1 //===- InstCombineCompares.cpp --------------------------------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements the visitICmp and visitFCmp functions.
11 //===----------------------------------------------------------------------===//
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APSInt.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/ConstantFolding.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
20 #include "llvm/IR/ConstantRange.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/GetElementPtrTypeIterator.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/PatternMatch.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/KnownBits.h"
29 using namespace PatternMatch;
31 #define DEBUG_TYPE "instcombine"
33 // How many times is a select replaced by one of its operands?
34 STATISTIC(NumSel, "Number of select opts");
37 /// Compute Result = In1+In2, returning true if the result overflowed for this
39 static bool addWithOverflow(APInt &Result, const APInt &In1,
40 const APInt &In2, bool IsSigned = false) {
43 Result = In1.sadd_ov(In2, Overflow);
45 Result = In1.uadd_ov(In2, Overflow);
50 /// Compute Result = In1-In2, returning true if the result overflowed for this
52 static bool subWithOverflow(APInt &Result, const APInt &In1,
53 const APInt &In2, bool IsSigned = false) {
56 Result = In1.ssub_ov(In2, Overflow);
58 Result = In1.usub_ov(In2, Overflow);
63 /// Given an icmp instruction, return true if any use of this comparison is a
64 /// branch on sign bit comparison.
65 static bool hasBranchUse(ICmpInst &I) {
66 for (auto *U : I.users())
67 if (isa<BranchInst>(U))
72 /// Given an exploded icmp instruction, return true if the comparison only
73 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
74 /// result of the comparison is true when the input value is signed.
75 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
78 case ICmpInst::ICMP_SLT: // True if LHS s< 0
80 return RHS.isNullValue();
81 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
83 return RHS.isAllOnesValue();
84 case ICmpInst::ICMP_SGT: // True if LHS s> -1
86 return RHS.isAllOnesValue();
87 case ICmpInst::ICMP_UGT:
88 // True if LHS u> RHS and RHS == high-bit-mask - 1
90 return RHS.isMaxSignedValue();
91 case ICmpInst::ICMP_UGE:
92 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
94 return RHS.isSignMask();
100 /// Returns true if the exploded icmp can be expressed as a signed comparison
101 /// to zero and updates the predicate accordingly.
102 /// The signedness of the comparison is preserved.
103 /// TODO: Refactor with decomposeBitTestICmp()?
104 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
105 if (!ICmpInst::isSigned(Pred))
109 return ICmpInst::isRelational(Pred);
111 if (C.isOneValue()) {
112 if (Pred == ICmpInst::ICMP_SLT) {
113 Pred = ICmpInst::ICMP_SLE;
116 } else if (C.isAllOnesValue()) {
117 if (Pred == ICmpInst::ICMP_SGT) {
118 Pred = ICmpInst::ICMP_SGE;
126 /// Given a signed integer type and a set of known zero and one bits, compute
127 /// the maximum and minimum values that could have the specified known zero and
128 /// known one bits, returning them in Min/Max.
129 /// TODO: Move to method on KnownBits struct?
130 static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known,
131 APInt &Min, APInt &Max) {
132 assert(Known.getBitWidth() == Min.getBitWidth() &&
133 Known.getBitWidth() == Max.getBitWidth() &&
134 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
135 APInt UnknownBits = ~(Known.Zero|Known.One);
137 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
138 // bit if it is unknown.
140 Max = Known.One|UnknownBits;
142 if (UnknownBits.isNegative()) { // Sign bit is unknown
148 /// Given an unsigned integer type and a set of known zero and one bits, compute
149 /// the maximum and minimum values that could have the specified known zero and
150 /// known one bits, returning them in Min/Max.
151 /// TODO: Move to method on KnownBits struct?
152 static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known,
153 APInt &Min, APInt &Max) {
154 assert(Known.getBitWidth() == Min.getBitWidth() &&
155 Known.getBitWidth() == Max.getBitWidth() &&
156 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
157 APInt UnknownBits = ~(Known.Zero|Known.One);
159 // The minimum value is when the unknown bits are all zeros.
161 // The maximum value is when the unknown bits are all ones.
162 Max = Known.One|UnknownBits;
165 /// This is called when we see this pattern:
166 /// cmp pred (load (gep GV, ...)), cmpcst
167 /// where GV is a global variable with a constant initializer. Try to simplify
168 /// this into some simple computation that does not need the load. For example
169 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
171 /// If AndCst is non-null, then the loaded value is masked with that constant
172 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
173 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
176 ConstantInt *AndCst) {
177 Constant *Init = GV->getInitializer();
178 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
181 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
182 // Don't blow up on huge arrays.
183 if (ArrayElementCount > MaxArraySizeForCombine)
186 // There are many forms of this optimization we can handle, for now, just do
187 // the simple index into a single-dimensional array.
189 // Require: GEP GV, 0, i {{, constant indices}}
190 if (GEP->getNumOperands() < 3 ||
191 !isa<ConstantInt>(GEP->getOperand(1)) ||
192 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
193 isa<Constant>(GEP->getOperand(2)))
196 // Check that indices after the variable are constants and in-range for the
197 // type they index. Collect the indices. This is typically for arrays of
199 SmallVector<unsigned, 4> LaterIndices;
201 Type *EltTy = Init->getType()->getArrayElementType();
202 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
203 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
204 if (!Idx) return nullptr; // Variable index.
206 uint64_t IdxVal = Idx->getZExtValue();
207 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
209 if (StructType *STy = dyn_cast<StructType>(EltTy))
210 EltTy = STy->getElementType(IdxVal);
211 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
212 if (IdxVal >= ATy->getNumElements()) return nullptr;
213 EltTy = ATy->getElementType();
215 return nullptr; // Unknown type.
218 LaterIndices.push_back(IdxVal);
221 enum { Overdefined = -3, Undefined = -2 };
223 // Variables for our state machines.
225 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
226 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
227 // and 87 is the second (and last) index. FirstTrueElement is -2 when
228 // undefined, otherwise set to the first true element. SecondTrueElement is
229 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
230 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
232 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
233 // form "i != 47 & i != 87". Same state transitions as for true elements.
234 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
236 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
237 /// define a state machine that triggers for ranges of values that the index
238 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
239 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
240 /// index in the range (inclusive). We use -2 for undefined here because we
241 /// use relative comparisons and don't want 0-1 to match -1.
242 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
244 // MagicBitvector - This is a magic bitvector where we set a bit if the
245 // comparison is true for element 'i'. If there are 64 elements or less in
246 // the array, this will fully represent all the comparison results.
247 uint64_t MagicBitvector = 0;
249 // Scan the array and see if one of our patterns matches.
250 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
251 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
252 Constant *Elt = Init->getAggregateElement(i);
253 if (!Elt) return nullptr;
255 // If this is indexing an array of structures, get the structure element.
256 if (!LaterIndices.empty())
257 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
259 // If the element is masked, handle it.
260 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
262 // Find out if the comparison would be true or false for the i'th element.
263 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
264 CompareRHS, DL, &TLI);
265 // If the result is undef for this element, ignore it.
266 if (isa<UndefValue>(C)) {
267 // Extend range state machines to cover this element in case there is an
268 // undef in the middle of the range.
269 if (TrueRangeEnd == (int)i-1)
271 if (FalseRangeEnd == (int)i-1)
276 // If we can't compute the result for any of the elements, we have to give
277 // up evaluating the entire conditional.
278 if (!isa<ConstantInt>(C)) return nullptr;
280 // Otherwise, we know if the comparison is true or false for this element,
281 // update our state machines.
282 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
284 // State machine for single/double/range index comparison.
286 // Update the TrueElement state machine.
287 if (FirstTrueElement == Undefined)
288 FirstTrueElement = TrueRangeEnd = i; // First true element.
290 // Update double-compare state machine.
291 if (SecondTrueElement == Undefined)
292 SecondTrueElement = i;
294 SecondTrueElement = Overdefined;
296 // Update range state machine.
297 if (TrueRangeEnd == (int)i-1)
300 TrueRangeEnd = Overdefined;
303 // Update the FalseElement state machine.
304 if (FirstFalseElement == Undefined)
305 FirstFalseElement = FalseRangeEnd = i; // First false element.
307 // Update double-compare state machine.
308 if (SecondFalseElement == Undefined)
309 SecondFalseElement = i;
311 SecondFalseElement = Overdefined;
313 // Update range state machine.
314 if (FalseRangeEnd == (int)i-1)
317 FalseRangeEnd = Overdefined;
321 // If this element is in range, update our magic bitvector.
322 if (i < 64 && IsTrueForElt)
323 MagicBitvector |= 1ULL << i;
325 // If all of our states become overdefined, bail out early. Since the
326 // predicate is expensive, only check it every 8 elements. This is only
327 // really useful for really huge arrays.
328 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
329 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
330 FalseRangeEnd == Overdefined)
334 // Now that we've scanned the entire array, emit our new comparison(s). We
335 // order the state machines in complexity of the generated code.
336 Value *Idx = GEP->getOperand(2);
338 // If the index is larger than the pointer size of the target, truncate the
339 // index down like the GEP would do implicitly. We don't have to do this for
340 // an inbounds GEP because the index can't be out of range.
341 if (!GEP->isInBounds()) {
342 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
343 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
344 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
345 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
348 // If the comparison is only true for one or two elements, emit direct
350 if (SecondTrueElement != Overdefined) {
351 // None true -> false.
352 if (FirstTrueElement == Undefined)
353 return replaceInstUsesWith(ICI, Builder.getFalse());
355 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
357 // True for one element -> 'i == 47'.
358 if (SecondTrueElement == Undefined)
359 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
361 // True for two elements -> 'i == 47 | i == 72'.
362 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
363 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
364 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
365 return BinaryOperator::CreateOr(C1, C2);
368 // If the comparison is only false for one or two elements, emit direct
370 if (SecondFalseElement != Overdefined) {
371 // None false -> true.
372 if (FirstFalseElement == Undefined)
373 return replaceInstUsesWith(ICI, Builder.getTrue());
375 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
377 // False for one element -> 'i != 47'.
378 if (SecondFalseElement == Undefined)
379 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
381 // False for two elements -> 'i != 47 & i != 72'.
382 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
383 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
384 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
385 return BinaryOperator::CreateAnd(C1, C2);
388 // If the comparison can be replaced with a range comparison for the elements
389 // where it is true, emit the range check.
390 if (TrueRangeEnd != Overdefined) {
391 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
393 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
394 if (FirstTrueElement) {
395 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
396 Idx = Builder.CreateAdd(Idx, Offs);
399 Value *End = ConstantInt::get(Idx->getType(),
400 TrueRangeEnd-FirstTrueElement+1);
401 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
404 // False range check.
405 if (FalseRangeEnd != Overdefined) {
406 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
407 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
408 if (FirstFalseElement) {
409 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
410 Idx = Builder.CreateAdd(Idx, Offs);
413 Value *End = ConstantInt::get(Idx->getType(),
414 FalseRangeEnd-FirstFalseElement);
415 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
418 // If a magic bitvector captures the entire comparison state
419 // of this load, replace it with computation that does:
420 // ((magic_cst >> i) & 1) != 0
424 // Look for an appropriate type:
425 // - The type of Idx if the magic fits
426 // - The smallest fitting legal type
427 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
430 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
433 Value *V = Builder.CreateIntCast(Idx, Ty, false);
434 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
435 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
436 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
443 /// Return a value that can be used to compare the *offset* implied by a GEP to
444 /// zero. For example, if we have &A[i], we want to return 'i' for
445 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
446 /// are involved. The above expression would also be legal to codegen as
447 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
448 /// This latter form is less amenable to optimization though, and we are allowed
449 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
451 /// If we can't emit an optimized form for this expression, this returns null.
453 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC,
454 const DataLayout &DL) {
455 gep_type_iterator GTI = gep_type_begin(GEP);
457 // Check to see if this gep only has a single variable index. If so, and if
458 // any constant indices are a multiple of its scale, then we can compute this
459 // in terms of the scale of the variable index. For example, if the GEP
460 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
461 // because the expression will cross zero at the same point.
462 unsigned i, e = GEP->getNumOperands();
464 for (i = 1; i != e; ++i, ++GTI) {
465 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
466 // Compute the aggregate offset of constant indices.
467 if (CI->isZero()) continue;
469 // Handle a struct index, which adds its field offset to the pointer.
470 if (StructType *STy = GTI.getStructTypeOrNull()) {
471 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
473 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
474 Offset += Size*CI->getSExtValue();
477 // Found our variable index.
482 // If there are no variable indices, we must have a constant offset, just
483 // evaluate it the general way.
484 if (i == e) return nullptr;
486 Value *VariableIdx = GEP->getOperand(i);
487 // Determine the scale factor of the variable element. For example, this is
488 // 4 if the variable index is into an array of i32.
489 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
491 // Verify that there are no other variable indices. If so, emit the hard way.
492 for (++i, ++GTI; i != e; ++i, ++GTI) {
493 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
494 if (!CI) return nullptr;
496 // Compute the aggregate offset of constant indices.
497 if (CI->isZero()) continue;
499 // Handle a struct index, which adds its field offset to the pointer.
500 if (StructType *STy = GTI.getStructTypeOrNull()) {
501 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
503 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
504 Offset += Size*CI->getSExtValue();
508 // Okay, we know we have a single variable index, which must be a
509 // pointer/array/vector index. If there is no offset, life is simple, return
511 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
512 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
514 // Cast to intptrty in case a truncation occurs. If an extension is needed,
515 // we don't need to bother extending: the extension won't affect where the
516 // computation crosses zero.
