// false. Otherwise, return None if we can't infer anything.
5017 static Optional<bool> isImpliedCondICmps(const ICmpInst *LHS,
5018 const ICmpInst *RHS,
5019 const DataLayout &DL, bool LHSIsTrue,
5021 Value *ALHS = LHS->getOperand(0);
5022 Value *ARHS = LHS->getOperand(1);
5023 // The rest of the logic assumes the LHS condition is true. If that's not the
5024 // case, invert the predicate to make it so.
5025 ICmpInst::Predicate APred =
5026 LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate();
5028 Value *BLHS = RHS->getOperand(0);
5029 Value *BRHS = RHS->getOperand(1);
5030 ICmpInst::Predicate BPred = RHS->getPredicate();
5032 // Can we infer anything when the two compares have matching operands?
5034 if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, IsSwappedOps)) {
5035 if (Optional<bool> Implication = isImpliedCondMatchingOperands(
5036 APred, ALHS, ARHS, BPred, BLHS, BRHS, IsSwappedOps))
5038 // No amount of additional analysis will infer the second condition, so
5043 // Can we infer anything when the LHS operands match and the RHS operands are
5044 // constants (not necessarily matching)?
5045 if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) {
5046 if (Optional<bool> Implication = isImpliedCondMatchingImmOperands(
5047 APred, ALHS, cast<ConstantInt>(ARHS), BPred, BLHS,
5048 cast<ConstantInt>(BRHS)))
5050 // No amount of additional analysis will infer the second condition, so
5056 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth);
5060 /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is
5061 /// false. Otherwise, return None if we can't infer anything. We expect the
5062 /// RHS to be an icmp and the LHS to be an 'and' or an 'or' instruction.
5063 static Optional<bool> isImpliedCondAndOr(const BinaryOperator *LHS,
5064 const ICmpInst *RHS,
5065 const DataLayout &DL, bool LHSIsTrue,
5067 // The LHS must be an 'or' or an 'and' instruction.
5068 assert((LHS->getOpcode() == Instruction::And ||
5069 LHS->getOpcode() == Instruction::Or) &&
5070 "Expected LHS to be 'and' or 'or'.");
5072 assert(Depth <= MaxDepth && "Hit recursion limit");
5074 // If the result of an 'or' is false, then we know both legs of the 'or' are
5075 // false. Similarly, if the result of an 'and' is true, then we know both
5076 // legs of the 'and' are true.
5078 if ((!LHSIsTrue && match(LHS, m_Or(m_Value(ALHS), m_Value(ARHS)))) ||
5079 (LHSIsTrue && match(LHS, m_And(m_Value(ALHS), m_Value(ARHS))))) {
5080 // FIXME: Make this non-recursion.
5081 if (Optional<bool> Implication =
5082 isImpliedCondition(ALHS, RHS, DL, LHSIsTrue, Depth + 1))
5084 if (Optional<bool> Implication =
5085 isImpliedCondition(ARHS, RHS, DL, LHSIsTrue, Depth + 1))
5092 Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
5093 const DataLayout &DL, bool LHSIsTrue,
5095 // Bail out when we hit the limit.
5096 if (Depth == MaxDepth)
5099 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for
5101 if (LHS->getType() != RHS->getType())
5104 Type *OpTy = LHS->getType();
5105 assert(OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!");
5107 // LHS ==> RHS by definition
5111 // FIXME: Extending the code below to handle vectors.
5112 if (OpTy->isVectorTy())
5115 assert(OpTy->isIntegerTy(1) && "implied by above");
5117 // Both LHS and RHS are icmps.
5118 const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS);
5119 const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS);
5120 if (LHSCmp && RHSCmp)
5121 return isImpliedCondICmps(LHSCmp, RHSCmp, DL, LHSIsTrue, Depth);
5123 // The LHS should be an 'or' or an 'and' instruction. We expect the RHS to be
5124 // an icmp. FIXME: Add support for and/or on the RHS.
5125 const BinaryOperator *LHSBO = dyn_cast<BinaryOperator>(LHS);
5126 if (LHSBO && RHSCmp) {
5127 if ((LHSBO->getOpcode() == Instruction::And ||
5128 LHSBO->getOpcode() == Instruction::Or))
5129 return isImpliedCondAndOr(LHSBO, RHSCmp, DL, LHSIsTrue, Depth);