517 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
518 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
523 // Otherwise, there is an index. The computation we will do will be modulo
525 Offset = SignExtend64(Offset, IntPtrWidth);
526 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
528 // To do this transformation, any constant index must be a multiple of the
529 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
530 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
531 // multiple of the variable scale.
532 int64_t NewOffs = Offset / (int64_t)VariableScale;
533 if (Offset != NewOffs*(int64_t)VariableScale)
536 // Okay, we can do this evaluation. Start by converting the index to intptr.
537 if (VariableIdx->getType() != IntPtrTy)
538 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
540 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
541 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
544 /// Returns true if we can rewrite Start as a GEP with pointer Base
545 /// and some integer offset. The nodes that need to be re-written
546 /// for this transformation will be added to Explored.
547 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
548 const DataLayout &DL,
549 SetVector<Value *> &Explored) {
550 SmallVector<Value *, 16> WorkList(1, Start);
551 Explored.insert(Base);
553 // The following traversal gives us an order which can be used
554 // when doing the final transformation. Since in the final
555 // transformation we create the PHI replacement instructions first,
556 // we don't have to get them in any particular order.
558 // However, for other instructions we will have to traverse the
559 // operands of an instruction first, which means that we have to
560 // do a post-order traversal.
561 while (!WorkList.empty()) {
562 SetVector<PHINode *> PHIs;
564 while (!WorkList.empty()) {
565 if (Explored.size() >= 100)
568 Value *V = WorkList.back();
570 if (Explored.count(V) != 0) {
575 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
576 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
577 // We've found some value that we can't explore which is different from
578 // the base. Therefore we can't do this transformation.
581 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
582 auto *CI = dyn_cast<CastInst>(V);
583 if (!CI->isNoopCast(DL))
586 if (Explored.count(CI->getOperand(0)) == 0)
587 WorkList.push_back(CI->getOperand(0));
590 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
591 // We're limiting the GEP to having one index. This will preserve
592 // the original pointer type. We could handle more cases in the
594 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
595 GEP->getType() != Start->getType())
598 if (Explored.count(GEP->getOperand(0)) == 0)
599 WorkList.push_back(GEP->getOperand(0));
602 if (WorkList.back() == V) {
604 // We've finished visiting this node, mark it as such.
608 if (auto *PN = dyn_cast<PHINode>(V)) {
609 // We cannot transform PHIs on unsplittable basic blocks.
610 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
617 // Explore the PHI nodes further.
618 for (auto *PN : PHIs)
619 for (Value *Op : PN->incoming_values())
620 if (Explored.count(Op) == 0)
621 WorkList.push_back(Op);
624 // Make sure that we can do this. Since we can't insert GEPs in a basic
625 // block before a PHI node, we can't easily do this transformation if
626 // we have PHI node users of transformed instructions.
627 for (Value *Val : Explored) {
628 for (Value *Use : Val->uses()) {
630 auto *PHI = dyn_cast<PHINode>(Use);
631 auto *Inst = dyn_cast<Instruction>(Val);
633 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
634 Explored.count(PHI) == 0)
637 if (PHI->getParent() == Inst->getParent())
644 // Sets the appropriate insert point on Builder where we can add
645 // a replacement Instruction for V (if that is possible).
646 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
647 bool Before = true) {
648 if (auto *PHI = dyn_cast<PHINode>(V)) {
649 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
652 if (auto *I = dyn_cast<Instruction>(V)) {
654 I = &*std::next(I->getIterator());
655 Builder.SetInsertPoint(I);
658 if (auto *A = dyn_cast<Argument>(V)) {
659 // Set the insertion point in the entry block.
660 BasicBlock &Entry = A->getParent()->getEntryBlock();
661 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
664 // Otherwise, this is a constant and we don't need to set a new
666 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
669 /// Returns a re-written value of Start as an indexed GEP using Base as a
671 static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
672 const DataLayout &DL,
673 SetVector<Value *> &Explored) {
674 // Perform all the substitutions. This is a bit tricky because we can
675 // have cycles in our use-def chains.
676 // 1. Create the PHI nodes without any incoming values.
677 // 2. Create all the other values.
678 // 3. Add the edges for the PHI nodes.
679 // 4. Emit GEPs to get the original pointers.
680 // 5. Remove the original instructions.
681 Type *IndexType = IntegerType::get(
682 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
684 DenseMap<Value *, Value *> NewInsts;
685 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
687 // Create the new PHI nodes, without adding any incoming values.
688 for (Value *Val : Explored) {
691 // Create empty phi nodes. This avoids cyclic dependencies when creating
692 // the remaining instructions.
693 if (auto *PHI = dyn_cast<PHINode>(Val))
694 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
695 PHI->getName() + ".idx", PHI);
697 IRBuilder<> Builder(Base->getContext());
699 // Create all the other instructions.
700 for (Value *Val : Explored) {
702 if (NewInsts.find(Val) != NewInsts.end())
705 if (auto *CI = dyn_cast<CastInst>(Val)) {
706 // Don't get rid of the intermediate variable here; the store can grow
707 // the map which will invalidate the reference to the input value.
708 Value *V = NewInsts[CI->getOperand(0)];
712 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
713 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
714 : GEP->getOperand(1);
715 setInsertionPoint(Builder, GEP);
716 // Indices might need to be sign extended. GEPs will magically do
717 // this, but we need to do it ourselves here.
718 if (Index->getType()->getScalarSizeInBits() !=
719 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
720 Index = Builder.CreateSExtOrTrunc(
721 Index, NewInsts[GEP->getOperand(0)]->getType(),
722 GEP->getOperand(0)->getName() + ".sext");
725 auto *Op = NewInsts[GEP->getOperand(0)];
726 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
727 NewInsts[GEP] = Index;
729 NewInsts[GEP] = Builder.CreateNSWAdd(
730 Op, Index, GEP->getOperand(0)->getName() + ".add");
733 if (isa<PHINode>(Val))
736 llvm_unreachable("Unexpected instruction type");
739 // Add the incoming values to the PHI nodes.
740 for (Value *Val : Explored) {
743 // All the instructions have been created, we can now add edges to the
745 if (auto *PHI = dyn_cast<PHINode>(Val)) {
746 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
747 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
748 Value *NewIncoming = PHI->getIncomingValue(I);
750 if (NewInsts.find(NewIncoming) != NewInsts.end())
751 NewIncoming = NewInsts[NewIncoming];
753 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
758 for (Value *Val : Explored) {
762 // Depending on the type, for external users we have to emit
763 // a GEP or a GEP + ptrtoint.
764 setInsertionPoint(Builder, Val, false);
766 // If required, create an inttoptr instruction for Base.
767 Value *NewBase = Base;
768 if (!Base->getType()->isPointerTy())
769 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
770 Start->getName() + "to.ptr");
772 Value *GEP = Builder.CreateInBoundsGEP(
773 Start->getType()->getPointerElementType(), NewBase,
774 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
776 if (!Val->getType()->isPointerTy()) {
777 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
778 Val->getName() + ".conv");
781 Val->replaceAllUsesWith(GEP);
784 return NewInsts[Start];
787 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
788 /// the input Value as a constant indexed GEP. Returns a pair containing
789 /// the GEPs Pointer and Index.
790 static std::pair<Value *, Value *>
791 getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
792 Type *IndexType = IntegerType::get(V->getContext(),
793 DL.getIndexTypeSizeInBits(V->getType()));
795 Constant *Index = ConstantInt::getNullValue(IndexType);
797 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
798 // We accept only inbouds GEPs here to exclude the possibility of
800 if (!GEP->isInBounds())
802 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
803 GEP->getType() == V->getType()) {
804 V = GEP->getOperand(0);
805 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
806 Index = ConstantExpr::getAdd(
807 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
812 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
813 if (!CI->isNoopCast(DL))
815 V = CI->getOperand(0);
818 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
819 if (!CI->isNoopCast(DL))
821 V = CI->getOperand(0);
829 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
830 /// We can look through PHIs, GEPs and casts in order to determine a common base
831 /// between GEPLHS and RHS.
832 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
833 ICmpInst::Predicate Cond,
834 const DataLayout &DL) {
835 // FIXME: Support vector of pointers.
836 if (GEPLHS->getType()->isVectorTy())
839 if (!GEPLHS->hasAllConstantIndices())
842 // Make sure the pointers have the same type.
843 if (GEPLHS->getType() != RHS->getType())
846 Value *PtrBase, *Index;
847 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
849 // The set of nodes that will take part in this transformation.
850 SetVector<Value *> Nodes;
852 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
855 // We know we can re-write this as
856 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
857 // Since we've only looked through inbouds GEPs we know that we
858 // can't have overflow on either side. We can therefore re-write
860 // OFFSET1 cmp OFFSET2
861 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
863 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
864 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
865 // offset. Since Index is the offset of LHS to the base pointer, we will now
866 // compare the offsets instead of comparing the pointers.
867 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
870 /// Fold comparisons between a GEP instruction and something else. At this point
871 /// we know that the GEP is on the LHS of the comparison.
872 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
873 ICmpInst::Predicate Cond,
875 // Don't transform signed compares of GEPs into index compares. Even if the
876 // GEP is inbounds, the final add of the base pointer can have signed overflow
877 // and would change the result of the icmp.
878 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
879 // the maximum signed value for the pointer type.
880 if (ICmpInst::isSigned(Cond))
883 // Look through bitcasts and addrspacecasts. We do not however want to remove
885 if (!isa<GetElementPtrInst>(RHS))
886 RHS = RHS->stripPointerCasts();
888 Value *PtrBase = GEPLHS->getOperand(0);
889 // FIXME: Support vector pointer GEPs.
890 if (PtrBase == RHS && GEPLHS->isInBounds() &&
891 !GEPLHS->getType()->isVectorTy()) {
892 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
893 // This transformation (ignoring the base and scales) is valid because we
894 // know pointers can't overflow since the gep is inbounds. See if we can
895 // output an optimized form.
896 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
898 // If not, synthesize the offset the hard way.
900 Offset = EmitGEPOffset(GEPLHS);
901 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
902 Constant::getNullValue(Offset->getType()));
903 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
904 // If the base pointers are different, but the indices are the same, just
905 // compare the base pointer.
906 if (PtrBase != GEPRHS->getOperand(0)) {
907 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
908 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
909 GEPRHS->getOperand(0)->getType();
911 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
912 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
913 IndicesTheSame = false;
917 // If all indices are the same, just compare the base pointers.
918 Type *BaseType = GEPLHS->getOperand(0)->getType();
919 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
920 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
922 // If we're comparing GEPs with two base pointers that only differ in type
923 // and both GEPs have only constant indices or just one use, then fold
924 // the compare with the adjusted indices.
925 // FIXME: Support vector of pointers.
926 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
927 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
928 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
929 PtrBase->stripPointerCasts() ==
930 GEPRHS->getOperand(0)->stripPointerCasts() &&
931 !GEPLHS->getType()->isVectorTy()) {
932 Value *LOffset = EmitGEPOffset(GEPLHS);
933 Value *ROffset = EmitGEPOffset(GEPRHS);
935 // If we looked through an addrspacecast between different sized address
936 // spaces, the LHS and RHS pointers are different sized
937 // integers. Truncate to the smaller one.
938 Type *LHSIndexTy = LOffset->getType();
939 Type *RHSIndexTy = ROffset->getType();
940 if (LHSIndexTy != RHSIndexTy) {
941 if (LHSIndexTy->getPrimitiveSizeInBits() <
942 RHSIndexTy->getPrimitiveSizeInBits()) {
943 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
945 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
948 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
950 return replaceInstUsesWith(I, Cmp);
953 // Otherwise, the base pointers are different and the indices are
954 // different. Try convert this to an indexed compare by looking through
956 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
959 // If one of the GEPs has all zero indices, recurse.
960 if (GEPLHS->hasAllZeroIndices())
961 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
962 ICmpInst::getSwappedPredicate(Cond), I);
964 // If the other GEP has all zero indices, recurse.
965 if (GEPRHS->hasAllZeroIndices())
966 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
968 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
969 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
970 // If the GEPs only differ by one index, compare it.
971 unsigned NumDifferences = 0; // Keep track of # differences.
972 unsigned DiffOperand = 0; // The operand that differs.
973 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
974 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
975 Type *LHSType = GEPLHS->getOperand(i)->getType();
976 Type *RHSType = GEPRHS->getOperand(i)->getType();
977 // FIXME: Better support for vector of pointers.
978 if (LHSType->getPrimitiveSizeInBits() !=
979 RHSType->getPrimitiveSizeInBits() ||
980 (GEPLHS->getType()->isVectorTy() &&
981 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
982 // Irreconcilable differences.
987 if (NumDifferences++) break;
991 if (NumDifferences == 0) // SAME GEP?
992 return replaceInstUsesWith(I, // No comparison is needed here.
993 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
995 else if (NumDifferences == 1 && GEPsInBounds) {
996 Value *LHSV = GEPLHS->getOperand(DiffOperand);
997 Value *RHSV = GEPRHS->getOperand(DiffOperand);
998 // Make sure we do a signed comparison here.
999 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1003 // Only lower this if the icmp is the only user of the GEP or if we expect
1004 // the result to fold to a constant!
1005 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1006 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1007 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1008 Value *L = EmitGEPOffset(GEPLHS);
1009 Value *R = EmitGEPOffset(GEPRHS);
1010 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1014 // Try convert this to an indexed compare by looking through PHIs/casts as a
1016 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1019 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1020 const AllocaInst *Alloca,
1021 const Value *Other) {
1022 assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1024 // It would be tempting to fold away comparisons between allocas and any
1025 // pointer not based on that alloca (e.g. an argument). However, even
1026 // though such pointers cannot alias, they can still compare equal.
1028 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1029 // doesn't escape we can argue that it's impossible to guess its value, and we
1030 // can therefore act as if any such guesses are wrong.
1032 // The code below checks that the alloca doesn't escape, and that it's only
1033 // used in a comparison once (the current instruction). The
1034 // single-comparison-use condition ensures that we're trivially folding all
1035 // comparisons against the alloca consistently, and avoids the risk of
1036 // erroneously folding a comparison of the pointer with itself.
1038 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1040 SmallVector<const Use *, 32> Worklist;
1041 for (const Use &U : Alloca->uses()) {
1042 if (Worklist.size() >= MaxIter)
1044 Worklist.push_back(&U);
1047 unsigned NumCmps = 0;
1048 while (!Worklist.empty()) {
1049 assert(Worklist.size() <= MaxIter);
1050 const Use *U = Worklist.pop_back_val();
1051 const Value *V = U->getUser();
1054 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1055 isa<SelectInst>(V)) {
1057 } else if (isa<LoadInst>(V)) {
1058 // Loading from the pointer doesn't escape it.
1060 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1061 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1062 if (SI->getValueOperand() == U->get())
1065 } else if (isa<ICmpInst>(V)) {
1067 return nullptr; // Found more than one cmp.
1069 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1070 switch (Intrin->getIntrinsicID()) {
1071 // These intrinsics don't escape or compare the pointer. Memset is safe
1072 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1073 // we don't allow stores, so src cannot point to V.
1074 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1075 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1083 for (const Use &U : V->uses()) {
1084 if (Worklist.size() >= MaxIter)
1086 Worklist.push_back(&U);
1090 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1091 return replaceInstUsesWith(
1093 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1096 /// Fold "icmp pred (X+C), X".
1097 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1098 ICmpInst::Predicate Pred) {
1099 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1100 // so the values can never be equal. Similarly for all other "or equals"
1102 assert(!!C && "C should not be zero!");
1104 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1105 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1106 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1107 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1108 Constant *R = ConstantInt::get(X->getType(),
1109 APInt::getMaxValue(C.getBitWidth()) - C);
1110 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1113 // (X+1) >u X --> X <u (0-1) --> X != 255
1114 // (X+2) >u X --> X <u (0-2) --> X <u 254
1115 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1116 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1117 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1118 ConstantInt::get(X->getType(), -C));
1120 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1122 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1123 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1124 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1125 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1126 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1127 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1128 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1129 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1130 ConstantInt::get(X->getType(), SMax - C));
1132 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1133 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1134 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1135 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1136 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1137 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1139 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1140 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1141 ConstantInt::get(X->getType(), SMax - (C - 1)));
1144 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1145 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1146 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1147 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1150 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1152 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1153 if (I.getPredicate() == I.ICMP_NE)
1154 Pred = CmpInst::getInversePredicate(Pred);
1155 return new ICmpInst(Pred, LHS, RHS);
1158 // Don't bother doing any work for cases which InstSimplify handles.
1159 if (AP2.isNullValue())
1162 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1164 if (AP2.isAllOnesValue())
1166 if (AP2.isNegative() != AP1.isNegative())
1173 // 'A' must be large enough to shift out the highest set bit.
1174 return getICmp(I.ICMP_UGT, A,
1175 ConstantInt::get(A->getType(), AP2.logBase2()));
1178 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1181 if (IsAShr && AP1.isNegative())
1182 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1184 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1187 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1188 // There are multiple solutions if we are comparing against -1 and the LHS
1189 // of the ashr is not a power of two.
1190 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1191 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1192 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1193 } else if (AP1 == AP2.lshr(Shift)) {
1194 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1198 // Shifting const2 will never be equal to const1.
1199 // FIXME: This should always be handled by InstSimplify?
1200 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1201 return replaceInstUsesWith(I, TorF);
1204 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1205 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1206 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1209 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1211 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1212 if (I.getPredicate() == I.ICMP_NE)
1213 Pred = CmpInst::getInversePredicate(Pred);
1214 return new ICmpInst(Pred, LHS, RHS);
1217 // Don't bother doing any work for cases which InstSimplify handles.
1218 if (AP2.isNullValue())
1221 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1223 if (!AP1 && AP2TrailingZeros != 0)
1226 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1229 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1231 // Get the distance between the lowest bits that are set.
1232 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1234 if (Shift > 0 && AP2.shl(Shift) == AP1)
1235 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1237 // Shifting const2 will never be equal to const1.
1238 // FIXME: This should always be handled by InstSimplify?
1239 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1240 return replaceInstUsesWith(I, TorF);
1243 /// The caller has matched a pattern of the form:
1244 /// I = icmp ugt (add (add A, B), CI2), CI1
1245 /// If this is of the form:
1247 /// if (sum+128 >u 255)
1248 /// Then replace it with llvm.sadd.with.overflow.i8.
1250 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1251 ConstantInt *CI2, ConstantInt *CI1,
1253 // The transformation we're trying to do here is to transform this into an
1254 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1255 // with a narrower add, and discard the add-with-constant that is part of the
1256 // range check (if we can't eliminate it, this isn't profitable).
1258 // In order to eliminate the add-with-constant, the compare can be its only
1260 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1261 if (!AddWithCst->hasOneUse())
1264 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1265 if (!CI2->getValue().isPowerOf2())
1267 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1268 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1271 // The width of the new add formed is 1 more than the bias.
1274 // Check to see that CI1 is an all-ones value with NewWidth bits.
1275 if (CI1->getBitWidth() == NewWidth ||
1276 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1279 // This is only really a signed overflow check if the inputs have been
1280 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1281 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1282 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1283 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1284 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1287 // In order to replace the original add with a narrower
1288 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1289 // and truncates that discard the high bits of the add. Verify that this is
1291 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1292 for (User *U : OrigAdd->users()) {
1293 if (U == AddWithCst)
1296 // Only accept truncates for now. We would really like a nice recursive
1297 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1298 // chain to see which bits of a value are actually demanded. If the
1299 // original add had another add which was then immediately truncated, we
1300 // could still do the transformation.
1301 TruncInst *TI = dyn_cast<TruncInst>(U);
1302 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1306 // If the pattern matches, truncate the inputs to the narrower type and
1307 // use the sadd_with_overflow intrinsic to efficiently compute both the
1308 // result and the overflow bit.
1309 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1310 Function *F = Intrinsic::getDeclaration(
1311 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1313 InstCombiner::BuilderTy &Builder = IC.Builder;
1315 // Put the new code above the original add, in case there are any uses of the
1316 // add between the add and the compare.
1317 Builder.SetInsertPoint(OrigAdd);
1319 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1320 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1321 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1322 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1323 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1325 // The inner add was the result of the narrow add, zero extended to the
1326 // wider type. Replace it with the result computed by the intrinsic.
1327 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1329 // The original icmp gets replaced with the overflow value.
1330 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1333 // Handle icmp pred X, 0
1334 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1335 CmpInst::Predicate Pred = Cmp.getPredicate();
1336 if (!match(Cmp.getOperand(1), m_Zero()))
1339 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1340 if (Pred == ICmpInst::ICMP_SGT) {
1342 SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1343 if (SPR.Flavor == SPF_SMIN) {
1344 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1345 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1346 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1347 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1352 // icmp eq/ne (urem %x, %y), 0
1353 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1356 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1357 ICmpInst::isEquality(Pred)) {
1358 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1359 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1360 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1361 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1367 /// Fold icmp Pred X, C.
1368 /// TODO: This code structure does not make sense. The saturating add fold
1369 /// should be moved to some other helper and extended as noted below (it is also
1370 /// possible that code has been made unnecessary - do we canonicalize IR to
1371 /// overflow/saturating intrinsics or not?).
1372 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1373 // Match the following pattern, which is a common idiom when writing
1374 // overflow-safe integer arithmetic functions. The source performs an addition
1375 // in wider type and explicitly checks for overflow using comparisons against
1376 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1378 // TODO: This could probably be generalized to handle other overflow-safe
1379 // operations if we worked out the formulas to compute the appropriate magic
1383 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1384 CmpInst::Predicate Pred = Cmp.getPredicate();
1385 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1387 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1388 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1389 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1390 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1396 /// Canonicalize icmp instructions based on dominating conditions.
1397 Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1398 // This is a cheap/incomplete check for dominance - just match a single
1399 // predecessor with a conditional branch.
1400 BasicBlock *CmpBB = Cmp.getParent();
1401 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1406 BasicBlock *TrueBB, *FalseBB;
1407 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1410 assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1411 "Predecessor block does not point to successor?");
1413 // The branch should get simplified. Don't bother simplifying this condition.
1414 if (TrueBB == FalseBB)
1417 // Try to simplify this compare to T/F based on the dominating condition.
1418 Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1420 return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1422 CmpInst::Predicate Pred = Cmp.getPredicate();
1423 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1424 ICmpInst::Predicate DomPred;
1425 const APInt *C, *DomC;
1426 if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1427 match(Y, m_APInt(C))) {
1428 // We have 2 compares of a variable with constants. Calculate the constant
1429 // ranges of those compares to see if we can transform the 2nd compare:
1431 // DomCond = icmp DomPred X, DomC
1432 // br DomCond, CmpBB, FalseBB
1434 // Cmp = icmp Pred X, C
1435 ConstantRange CR = ConstantRange::makeAllowedICmpRegion(Pred, *C);
1436 ConstantRange DominatingCR =
1437 (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1438 : ConstantRange::makeExactICmpRegion(
1439 CmpInst::getInversePredicate(DomPred), *DomC);
1440 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1441 ConstantRange Difference = DominatingCR.difference(CR);
1442 if (Intersection.isEmptySet())
1443 return replaceInstUsesWith(Cmp, Builder.getFalse());
1444 if (Difference.isEmptySet())
1445 return replaceInstUsesWith(Cmp, Builder.getTrue());
1447 // Canonicalizing a sign bit comparison that gets used in a branch,
1448 // pessimizes codegen by generating branch on zero instruction instead
1449 // of a test and branch. So we avoid canonicalizing in such situations
1450 // because test and branch instruction has better branch displacement
1451 // than compare and branch instruction.
1453 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1454 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1457 if (const APInt *EqC = Intersection.getSingleElement())
1458 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1459 if (const APInt *NeC = Difference.getSingleElement())
1460 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1466 /// Fold icmp (trunc X, Y), C.
1467 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1470 ICmpInst::Predicate Pred = Cmp.getPredicate();
1471 Value *X = Trunc->getOperand(0);
1472 if (C.isOneValue() && C.getBitWidth() > 1) {
1473 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1475 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1476 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1477 ConstantInt::get(V->getType(), 1));
1480 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1481 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1482 // of the high bits truncated out of x are known.
1483 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1484 SrcBits = X->getType()->getScalarSizeInBits();
1485 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1487 // If all the high bits are known, we can do this xform.
1488 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1489 // Pull in the high bits from known-ones set.
1490 APInt NewRHS = C.zext(SrcBits);
1491 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1492 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1499 /// Fold icmp (xor X, Y), C.
1500 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1501 BinaryOperator *Xor,
1503 Value *X = Xor->getOperand(0);
1504 Value *Y = Xor->getOperand(1);
1506 if (!match(Y, m_APInt(XorC)))
1509 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1511 ICmpInst::Predicate Pred = Cmp.getPredicate();
1512 bool TrueIfSigned = false;
1513 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1515 // If the sign bit of the XorCst is not set, there is no change to
1516 // the operation, just stop using the Xor.
1517 if (!XorC->isNegative()) {
1518 Cmp.setOperand(0, X);
1523 // Emit the opposite comparison.
1525 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1526 ConstantInt::getAllOnesValue(X->getType()));
1528 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1529 ConstantInt::getNullValue(X->getType()));
1532 if (Xor->hasOneUse()) {
1533 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1534 if (!Cmp.isEquality() && XorC->isSignMask()) {
1535 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1536 : Cmp.getSignedPredicate();
1537 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1540 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1541 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1542 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1543 : Cmp.getSignedPredicate();
1544 Pred = Cmp.getSwappedPredicate(Pred);
1545 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1549 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1550 if (Pred == ICmpInst::ICMP_UGT) {
1551 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1552 if (*XorC == ~C && (C + 1).isPowerOf2())
1553 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1554 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1555 if (*XorC == C && (C + 1).isPowerOf2())
1556 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1558 if (Pred == ICmpInst::ICMP_ULT) {
1559 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1560 if (*XorC == -C && C.isPowerOf2())
1561 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1562 ConstantInt::get(X->getType(), ~C));
1563 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1564 if (*XorC == C && (-C).isPowerOf2())
1565 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1566 ConstantInt::get(X->getType(), ~C));
1571 /// Fold icmp (and (sh X, Y), C2), C1.
1572 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1573 const APInt &C1, const APInt &C2) {
1574 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1575 if (!Shift || !Shift->isShift())
1578 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1579 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1580 // code produced by the clang front-end, for bitfield access.
1581 // This seemingly simple opportunity to fold away a shift turns out to be
1582 // rather complicated. See PR17827 for details.
1583 unsigned ShiftOpcode = Shift->getOpcode();
1584 bool IsShl = ShiftOpcode == Instruction::Shl;
1586 if (match(Shift->getOperand(1), m_APInt(C3))) {
1587 bool CanFold = false;
1588 if (ShiftOpcode == Instruction::Shl) {
1589 // For a left shift, we can fold if the comparison is not signed. We can
1590 // also fold a signed comparison if the mask value and comparison value
1591 // are not negative. These constraints may not be obvious, but we can
1592 // prove that they are correct using an SMT solver.
1593 if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1596 bool IsAshr = ShiftOpcode == Instruction::AShr;
1597 // For a logical right shift, we can fold if the comparison is not signed.
1598 // We can also fold a signed comparison if the shifted mask value and the
1599 // shifted comparison value are not negative. These constraints may not be
1600 // obvious, but we can prove that they are correct using an SMT solver.
1601 // For an arithmetic shift right we can do the same, if we ensure
1602 // the And doesn't use any bits being shifted in. Normally these would
1603 // be turned into lshr by SimplifyDemandedBits, but not if there is an
1605 if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1606 if (!Cmp.isSigned() ||
1607 (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1613 APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1614 APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1615 // Check to see if we are shifting out any of the bits being compared.
1616 if (SameAsC1 != C1) {
1617 // If we shifted bits out, the fold is not going to work out. As a
1618 // special case, check to see if this means that the result is always
1619 // true or false now.
1620 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1621 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1622 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1623 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1625 Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1626 APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1627 And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1628 And->setOperand(0, Shift->getOperand(0));
1629 Worklist.Add(Shift); // Shift is dead.
1635 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1636 // preferable because it allows the C2 << Y expression to be hoisted out of a
1637 // loop if Y is invariant and X is not.
1638 if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1639 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1642 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1643 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1645 // Compute X & (C2 << Y).
1646 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1647 Cmp.setOperand(0, NewAnd);
1654 /// Fold icmp (and X, C2), C1.
1655 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1656 BinaryOperator *And,
1658 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1660 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1661 // TODO: We canonicalize to the longer form for scalars because we have
1662 // better analysis/folds for icmp, and codegen may be better with icmp.
1663 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1664 match(And->getOperand(1), m_One()))
1665 return new TruncInst(And->getOperand(0), Cmp.getType());
1669 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1672 // Don't perform the following transforms if the AND has multiple uses
1673 if (!And->hasOneUse())
1676 if (Cmp.isEquality() && C1.isNullValue()) {
1677 // Restrict this fold to single-use 'and' (PR10267).
1678 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1679 if (C2->isSignMask()) {
1680 Constant *Zero = Constant::getNullValue(X->getType());
1681 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1682 return new ICmpInst(NewPred, X, Zero);
1685 // Restrict this fold only for single-use 'and' (PR10267).
1686 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1687 if ((~(*C2) + 1).isPowerOf2()) {
1689 ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1690 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1691 return new ICmpInst(NewPred, X, NegBOC);
1695 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1696 // the input width without changing the value produced, eliminate the cast:
1698 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1700 // We can do this transformation if the constants do not have their sign bits
1701 // set or if it is an equality comparison. Extending a relational comparison
1702 // when we're checking the sign bit would not work.
1704 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1705 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1706 // TODO: Is this a good transform for vectors? Wider types may reduce
1707 // throughput. Should this transform be limited (even for scalars) by using
1708 // shouldChangeType()?
1709 if (!Cmp.getType()->isVectorTy()) {
1710 Type *WideType = W->getType();
1711 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1712 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1713 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1714 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1715 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1719 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1722 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1723 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1725 // iff pred isn't signed
1726 if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1727 match(And->getOperand(1), m_One())) {
1728 Constant *One = cast<Constant>(And->getOperand(1));
1729 Value *Or = And->getOperand(0);
1730 Value *A, *B, *LShr;
1731 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1732 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1733 unsigned UsesRemoved = 0;
1734 if (And->hasOneUse())
1736 if (Or->hasOneUse())
1738 if (LShr->hasOneUse())
1741 // Compute A & ((1 << B) | 1)
1742 Value *NewOr = nullptr;
1743 if (auto *C = dyn_cast<Constant>(B)) {
1744 if (UsesRemoved >= 1)
1745 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1747 if (UsesRemoved >= 3)
1748 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1750 One, Or->getName());
1753 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1754 Cmp.setOperand(0, NewAnd);
1763 /// Fold icmp (and X, Y), C.
1764 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1765 BinaryOperator *And,
1767 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1770 // TODO: These all require that Y is constant too, so refactor with the above.
1772 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1773 Value *X = And->getOperand(0);
1774 Value *Y = And->getOperand(1);
1775 if (auto *LI = dyn_cast<LoadInst>(X))
1776 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1777 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1778 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1779 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1780 ConstantInt *C2 = cast<ConstantInt>(Y);
1781 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1785 if (!Cmp.isEquality())
1788 // X & -C == -C -> X > u ~C
1789 // X & -C != -C -> X <= u ~C
1790 // iff C is a power of 2
1791 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1792 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1793 : CmpInst::ICMP_ULE;
1794 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1797 // (X & C2) == 0 -> (trunc X) >= 0
1798 // (X & C2) != 0 -> (trunc X) < 0
1799 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1801 if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1802 int32_t ExactLogBase2 = C2->exactLogBase2();
1803 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1804 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1805 if (And->getType()->isVectorTy())
1806 NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1807 Value *Trunc = Builder.CreateTrunc(X, NTy);
1808 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1809 : CmpInst::ICMP_SLT;
1810 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1817 /// Fold icmp (or X, Y), C.
1818 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1820 ICmpInst::Predicate Pred = Cmp.getPredicate();
1821 if (C.isOneValue()) {
1822 // icmp slt signum(V) 1 --> icmp slt V, 1
1824 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1825 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1826 ConstantInt::get(V->getType(), 1));
1829 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1830 if (Cmp.isEquality() && Cmp.getOperand(1) == OrOp1) {
1831 // X | C == C --> X <=u C
1832 // X | C != C --> X >u C
1833 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1834 if ((C + 1).isPowerOf2()) {
1835 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1836 return new ICmpInst(Pred, OrOp0, OrOp1);
1838 // More general: are all bits outside of a mask constant set or not set?
1839 // X | C == C --> (X & ~C) == 0
1840 // X | C != C --> (X & ~C) != 0
1841 if (Or->hasOneUse()) {
1842 Value *A = Builder.CreateAnd(OrOp0, ~C);
1843 return new ICmpInst(Pred, A, ConstantInt::getNullValue(OrOp0->getType()));
1847 if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1851 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1852 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1853 // -> and (icmp eq P, null), (icmp eq Q, null).
1855 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1857 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1858 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1859 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1862 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1863 // a shorter form that has more potential to be folded even further.
1864 Value *X1, *X2, *X3, *X4;
1865 if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1866 match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1867 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1868 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1869 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1870 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1871 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1872 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1878 /// Fold icmp (mul X, Y), C.
1879 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1880 BinaryOperator *Mul,
1883 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1886 // If this is a test of the sign bit and the multiply is sign-preserving with
1887 // a constant operand, use the multiply LHS operand instead.
1888 ICmpInst::Predicate Pred = Cmp.getPredicate();
1889 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1890 if (MulC->isNegative())
1891 Pred = ICmpInst::getSwappedPredicate(Pred);
1892 return new ICmpInst(Pred, Mul->getOperand(0),
1893 Constant::getNullValue(Mul->getType()));
1899 /// Fold icmp (shl 1, Y), C.
1900 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
1903 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1906 Type *ShiftType = Shl->getType();
1907 unsigned TypeBits = C.getBitWidth();
1908 bool CIsPowerOf2 = C.isPowerOf2();
1909 ICmpInst::Predicate Pred = Cmp.getPredicate();
1910 if (Cmp.isUnsigned()) {
1911 // (1 << Y) pred C -> Y pred Log2(C)
1913 // (1 << Y) < 30 -> Y <= 4
1914 // (1 << Y) <= 30 -> Y <= 4
1915 // (1 << Y) >= 30 -> Y > 4
1916 // (1 << Y) > 30 -> Y > 4
1917 if (Pred == ICmpInst::ICMP_ULT)
1918 Pred = ICmpInst::ICMP_ULE;
1919 else if (Pred == ICmpInst::ICMP_UGE)
1920 Pred = ICmpInst::ICMP_UGT;
1923 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1924 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1925 unsigned CLog2 = C.logBase2();
1926 if (CLog2 == TypeBits - 1) {
1927 if (Pred == ICmpInst::ICMP_UGE)
1928 Pred = ICmpInst::ICMP_EQ;
1929 else if (Pred == ICmpInst::ICMP_ULT)
1930 Pred = ICmpInst::ICMP_NE;
1932 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1933 } else if (Cmp.isSigned()) {
1934 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1935 if (C.isAllOnesValue()) {
1936 // (1 << Y) <= -1 -> Y == 31
1937 if (Pred == ICmpInst::ICMP_SLE)
1938 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1940 // (1 << Y) > -1 -> Y != 31
1941 if (Pred == ICmpInst::ICMP_SGT)
1942 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1944 // (1 << Y) < 0 -> Y == 31
1945 // (1 << Y) <= 0 -> Y == 31
1946 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1947 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1949 // (1 << Y) >= 0 -> Y != 31
1950 // (1 << Y) > 0 -> Y != 31
1951 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1952 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1954 } else if (Cmp.isEquality() && CIsPowerOf2) {
1955 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
1961 /// Fold icmp (shl X, Y), C.
1962 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1963 BinaryOperator *Shl,
1965 const APInt *ShiftVal;
1966 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1967 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
1969 const APInt *ShiftAmt;
1970 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1971 return foldICmpShlOne(Cmp, Shl, C);
1973 // Check that the shift amount is in range. If not, don't perform undefined
1974 // shifts. When the shift is visited, it will be simplified.
1975 unsigned TypeBits = C.getBitWidth();
1976 if (ShiftAmt->uge(TypeBits))
1979 ICmpInst::Predicate Pred = Cmp.getPredicate();
1980 Value *X = Shl->getOperand(0);
1981 Type *ShType = Shl->getType();
1983 // NSW guarantees that we are only shifting out sign bits from the high bits,
1984 // so we can ASHR the compare constant without needing a mask and eliminate
1986 if (Shl->hasNoSignedWrap()) {
1987 if (Pred == ICmpInst::ICMP_SGT) {
1988 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
1989 APInt ShiftedC = C.ashr(*ShiftAmt);
1990 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1992 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1993 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
1994 APInt ShiftedC = C.ashr(*ShiftAmt);
1995 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1997 if (Pred == ICmpInst::ICMP_SLT) {
1998 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
1999 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2000 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2001 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2002 assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2003 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2004 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2006 // If this is a signed comparison to 0 and the shift is sign preserving,
2007 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2008 // do that if we're sure to not continue on in this function.
2009 if (isSignTest(Pred, C))
2010 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2013 // NUW guarantees that we are only shifting out zero bits from the high bits,
2014 // so we can LSHR the compare constant without needing a mask and eliminate
2016 if (Shl->hasNoUnsignedWrap()) {
2017 if (Pred == ICmpInst::ICMP_UGT) {
2018 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2019 APInt ShiftedC = C.lshr(*ShiftAmt);
2020 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2022 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2023 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2024 APInt ShiftedC = C.lshr(*ShiftAmt);
2025 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2027 if (Pred == ICmpInst::ICMP_ULT) {
2028 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2029 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2030 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2031 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2032 assert(C.ugt(0) && "ult 0 should have been eliminated");
2033 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2034 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2038 if (Cmp.isEquality() && Shl->hasOneUse()) {
2039 // Strength-reduce the shift into an 'and'.
2040 Constant *Mask = ConstantInt::get(
2042 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2043 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2044 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2045 return new ICmpInst(Pred, And, LShrC);
2048 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2049 bool TrueIfSigned = false;
2050 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2051 // (X << 31) <s 0 --> (X & 1) != 0
2052 Constant *Mask = ConstantInt::get(
2054 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2055 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2056 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2057 And, Constant::getNullValue(ShType));
2060 // Simplify 'shl' inequality test into 'and' equality test.
2061 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2062 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2063 if ((C + 1).isPowerOf2() &&
2064 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2065 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2066 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2067 : ICmpInst::ICMP_NE,
2068 And, Constant::getNullValue(ShType));
2070 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2071 if (C.isPowerOf2() &&
2072 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2074 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2075 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2076 : ICmpInst::ICMP_NE,
2077 And, Constant::getNullValue(ShType));
2081 // Transform (icmp pred iM (shl iM %v, N), C)
2082 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2083 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2084 // This enables us to get rid of the shift in favor of a trunc that may be
2085 // free on the target. It has the additional benefit of comparing to a
2086 // smaller constant that may be more target-friendly.
2087 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2088 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2089 DL.isLegalInteger(TypeBits - Amt)) {
2090 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2091 if (ShType->isVectorTy())
2092 TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2094 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2095 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2101 /// Fold icmp ({al}shr X, Y), C.
2102 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2103 BinaryOperator *Shr,
2105 // An exact shr only shifts out zero bits, so:
2106 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2107 Value *X = Shr->getOperand(0);
2108 CmpInst::Predicate Pred = Cmp.getPredicate();
2109 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2111 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2113 const APInt *ShiftVal;
2114 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2115 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2117 const APInt *ShiftAmt;
2118 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2121 // Check that the shift amount is in range. If not, don't perform undefined
2122 // shifts. When the shift is visited it will be simplified.
2123 unsigned TypeBits = C.getBitWidth();
2124 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2125 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2128 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2129 bool IsExact = Shr->isExact();
2130 Type *ShrTy = Shr->getType();
2131 // TODO: If we could guarantee that InstSimplify would handle all of the
2132 // constant-value-based preconditions in the folds below, then we could assert
2133 // those conditions rather than checking them. This is difficult because of
2134 // undef/poison (PR34838).
2136 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2137 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2138 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2139 APInt ShiftedC = C.shl(ShAmtVal);
2140 if (ShiftedC.ashr(ShAmtVal) == C)
2141 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2143 if (Pred == CmpInst::ICMP_SGT) {
2144 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2145 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2146 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2147 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2148 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2151 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2152 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2153 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2154 APInt ShiftedC = C.shl(ShAmtVal);
2155 if (ShiftedC.lshr(ShAmtVal) == C)
2156 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2158 if (Pred == CmpInst::ICMP_UGT) {
2159 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2160 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2161 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2162 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2166 if (!Cmp.isEquality())
2169 // Handle equality comparisons of shift-by-constant.
2171 // If the comparison constant changes with the shift, the comparison cannot
2172 // succeed (bits of the comparison constant cannot match the shifted value).
2173 // This should be known by InstSimplify and already be folded to true/false.
2174 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2175 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2176 "Expected icmp+shr simplify did not occur.");
2178 // If the bits shifted out are known zero, compare the unshifted value:
2179 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2181 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2183 if (Shr->hasOneUse()) {
2184 // Canonicalize the shift into an 'and':
2185 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2186 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2187 Constant *Mask = ConstantInt::get(ShrTy, Val);
2188 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2189 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2195 /// Fold icmp (udiv X, Y), C.
2196 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2197 BinaryOperator *UDiv,
2200 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2203 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2205 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2206 Value *Y = UDiv->getOperand(1);
2207 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2208 assert(!C.isMaxValue() &&
2209 "icmp ugt X, UINT_MAX should have been simplified already.");
2210 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2211 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2214 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2215 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2216 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2217 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2218 ConstantInt::get(Y->getType(), C2->udiv(C)));
2224 /// Fold icmp ({su}div X, Y), C.
2225 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2226 BinaryOperator *Div,
2228 // Fold: icmp pred ([us]div X, C2), C -> range test
2229 // Fold this div into the comparison, producing a range check.
2230 // Determine, based on the divide type, what the range is being
2231 // checked. If there is an overflow on the low or high side, remember
2232 // it, otherwise compute the range [low, hi) bounding the new value.
2233 // See: InsertRangeTest above for the kinds of replacements possible.
2235 if (!match(Div->getOperand(1), m_APInt(C2)))
2238 // FIXME: If the operand types don't match the type of the divide
2239 // then don't attempt this transform. The code below doesn't have the
2240 // logic to deal with a signed divide and an unsigned compare (and
2241 // vice versa). This is because (x /s C2) <s C produces different
2242 // results than (x /s C2) <u C or (x /u C2) <s C or even
2243 // (x /u C2) <u C. Simply casting the operands and result won't
2244 // work. :( The if statement below tests that condition and bails
2246 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2247 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2250 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2251 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2252 // division-by-constant cases should be present, we can not assert that they
2253 // have happened before we reach this icmp instruction.
2254 if (C2->isNullValue() || C2->isOneValue() ||
2255 (DivIsSigned && C2->isAllOnesValue()))
2258 // Compute Prod = C * C2. We are essentially solving an equation of
2259 // form X / C2 = C. We solve for X by multiplying C2 and C.
2260 // By solving for X, we can turn this into a range check instead of computing
2262 APInt Prod = C * *C2;
2264 // Determine if the product overflows by seeing if the product is not equal to
2265 // the divide. Make sure we do the same kind of divide as in the LHS
2266 // instruction that we're folding.
2267 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2269 ICmpInst::Predicate Pred = Cmp.getPredicate();
2271 // If the division is known to be exact, then there is no remainder from the
2272 // divide, so the covered range size is unit, otherwise it is the divisor.
2273 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2275 // Figure out the interval that is being checked. For example, a comparison
2276 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2277 // Compute this interval based on the constants involved and the signedness of
2278 // the compare/divide. This computes a half-open interval, keeping track of
2279 // whether either value in the interval overflows. After analysis each
2280 // overflow variable is set to 0 if it's corresponding bound variable is valid
2281 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2282 int LoOverflow = 0, HiOverflow = 0;
2283 APInt LoBound, HiBound;
2285 if (!DivIsSigned) { // udiv
2286 // e.g. X/5 op 3 --> [15, 20)
2288 HiOverflow = LoOverflow = ProdOV;
2290 // If this is not an exact divide, then many values in the range collapse
2291 // to the same result value.
2292 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2294 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2295 if (C.isNullValue()) { // (X / pos) op 0
2296 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2297 LoBound = -(RangeSize - 1);
2298 HiBound = RangeSize;
2299 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2300 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2301 HiOverflow = LoOverflow = ProdOV;
2303 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2304 } else { // (X / pos) op neg
2305 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2307 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2309 APInt DivNeg = -RangeSize;
2310 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2313 } else if (C2->isNegative()) { // Divisor is < 0.
2316 if (C.isNullValue()) { // (X / neg) op 0
2317 // e.g. X/-5 op 0 --> [-4, 5)
2318 LoBound = RangeSize + 1;
2319 HiBound = -RangeSize;
2320 if (HiBound == *C2) { // -INTMIN = INTMIN
2321 HiOverflow = 1; // [INTMIN+1, overflow)
2322 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2324 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2325 // e.g. X/-5 op 3 --> [-19, -14)
2327 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2329 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2330 } else { // (X / neg) op neg
2331 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2332 LoOverflow = HiOverflow = ProdOV;
2334 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2337 // Dividing by a negative swaps the condition. LT <-> GT
2338 Pred = ICmpInst::getSwappedPredicate(Pred);
2341 Value *X = Div->getOperand(0);
2343 default: llvm_unreachable("Unhandled icmp opcode!");
2344 case ICmpInst::ICMP_EQ:
2345 if (LoOverflow && HiOverflow)
2346 return replaceInstUsesWith(Cmp, Builder.getFalse());
2348 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2349 ICmpInst::ICMP_UGE, X,
2350 ConstantInt::get(Div->getType(), LoBound));
2352 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2353 ICmpInst::ICMP_ULT, X,
2354 ConstantInt::get(Div->getType(), HiBound));
2355 return replaceInstUsesWith(
2356 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2357 case ICmpInst::ICMP_NE:
2358 if (LoOverflow && HiOverflow)
2359 return replaceInstUsesWith(Cmp, Builder.getTrue());
2361 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2362 ICmpInst::ICMP_ULT, X,
2363 ConstantInt::get(Div->getType(), LoBound));
2365 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2366 ICmpInst::ICMP_UGE, X,
2367 ConstantInt::get(Div->getType(), HiBound));
2368 return replaceInstUsesWith(Cmp,
2369 insertRangeTest(X, LoBound, HiBound,
2370 DivIsSigned, false));
2371 case ICmpInst::ICMP_ULT:
2372 case ICmpInst::ICMP_SLT:
2373 if (LoOverflow == +1) // Low bound is greater than input range.
2374 return replaceInstUsesWith(Cmp, Builder.getTrue());
2375 if (LoOverflow == -1) // Low bound is less than input range.
2376 return replaceInstUsesWith(Cmp, Builder.getFalse());
2377 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2378 case ICmpInst::ICMP_UGT:
2379 case ICmpInst::ICMP_SGT:
2380 if (HiOverflow == +1) // High bound greater than input range.
2381 return replaceInstUsesWith(Cmp, Builder.getFalse());
2382 if (HiOverflow == -1) // High bound less than input range.
2383 return replaceInstUsesWith(Cmp, Builder.getTrue());
2384 if (Pred == ICmpInst::ICMP_UGT)
2385 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2386 ConstantInt::get(Div->getType(), HiBound));
2387 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2388 ConstantInt::get(Div->getType(), HiBound));
2394 /// Fold icmp (sub X, Y), C.
2395 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2396 BinaryOperator *Sub,
2398 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2399 ICmpInst::Predicate Pred = Cmp.getPredicate();
2403 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2404 if (match(X, m_APInt(C2)) &&
2405 ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2406 (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2407 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2408 return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2409 ConstantInt::get(Y->getType(), SubResult));
2411 // The following transforms are only worth it if the only user of the subtract
2413 if (!Sub->hasOneUse())
2416 if (Sub->hasNoSignedWrap()) {
2417 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2418 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2419 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2421 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2422 if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2423 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2425 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2426 if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2427 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2429 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2430 if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2431 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2434 if (!match(X, m_APInt(C2)))
2437 // C2 - Y <u C -> (Y | (C - 1)) == C2
2438 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2439 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2440 (*C2 & (C - 1)) == (C - 1))
2441 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2443 // C2 - Y >u C -> (Y | C) != C2
2444 // iff C2 & C == C and C + 1 is a power of 2
2445 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2446 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2451 /// Fold icmp (add X, Y), C.
2452 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2453 BinaryOperator *Add,
2455 Value *Y = Add->getOperand(1);
2457 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2460 // Fold icmp pred (add X, C2), C.
2461 Value *X = Add->getOperand(0);
2462 Type *Ty = Add->getType();
2463 CmpInst::Predicate Pred = Cmp.getPredicate();
2465 if (!Add->hasOneUse())
2468 // If the add does not wrap, we can always adjust the compare by subtracting
2469 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2470 // are canonicalized to SGT/SLT/UGT/ULT.
2471 if ((Add->hasNoSignedWrap() &&
2472 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2473 (Add->hasNoUnsignedWrap() &&
2474 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2477 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2478 // If there is overflow, the result must be true or false.
2479 // TODO: Can we assert there is no overflow because InstSimplify always
2480 // handles those cases?
2482 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2483 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2486 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2487 const APInt &Upper = CR.getUpper();
2488 const APInt &Lower = CR.getLower();
2489 if (Cmp.isSigned()) {
2490 if (Lower.isSignMask())
2491 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2492 if (Upper.isSignMask())
2493 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2495 if (Lower.isMinValue())
2496 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2497 if (Upper.isMinValue())
2498 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2501 // X+C <u C2 -> (X & -C2) == C
2502 // iff C & (C2-1) == 0
2503 // C2 is a power of 2
2504 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2505 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2506 ConstantExpr::getNeg(cast<Constant>(Y)));
2508 // X+C >u C2 -> (X & ~C2) != C
2510 // C2+1 is a power of 2
2511 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2512 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2513 ConstantExpr::getNeg(cast<Constant>(Y)));
2518 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2519 Value *&RHS, ConstantInt *&Less,
2520 ConstantInt *&Equal,
2521 ConstantInt *&Greater) {
2522 // TODO: Generalize this to work with other comparison idioms or ensure
2523 // they get canonicalized into this form.
2525 // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32
2526 // Greater), where Equal, Less and Greater are placeholders for any three
2528 ICmpInst::Predicate PredA, PredB;
2529 if (match(SI->getTrueValue(), m_ConstantInt(Equal)) &&
2530 match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) &&
2531 PredA == ICmpInst::ICMP_EQ &&
2532 match(SI->getFalseValue(),
2533 m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)),
2534 m_ConstantInt(Less), m_ConstantInt(Greater))) &&
2535 PredB == ICmpInst::ICMP_SLT) {
2541 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2545 assert(C && "Cmp RHS should be a constant int!");
2546 // If we're testing a constant value against the result of a three way
2547 // comparison, the result can be expressed directly in terms of the
2548 // original values being compared. Note: We could possibly be more
2549 // aggressive here and remove the hasOneUse test. The original select is
2550 // really likely to simplify or sink when we remove a test of the result.
2551 Value *OrigLHS, *OrigRHS;
2552 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2553 if (Cmp.hasOneUse() &&
2554 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2556 assert(C1LessThan && C2Equal && C3GreaterThan);
2558 bool TrueWhenLessThan =
2559 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2561 bool TrueWhenEqual =
2562 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2564 bool TrueWhenGreaterThan =
2565 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2568 // This generates the new instruction that will replace the original Cmp
2569 // Instruction. Instead of enumerating the various combinations when
2570 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2571 // false, we rely on chaining of ORs and future passes of InstCombine to
2572 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2574 // When none of the three constants satisfy the predicate for the RHS (C),
2575 // the entire original Cmp can be simplified to a false.
2576 Value *Cond = Builder.getFalse();
2577 if (TrueWhenLessThan)
2578 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2581 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2583 if (TrueWhenGreaterThan)
2584 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2587 return replaceInstUsesWith(Cmp, Cond);
2592 static Instruction *foldICmpBitCast(ICmpInst &Cmp,
2593 InstCombiner::BuilderTy &Builder) {
2594 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2598 ICmpInst::Predicate Pred = Cmp.getPredicate();
2599 Value *Op1 = Cmp.getOperand(1);
2600 Value *BCSrcOp = Bitcast->getOperand(0);
2602 // Make sure the bitcast doesn't change the number of vector elements.
2603 if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2604 Bitcast->getDestTy()->getScalarSizeInBits()) {
2605 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2607 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2608 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2609 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2610 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2611 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2612 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2613 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2614 match(Op1, m_Zero()))
2615 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2617 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2618 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2619 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2621 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2622 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2623 return new ICmpInst(Pred, X,
2624 ConstantInt::getAllOnesValue(X->getType()));
2627 // Zero-equality checks are preserved through unsigned floating-point casts:
2628 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2629 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2630 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2631 if (Cmp.isEquality() && match(Op1, m_Zero()))
2632 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2635 // Test to see if the operands of the icmp are casted versions of other
2636 // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2637 if (Bitcast->getType()->isPointerTy() &&
2638 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2639 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2640 // so eliminate it as well.
2641 if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2642 Op1 = BC2->getOperand(0);
2644 Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2645 return new ICmpInst(Pred, BCSrcOp, Op1);
2648 // Folding: icmp <pred> iN X, C
2649 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2650 // and C is a splat of a K-bit pattern
2651 // and SC is a constant vector = <C', C', C', ..., C'>
2653 // %E = extractelement <M x iK> %vec, i32 C'
2654 // icmp <pred> iK %E, trunc(C)
2656 if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2657 !Bitcast->getType()->isIntegerTy() ||
2658 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2664 m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2665 // Check whether every element of Mask is the same constant
2666 if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2667 auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2668 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2669 if (C->isSplat(EltTy->getBitWidth())) {
2670 // Fold the icmp based on the value of C
2671 // If C is M copies of an iK sized bit pattern,
2673 // => %E = extractelement <N x iK> %vec, i32 Elem
2674 // icmp <pred> iK %SplatVal, <pattern>
2675 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2676 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2677 return new ICmpInst(Pred, Extract, NewC);
2684 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2685 /// where X is some kind of instruction.
2686 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2688 if (!match(Cmp.getOperand(1), m_APInt(C)))
2691 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2692 switch (BO->getOpcode()) {
2693 case Instruction::Xor:
2694 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2697 case Instruction::And:
2698 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2701 case Instruction::Or:
2702 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2705 case Instruction::Mul:
2706 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2709 case Instruction::Shl:
2710 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2713 case Instruction::LShr:
2714 case Instruction::AShr:
2715 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2718 case Instruction::UDiv:
2719 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2722 case Instruction::SDiv:
2723 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2726 case Instruction::Sub:
2727 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2730 case Instruction::Add:
2731 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2737 // TODO: These folds could be refactored to be part of the above calls.
2738 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2742 // Match against CmpInst LHS being instructions other than binary operators.
2744 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2745 // For now, we only support constant integers while folding the
2746 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2747 // similar to the cases handled by binary ops above.
2748 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2749 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2753 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2754 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2758 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2759 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2765 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2766 /// icmp eq/ne BO, C.
2767 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2770 // TODO: Some of these folds could work with arbitrary constants, but this
2771 // function is limited to scalar and vector splat constants.
2772 if (!Cmp.isEquality())
2775 ICmpInst::Predicate Pred = Cmp.getPredicate();
2776 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2777 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2778 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2780 switch (BO->getOpcode()) {
2781 case Instruction::SRem:
2782 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2783 if (C.isNullValue() && BO->hasOneUse()) {
2785 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2786 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2787 return new ICmpInst(Pred, NewRem,
2788 Constant::getNullValue(BO->getType()));
2792 case Instruction::Add: {
2793 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2795 if (match(BOp1, m_APInt(BOC))) {
2796 if (BO->hasOneUse()) {
2797 Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2798 return new ICmpInst(Pred, BOp0, SubC);
2800 } else if (C.isNullValue()) {
2801 // Replace ((add A, B) != 0) with (A != -B) if A or B is
2802 // efficiently invertible, or if the add has just this one use.
2803 if (Value *NegVal = dyn_castNegVal(BOp1))
2804 return new ICmpInst(Pred, BOp0, NegVal);
2805 if (Value *NegVal = dyn_castNegVal(BOp0))
2806 return new ICmpInst(Pred, NegVal, BOp1);
2807 if (BO->hasOneUse()) {
2808 Value *Neg = Builder.CreateNeg(BOp1);
2810 return new ICmpInst(Pred, BOp0, Neg);
2815 case Instruction::Xor:
2816 if (BO->hasOneUse()) {
2817 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2818 // For the xor case, we can xor two constants together, eliminating
2819 // the explicit xor.
2820 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2821 } else if (C.isNullValue()) {
2822 // Replace ((xor A, B) != 0) with (A != B)
2823 return new ICmpInst(Pred, BOp0, BOp1);
2827 case Instruction::Sub:
2828 if (BO->hasOneUse()) {
2830 if (match(BOp0, m_APInt(BOC))) {
2831 // Replace ((sub BOC, B) != C) with (B != BOC-C).
2832 Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2833 return new ICmpInst(Pred, BOp1, SubC);
2834 } else if (C.isNullValue()) {
2835 // Replace ((sub A, B) != 0) with (A != B).
2836 return new ICmpInst(Pred, BOp0, BOp1);
2840 case Instruction::Or: {
2842 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2843 // Comparing if all bits outside of a constant mask are set?
2844 // Replace (X | C) == -1 with (X & ~C) == ~C.
2845 // This removes the -1 constant.
2846 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2847 Value *And = Builder.CreateAnd(BOp0, NotBOC);
2848 return new ICmpInst(Pred, And, NotBOC);
2852 case Instruction::And: {
2854 if (match(BOp1, m_APInt(BOC))) {
2855 // If we have ((X & C) == C), turn it into ((X & C) != 0).
2856 if (C == *BOC && C.isPowerOf2())
2857 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2858 BO, Constant::getNullValue(RHS->getType()));
2862 case Instruction::Mul:
2863 if (C.isNullValue() && BO->hasNoSignedWrap()) {
2865 if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2866 // The trivial case (mul X, 0) is handled by InstSimplify.
2867 // General case : (mul X, C) != 0 iff X != 0
2868 // (mul X, C) == 0 iff X == 0
2869 return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2873 case Instruction::UDiv:
2874 if (C.isNullValue()) {
2875 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2876 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2877 return new ICmpInst(NewPred, BOp1, BOp0);
2886 /// Fold an equality icmp with LLVM intrinsic and constant operand.
2887 Instruction *InstCombiner::foldICmpEqIntrinsicWithConstant(ICmpInst &Cmp,
2890 Type *Ty = II->getType();
2891 unsigned BitWidth = C.getBitWidth();
2892 switch (II->getIntrinsicID()) {
2893 case Intrinsic::bswap:
2895 Cmp.setOperand(0, II->getArgOperand(0));
2896 Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
2899 case Intrinsic::ctlz:
2900 case Intrinsic::cttz: {
2901 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
2902 if (C == BitWidth) {
2904 Cmp.setOperand(0, II->getArgOperand(0));
2905 Cmp.setOperand(1, ConstantInt::getNullValue(Ty));
2909 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
2910 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
2911 // Limit to one use to ensure we don't increase instruction count.
2912 unsigned Num = C.getLimitedValue(BitWidth);
2913 if (Num != BitWidth && II->hasOneUse()) {
2914 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
2915 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
2916 : APInt::getHighBitsSet(BitWidth, Num + 1);
2917 APInt Mask2 = IsTrailing
2918 ? APInt::getOneBitSet(BitWidth, Num)
2919 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
2920 Cmp.setOperand(0, Builder.CreateAnd(II->getArgOperand(0), Mask1));
2921 Cmp.setOperand(1, ConstantInt::get(Ty, Mask2));
2928 case Intrinsic::ctpop: {
2929 // popcount(A) == 0 -> A == 0 and likewise for !=
2930 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
2931 bool IsZero = C.isNullValue();
2932 if (IsZero || C == BitWidth) {
2934 Cmp.setOperand(0, II->getArgOperand(0));
2936 IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty);
2937 Cmp.setOperand(1, NewOp);
2949 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
2950 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2953 if (Cmp.isEquality())
2954 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
2956 Type *Ty = II->getType();
2957 unsigned BitWidth = C.getBitWidth();
2958 switch (II->getIntrinsicID()) {
2959 case Intrinsic::ctlz: {
2960 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
2961 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
2962 unsigned Num = C.getLimitedValue();
2963 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
2964 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
2965 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
2968 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
2969 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
2970 C.uge(1) && C.ule(BitWidth)) {
2971 unsigned Num = C.getLimitedValue();
2972 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
2973 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
2974 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
2978 case Intrinsic::cttz: {
2979 // Limit to one use to ensure we don't increase instruction count.
2980 if (!II->hasOneUse())
2983 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
2984 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
2985 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
2986 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
2987 Builder.CreateAnd(II->getArgOperand(0), Mask),
2988 ConstantInt::getNullValue(Ty));
2991 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
2992 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
2993 C.uge(1) && C.ule(BitWidth)) {
2994 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
2995 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
2996 Builder.CreateAnd(II->getArgOperand(0), Mask),
2997 ConstantInt::getNullValue(Ty));
3008 /// Handle icmp with constant (but not simple integer constant) RHS.
3009 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3010 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3011 Constant *RHSC = dyn_cast<Constant>(Op1);
3012 Instruction *LHSI = dyn_cast<Instruction>(Op0);
3016 switch (LHSI->getOpcode()) {
3017 case Instruction::GetElementPtr:
3018 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3019 if (RHSC->isNullValue() &&
3020 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3021 return new ICmpInst(
3022 I.getPredicate(), LHSI->getOperand(0),
3023 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3025 case Instruction::PHI:
3026 // Only fold icmp into the PHI if the phi and icmp are in the same
3027 // block. If in the same block, we're encouraging jump threading. If
3028 // not, we are just pessimizing the code by making an i1 phi.
3029 if (LHSI->getParent() == I.getParent())
3030 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3033 case Instruction::Select: {
3034 // If either operand of the select is a constant, we can fold the
3035 // comparison into the select arms, which will cause one to be
3036 // constant folded and the select turned into a bitwise or.
3037 Value *Op1 = nullptr, *Op2 = nullptr;
3038 ConstantInt *CI = nullptr;
3039 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3040 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3041 CI = dyn_cast<ConstantInt>(Op1);
3043 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3044 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3045 CI = dyn_cast<ConstantInt>(Op2);
3048 // We only want to perform this transformation if it will not lead to
3049 // additional code. This is true if either both sides of the select
3050 // fold to a constant (in which case the icmp is replaced with a select
3051 // which will usually simplify) or this is the only user of the
3052 // select (in which case we are trading a select+icmp for a simpler
3053 // select+icmp) or all uses of the select can be replaced based on
3054 // dominance information ("Global cases").
3055 bool Transform = false;
3058 else if (Op1 || Op2) {
3060 if (LHSI->hasOneUse())
3063 else if (CI && !CI->isZero())
3064 // When Op1 is constant try replacing select with second operand.
3065 // Otherwise Op2 is constant and try replacing select with first
3068 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3072 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3075 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3077 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3081 case Instruction::IntToPtr:
3082 // icmp pred inttoptr(X), null -> icmp pred X, 0
3083 if (RHSC->isNullValue() &&
3084 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3085 return new ICmpInst(
3086 I.getPredicate(), LHSI->getOperand(0),
3087 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3090 case Instruction::Load:
3091 // Try to optimize things like "A[i] > 4" to index computations.
3092 if (GetElementPtrInst *GEP =
3093 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3094 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3095 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3096 !cast<LoadInst>(LHSI)->isVolatile())
3097 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3106 /// Some comparisons can be simplified.
3107 /// In this case, we are looking for comparisons that look like
3108 /// a check for a lossy truncation.
3110 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3111 /// Where Mask is some pattern that produces all-ones in low bits:
3113 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3115 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3116 /// The Mask can be a constant, too.
3117 /// For some predicates, the operands are commutative.
3118 /// For others, x can only be on a specific side.
3119 static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
3120 InstCombiner::BuilderTy &Builder) {
3121 ICmpInst::Predicate SrcPred;
3123 auto m_VariableMask = m_CombineOr(
3124 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
3125 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3126 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
3127 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3128 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3129 if (!match(&I, m_c_ICmp(SrcPred,
3130 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3134 ICmpInst::Predicate DstPred;
3136 case ICmpInst::Predicate::ICMP_EQ:
3137 // x & (-1 >> y) == x -> x u<= (-1 >> y)
3138 DstPred = ICmpInst::Predicate::ICMP_ULE;
3140 case ICmpInst::Predicate::ICMP_NE:
3141 // x & (-1 >> y) != x -> x u> (-1 >> y)
3142 DstPred = ICmpInst::Predicate::ICMP_UGT;
3144 case ICmpInst::Predicate::ICMP_UGT:
3145 // x u> x & (-1 >> y) -> x u> (-1 >> y)
3146 assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3147 DstPred = ICmpInst::Predicate::ICMP_UGT;
3149 case ICmpInst::Predicate::ICMP_UGE:
3150 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3151 assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3152 DstPred = ICmpInst::Predicate::ICMP_ULE;
3154 case ICmpInst::Predicate::ICMP_ULT:
3155 // x & (-1 >> y) u< x -> x u> (-1 >> y)
3156 assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3157 DstPred = ICmpInst::Predicate::ICMP_UGT;
3159 case ICmpInst::Predicate::ICMP_ULE:
3160 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3161 assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3162 DstPred = ICmpInst::Predicate::ICMP_ULE;
3164 case ICmpInst::Predicate::ICMP_SGT:
3165 // x s> x & (-1 >> y) -> x s> (-1 >> y)
3166 if (X != I.getOperand(0)) // X must be on LHS of comparison!
3167 return nullptr; // Ignore the other case.
3168 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3170 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3172 DstPred = ICmpInst::Predicate::ICMP_SGT;
3174 case ICmpInst::Predicate::ICMP_SGE:
3175 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3176 if (X != I.getOperand(1)) // X must be on RHS of comparison!
3177 return nullptr; // Ignore the other case.
3178 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3180 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3182 DstPred = ICmpInst::Predicate::ICMP_SLE;
3184 case ICmpInst::Predicate::ICMP_SLT:
3185 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3186 if (X != I.getOperand(1)) // X must be on RHS of comparison!
3187 return nullptr; // Ignore the other case.
3188 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3190 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3192 DstPred = ICmpInst::Predicate::ICMP_SGT;
3194 case ICmpInst::Predicate::ICMP_SLE:
3195 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3196 if (X != I.getOperand(0)) // X must be on LHS of comparison!
3197 return nullptr; // Ignore the other case.
3198 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3200 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3202 DstPred = ICmpInst::Predicate::ICMP_SLE;
3205 llvm_unreachable("All possible folds are handled.");
3208 return Builder.CreateICmp(DstPred, X, M);
3211 /// Some comparisons can be simplified.
3212 /// In this case, we are looking for comparisons that look like
3213 /// a check for a lossy signed truncation.
3214 /// Folds: (MaskedBits is a constant.)
3215 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3217 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3218 /// Where KeptBits = bitwidth(%x) - MaskedBits
3220 foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3221 InstCombiner::BuilderTy &Builder) {
3222 ICmpInst::Predicate SrcPred;
3224 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3225 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3226 if (!match(&I, m_c_ICmp(SrcPred,
3227 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3232 // Potential handling of non-splats: for each element:
3233 // * if both are undef, replace with constant 0.
3234 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3235 // * if both are not undef, and are different, bailout.
3236 // * else, only one is undef, then pick the non-undef one.
3238 // The shift amount must be equal.
3241 const APInt &MaskedBits = *C0;
3242 assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3244 ICmpInst::Predicate DstPred;
3246 case ICmpInst::Predicate::ICMP_EQ:
3247 // ((%x << MaskedBits) a>> MaskedBits) == %x
3249 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3250 DstPred = ICmpInst::Predicate::ICMP_ULT;
3252 case ICmpInst::Predicate::ICMP_NE:
3253 // ((%x << MaskedBits) a>> MaskedBits) != %x
3255 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3256 DstPred = ICmpInst::Predicate::ICMP_UGE;
3258 // FIXME: are more folds possible?
3263 auto *XType = X->getType();
3264 const unsigned XBitWidth = XType->getScalarSizeInBits();
3265 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3266 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3268 // KeptBits = bitwidth(%x) - MaskedBits
3269 const APInt KeptBits = BitWidth - MaskedBits;
3270 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3271 // ICmpCst = (1 << KeptBits)
3272 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3273 assert(ICmpCst.isPowerOf2());
3274 // AddCst = (1 << (KeptBits-1))
3275 const APInt AddCst = ICmpCst.lshr(1);
3276 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3278 // T0 = add %x, AddCst
3279 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3280 // T1 = T0 DstPred ICmpCst
3281 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3287 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3288 // we should move shifts to the same hand of 'and', i.e. rewrite as
3289 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3290 // We are only interested in opposite logical shifts here.
3291 // If we can, we want to end up creating 'lshr' shift.
3293 foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
3294 InstCombiner::BuilderTy &Builder) {
3295 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3296 !I.getOperand(0)->hasOneUse())
3299 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3300 auto m_AnyLShr = m_LShr(m_Value(), m_Value());
3302 // Look for an 'and' of two (opposite) logical shifts.
3303 // Pick the single-use shift as XShift.
3304 Instruction *XShift, *YShift;
3305 if (!match(I.getOperand(0),
3306 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3307 m_CombineAnd(m_AnyLogicalShift, m_Instruction(YShift)))))
3310 // If YShift is a 'lshr', swap the shifts around.
3311 if (match(YShift, m_AnyLShr))
3312 std::swap(XShift, YShift);
3314 // The shifts must be in opposite directions.
3315 auto XShiftOpcode = XShift->getOpcode();
3316 if (XShiftOpcode == YShift->getOpcode())
3317 return nullptr; // Do not care about same-direction shifts here.
3319 Value *X, *XShAmt, *Y, *YShAmt;
3320 match(XShift, m_BinOp(m_Value(X), m_Value(XShAmt)));
3321 match(YShift, m_BinOp(m_Value(Y), m_Value(YShAmt)));
3323 // If one of the values being shifted is a constant, then we will end with
3324 // and+icmp, and shift instr will be constant-folded. If they are not,
3325 // however, we will need to ensure that we won't increase instruction count.
3326 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3327 // At least one of the hands of the 'and' should be one-use shift.
3328 if (!match(I.getOperand(0),
3329 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3333 // Can we fold (XShAmt+YShAmt) ?
3334 Value *NewShAmt = SimplifyBinOp(Instruction::BinaryOps::Add, XShAmt, YShAmt,
3335 SQ.getWithInstruction(&I));
3338 // Is the new shift amount smaller than the bit width?
3339 // FIXME: could also rely on ConstantRange.
3340 unsigned BitWidth = X->getType()->getScalarSizeInBits();
3341 if (!match(NewShAmt, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3342 APInt(BitWidth, BitWidth))))
3344 // All good, we can do this fold. The shift is the same that was for X.
3345 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3346 ? Builder.CreateLShr(X, NewShAmt)
3347 : Builder.CreateShl(X, NewShAmt);
3348 Value *T1 = Builder.CreateAnd(T0, Y);
3349 return Builder.CreateICmp(I.getPredicate(), T1,
3350 Constant::getNullValue(X->getType()));
3353 /// Try to fold icmp (binop), X or icmp X, (binop).
3354 /// TODO: A large part of this logic is duplicated in InstSimplify's
3355 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3357 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
3358 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3360 // Special logic for binary operators.
3361 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3362 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3366 const CmpInst::Predicate Pred = I.getPredicate();
3369 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3370 // (Op1 + X) <u Op1 --> ~Op1 <u X
3371 // Op0 >u (Op0 + X) --> X >u ~Op0
3372 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3373 Pred == ICmpInst::ICMP_ULT)
3374 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3375 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3376 Pred == ICmpInst::ICMP_UGT)
3377 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3379 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3380 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3382 ICmpInst::isEquality(Pred) ||
3383 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3384 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3385 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3387 ICmpInst::isEquality(Pred) ||
3388 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3389 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3391 // Analyze the case when either Op0 or Op1 is an add instruction.
3392 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3393 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3394 if (BO0 && BO0->getOpcode() == Instruction::Add) {
3395 A = BO0->getOperand(0);
3396 B = BO0->getOperand(1);
3398 if (BO1 && BO1->getOpcode() == Instruction::Add) {
3399 C = BO1->getOperand(0);
3400 D = BO1->getOperand(1);
3403 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3404 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3405 return new ICmpInst(Pred, A == Op1 ? B : A,
3406 Constant::getNullValue(Op1->getType()));
3408 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3409 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3410 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3413 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3414 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3416 // Try not to increase register pressure.
3417 BO0->hasOneUse() && BO1->hasOneUse()) {
3418 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3421 // C + B == C + D -> B == D
3424 } else if (A == D) {
3425 // D + B == C + D -> B == C
3428 } else if (B == C) {
3429 // A + C == C + D -> A == D
3434 // A + D == C + D -> A == C
3438 return new ICmpInst(Pred, Y, Z);
3441 // icmp slt (X + -1), Y -> icmp sle X, Y
3442 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3443 match(B, m_AllOnes()))
3444 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3446 // icmp sge (X + -1), Y -> icmp sgt X, Y
3447 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3448 match(B, m_AllOnes()))
3449 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3451 // icmp sle (X + 1), Y -> icmp slt X, Y
3452 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3453 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3455 // icmp sgt (X + 1), Y -> icmp sge X, Y
3456 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3457 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3459 // icmp sgt X, (Y + -1) -> icmp sge X, Y
3460 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3461 match(D, m_AllOnes()))
3462 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3464 // icmp sle X, (Y + -1) -> icmp slt X, Y
3465 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3466 match(D, m_AllOnes()))
3467 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3469 // icmp sge X, (Y + 1) -> icmp sgt X, Y
3470 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3471 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3473 // icmp slt X, (Y + 1) -> icmp sle X, Y
3474 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3475 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3477 // TODO: The subtraction-related identities shown below also hold, but
3478 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3479 // wouldn't happen even if they were implemented.
3481 // icmp ult (X - 1), Y -> icmp ule X, Y
3482 // icmp uge (X - 1), Y -> icmp ugt X, Y
3483 // icmp ugt X, (Y - 1) -> icmp uge X, Y
3484 // icmp ule X, (Y - 1) -> icmp ult X, Y
3486 // icmp ule (X + 1), Y -> icmp ult X, Y
3487 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3488 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3490 // icmp ugt (X + 1), Y -> icmp uge X, Y
3491 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3492 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3494 // icmp uge X, (Y + 1) -> icmp ugt X, Y
3495 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3496 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3498 // icmp ult X, (Y + 1) -> icmp ule X, Y
3499 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3500 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3502 // if C1 has greater magnitude than C2:
3503 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3504 // s.t. C3 = C1 - C2
3506 // if C2 has greater magnitude than C1:
3507 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3508 // s.t. C3 = C2 - C1
3509 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3510 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3511 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3512 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3513 const APInt &AP1 = C1->getValue();
3514 const APInt &AP2 = C2->getValue();
3515 if (AP1.isNegative() == AP2.isNegative()) {
3516 APInt AP1Abs = C1->getValue().abs();
3517 APInt AP2Abs = C2->getValue().abs();
3518 if (AP1Abs.uge(AP2Abs)) {
3519 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3520 Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3521 return new ICmpInst(Pred, NewAdd, C);
3523 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3524 Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3525 return new ICmpInst(Pred, A, NewAdd);
3530 // Analyze the case when either Op0 or Op1 is a sub instruction.
3531 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3536 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3537 A = BO0->getOperand(0);
3538 B = BO0->getOperand(1);
3540 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3541 C = BO1->getOperand(0);
3542 D = BO1->getOperand(1);
3545 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3546 if (A == Op1 && NoOp0WrapProblem)
3547 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3548 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3549 if (C == Op0 && NoOp1WrapProblem)
3550 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3552 // (A - B) >u A --> A <u B
3553 if (A == Op1 && Pred == ICmpInst::ICMP_UGT)
3554 return new ICmpInst(ICmpInst::ICMP_ULT, A, B);
3555 // C <u (C - D) --> C <u D
3556 if (C == Op0 && Pred == ICmpInst::ICMP_ULT)
3557 return new ICmpInst(ICmpInst::ICMP_ULT, C, D);
3559 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3560 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3561 // Try not to increase register pressure.
3562 BO0->hasOneUse() && BO1->hasOneUse())
3563 return new ICmpInst(Pred, A, C);
3564 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3565 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3566 // Try not to increase register pressure.
3567 BO0->hasOneUse() && BO1->hasOneUse())
3568 return new ICmpInst(Pred, D, B);
3570 // icmp (0-X) < cst --> x > -cst
3571 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3573 if (match(BO0, m_Neg(m_Value(X))))
3574 if (Constant *RHSC = dyn_cast<Constant>(Op1))
3575 if (RHSC->isNotMinSignedValue())
3576 return new ICmpInst(I.getSwappedPredicate(), X,
3577 ConstantExpr::getNeg(RHSC));
3580 BinaryOperator *SRem = nullptr;
3581 // icmp (srem X, Y), Y
3582 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3584 // icmp Y, (srem X, Y)
3585 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3586 Op0 == BO1->getOperand(1))
3589 // We don't check hasOneUse to avoid increasing register pressure because
3590 // the value we use is the same value this instruction was already using.
3591 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3594 case ICmpInst::ICMP_EQ:
3595 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3596 case ICmpInst::ICMP_NE:
3597 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3598 case ICmpInst::ICMP_SGT:
3599 case ICmpInst::ICMP_SGE:
3600 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3601 Constant::getAllOnesValue(SRem->getType()));
3602 case ICmpInst::ICMP_SLT:
3603 case ICmpInst::ICMP_SLE:
3604 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3605 Constant::getNullValue(SRem->getType()));
3609 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3610 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3611 switch (BO0->getOpcode()) {
3614 case Instruction::Add:
3615 case Instruction::Sub:
3616 case Instruction::Xor: {
3617 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3618 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3621 if (match(BO0->getOperand(1), m_APInt(C))) {
3622 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3623 if (C->isSignMask()) {
3624 ICmpInst::Predicate NewPred =
3625 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3626 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3629 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3630 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3631 ICmpInst::Predicate NewPred =
3632 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3633 NewPred = I.getSwappedPredicate(NewPred);
3634 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3639 case Instruction::Mul: {
3640 if (!I.isEquality())
3644 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3646 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3647 // Mask = -1 >> count-trailing-zeros(C).
3648 if (unsigned TZs = C->countTrailingZeros()) {
3649 Constant *Mask = ConstantInt::get(
3651 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3652 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3653 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3654 return new ICmpInst(Pred, And1, And2);
3656 // If there are no trailing zeros in the multiplier, just eliminate
3657 // the multiplies (no masking is needed):
3658 // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3659 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3663 case Instruction::UDiv:
3664 case Instruction::LShr:
3665 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3667 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3669 case Instruction::SDiv:
3670 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3672 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3674 case Instruction::AShr:
3675 if (!BO0->isExact() || !BO1->isExact())
3677 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3679 case Instruction::Shl: {
3680 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3681 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3684 if (!NSW && I.isSigned())
3686 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3692 // Transform A & (L - 1) `ult` L --> L != 0
3693 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3694 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3696 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3697 auto *Zero = Constant::getNullValue(BO0->getType());
3698 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3702 if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
3703 return replaceInstUsesWith(I, V);
3705 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
3706 return replaceInstUsesWith(I, V);
3708 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
3709 return replaceInstUsesWith(I, V);
3714 /// Fold icmp Pred min|max(X, Y), X.
3715 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
3716 ICmpInst::Predicate Pred = Cmp.getPredicate();
3717 Value *Op0 = Cmp.getOperand(0);
3718 Value *X = Cmp.getOperand(1);
3720 // Canonicalize minimum or maximum operand to LHS of the icmp.
3721 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3722 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3723 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3724 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3726 Pred = Cmp.getSwappedPredicate();
3730 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3731 // smin(X, Y) == X --> X s<= Y
3732 // smin(X, Y) s>= X --> X s<= Y
3733 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3734 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3736 // smin(X, Y) != X --> X s> Y
3737 // smin(X, Y) s< X --> X s> Y
3738 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3739 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3741 // These cases should be handled in InstSimplify:
3742 // smin(X, Y) s<= X --> true
3743 // smin(X, Y) s> X --> false
3747 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3748 // smax(X, Y) == X --> X s>= Y
3749 // smax(X, Y) s<= X --> X s>= Y
3750 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3751 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3753 // smax(X, Y) != X --> X s< Y
3754 // smax(X, Y) s> X --> X s< Y
3755 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3756 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3758 // These cases should be handled in InstSimplify:
3759 // smax(X, Y) s>= X --> true
3760 // smax(X, Y) s< X --> false
3764 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3765 // umin(X, Y) == X --> X u<= Y
3766 // umin(X, Y) u>= X --> X u<= Y
3767 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3768 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3770 // umin(X, Y) != X --> X u> Y
3771 // umin(X, Y) u< X --> X u> Y
3772 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3773 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3775 // These cases should be handled in InstSimplify:
3776 // umin(X, Y) u<= X --> true
3777 // umin(X, Y) u> X --> false
3781 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3782 // umax(X, Y) == X --> X u>= Y
3783 // umax(X, Y) u<= X --> X u>= Y
3784 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3785 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3787 // umax(X, Y) != X --> X u< Y
3788 // umax(X, Y) u> X --> X u< Y
3789 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3790 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3792 // These cases should be handled in InstSimplify:
3793 // umax(X, Y) u>= X --> true
3794 // umax(X, Y) u< X --> false
3801 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3802 if (!I.isEquality())
3805 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3806 const CmpInst::Predicate Pred = I.getPredicate();
3807 Value *A, *B, *C, *D;
3808 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3809 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3810 Value *OtherVal = A == Op1 ? B : A;
3811 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3814 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3815 // A^c1 == C^c2 --> A == C^(c1^c2)
3816 ConstantInt *C1, *C2;
3817 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3819 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
3820 Value *Xor = Builder.CreateXor(C, NC);
3821 return new ICmpInst(Pred, A, Xor);
3824 // A^B == A^D -> B == D
3826 return new ICmpInst(Pred, B, D);
3828 return new ICmpInst(Pred, B, C);
3830 return new ICmpInst(Pred, A, D);
3832 return new ICmpInst(Pred, A, C);
3836 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3837 // A == (A^B) -> B == 0
3838 Value *OtherVal = A == Op0 ? B : A;
3839 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3842 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3843 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3844 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3845 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3851 } else if (A == D) {
3855 } else if (B == C) {
3859 } else if (B == D) {
3865 if (X) { // Build (X^Y) & Z
3866 Op1 = Builder.CreateXor(X, Y);
3867 Op1 = Builder.CreateAnd(Op1, Z);
3868 I.setOperand(0, Op1);
3869 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3874 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3875 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3877 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3878 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3879 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3880 match(Op1, m_ZExt(m_Value(A))))) {
3881 APInt Pow2 = Cst1->getValue() + 1;
3882 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3883 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3884 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
3887 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3888 // For lshr and ashr pairs.
3889 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3890 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3891 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3892 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3893 unsigned TypeBits = Cst1->getBitWidth();
3894 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3895 if (ShAmt < TypeBits && ShAmt != 0) {
3896 ICmpInst::Predicate NewPred =
3897 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
3898 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3899 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3900 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
3904 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3905 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3906 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3907 unsigned TypeBits = Cst1->getBitWidth();
3908 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3909 if (ShAmt < TypeBits && ShAmt != 0) {
3910 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3911 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3912 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
3913 I.getName() + ".mask");
3914 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
3918 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3919 // "icmp (and X, mask), cst"
3921 if (Op0->hasOneUse() &&
3922 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3923 match(Op1, m_ConstantInt(Cst1)) &&
3924 // Only do this when A has multiple uses. This is most important to do
3925 // when it exposes other optimizations.
3927 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3929 if (ShAmt < ASize) {
3931 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3934 APInt CmpV = Cst1->getValue().zext(ASize);
3937 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
3938 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
3942 // If both operands are byte-swapped or bit-reversed, just compare the
3944 // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
3945 // and handle more intrinsics.
3946 if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
3947 (match(Op0, m_BitReverse(m_Value(A))) &&
3948 match(Op1, m_BitReverse(m_Value(B)))))
3949 return new ICmpInst(Pred, A, B);
3951 // Canonicalize checking for a power-of-2-or-zero value:
3952 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
3953 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
3954 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
3956 !match(Op1, m_ZeroInt()))
3959 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
3960 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
3961 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
3964 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
3968 Type *Ty = A->getType();
3969 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
3970 return Pred == ICmpInst::ICMP_EQ
3971 ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
3972 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
3978 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3980 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3981 const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3982 Value *LHSCIOp = LHSCI->getOperand(0);
3983 Type *SrcTy = LHSCIOp->getType();
3984 Type *DestTy = LHSCI->getType();
3986 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3987 // integer type is the same size as the pointer type.
3988 const auto& CompatibleSizes = [&](Type* SrcTy, Type* DestTy) -> bool {
3989 if (isa<VectorType>(SrcTy)) {
3990 SrcTy = cast<VectorType>(SrcTy)->getElementType();
3991 DestTy = cast<VectorType>(DestTy)->getElementType();
3993 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
3995 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3996 CompatibleSizes(SrcTy, DestTy)) {
3997 Value *RHSOp = nullptr;
3998 if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3999 Value *RHSCIOp = RHSC->getOperand(0);
4000 if (RHSCIOp->getType()->getPointerAddressSpace() ==
4001 LHSCIOp->getType()->getPointerAddressSpace()) {
4002 RHSOp = RHSC->getOperand(0);
4003 // If the pointer types don't match, insert a bitcast.
4004 if (LHSCIOp->getType() != RHSOp->getType())
4005 RHSOp = Builder.CreateBitCast(RHSOp, LHSCIOp->getType());
4007 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4008 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4012 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
4015 // The code below only handles extension cast instructions, so far.
4017 if (LHSCI->getOpcode() != Instruction::ZExt &&
4018 LHSCI->getOpcode() != Instruction::SExt)
4021 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
4022 bool isSignedCmp = ICmp.isSigned();
4024 if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4025 // Not an extension from the same type?
4026 Value *RHSCIOp = CI->getOperand(0);
4027 if (RHSCIOp->getType() != LHSCIOp->getType())
4030 // If the signedness of the two casts doesn't agree (i.e. one is a sext
4031 // and the other is a zext), then we can't handle this.
4032 if (CI->getOpcode() != LHSCI->getOpcode())
4035 // Deal with equality cases early.
4036 if (ICmp.isEquality())
4037 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
4039 // A signed comparison of sign extended values simplifies into a
4040 // signed comparison.
4041 if (isSignedCmp && isSignedExt)
4042 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
4044 // The other three cases all fold into an unsigned comparison.
4045 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
4048 // If we aren't dealing with a constant on the RHS, exit early.
4049 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4053 // Compute the constant that would happen if we truncated to SrcTy then
4054 // re-extended to DestTy.
4055 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4056 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
4058 // If the re-extended constant didn't change...
4060 // Deal with equality cases early.
4061 if (ICmp.isEquality())
4062 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
4064 // A signed comparison of sign extended values simplifies into a
4065 // signed comparison.
4066 if (isSignedExt && isSignedCmp)
4067 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
4069 // The other three cases all fold into an unsigned comparison.
4070 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
4073 // The re-extended constant changed, partly changed (in the case of a vector),
4074 // or could not be determined to be equal (in the case of a constant
4075 // expression), so the constant cannot be represented in the shorter type.
4076 // Consequently, we cannot emit a simple comparison.
4077 // All the cases that fold to true or false will have already been handled
4078 // by SimplifyICmpInst, so only deal with the tricky case.
4080 if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
4083 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
4084 // should have been folded away previously and not enter in here.
4086 // We're performing an unsigned comp with a sign extended value.
4087 // This is true if the input is >= 0. [aka >s -1]
4088 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
4089 Value *Result = Builder.CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
4091 // Finally, return the value computed.
4092 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4093 return replaceInstUsesWith(ICmp, Result);
4095 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
4096 return BinaryOperator::CreateNot(Result);
4099 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4102 llvm_unreachable("Unsupported binary op");
4103 case Instruction::Add:
4104 case Instruction::Sub:
4105 return match(RHS, m_Zero());
4106 case Instruction::Mul:
4107 return match(RHS, m_One());
4111 OverflowResult InstCombiner::computeOverflow(
4112 Instruction::BinaryOps BinaryOp, bool IsSigned,
4113 Value *LHS, Value *RHS, Instruction *CxtI) const {
4116 llvm_unreachable("Unsupported binary op");
4117 case Instruction::Add:
4119 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4121 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4122 case Instruction::Sub:
4124 return computeOverflowForSignedSub(LHS, RHS, CxtI);
4126 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4127 case Instruction::Mul:
4129 return computeOverflowForSignedMul(LHS, RHS, CxtI);
4131 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4135 bool InstCombiner::OptimizeOverflowCheck(
4136 Instruction::BinaryOps BinaryOp, bool IsSigned, Value *LHS, Value *RHS,
4137 Instruction &OrigI, Value *&Result, Constant *&Overflow) {
4138 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4139 std::swap(LHS, RHS);
4141 // If the overflow check was an add followed by a compare, the insertion point
4142 // may be pointing to the compare. We want to insert the new instructions
4143 // before the add in case there are uses of the add between the add and the
4145 Builder.SetInsertPoint(&OrigI);
4147 if (isNeutralValue(BinaryOp, RHS)) {
4149 Overflow = Builder.getFalse();
4153 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4154 case OverflowResult::MayOverflow:
4156 case OverflowResult::AlwaysOverflowsLow:
4157 case OverflowResult::AlwaysOverflowsHigh:
4158 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4159 Result->takeName(&OrigI);
4160 Overflow = Builder.getTrue();
4162 case OverflowResult::NeverOverflows:
4163 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4164 Result->takeName(&OrigI);
4165 Overflow = Builder.getFalse();
4166 if (auto *Inst = dyn_cast<Instruction>(Result)) {
4168 Inst->setHasNoSignedWrap();
4170 Inst->setHasNoUnsignedWrap();
4175 llvm_unreachable("Unexpected overflow result");
4178 /// Recognize and process idiom involving test for multiplication
4181 /// The caller has matched a pattern of the form:
4182 /// I = cmp u (mul(zext A, zext B), V
4183 /// The function checks if this is a test for overflow and if so replaces
4184 /// multiplication with call to 'mul.with.overflow' intrinsic.
4186 /// \param I Compare instruction.
4187 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4188 /// the compare instruction. Must be of integer type.
4189 /// \param OtherVal The other argument of compare instruction.
4190 /// \returns Instruction which must replace the compare instruction, NULL if no
4191 /// replacement required.
4192 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
4193 Value *OtherVal, InstCombiner &IC) {
4194 // Don't bother doing this transformation for pointers, don't do it for
4196 if (!isa<IntegerType>(MulVal->getType()))
4199 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
4200 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
4201 auto *MulInstr = dyn_cast<Instruction>(MulVal);
4204 assert(MulInstr->getOpcode() == Instruction::Mul);
4206 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4207 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4208 assert(LHS->getOpcode() == Instruction::ZExt);
4209 assert(RHS->getOpcode() == Instruction::ZExt);
4210 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4212 // Calculate type and width of the result produced by mul.with.overflow.
4213 Type *TyA = A->getType(), *TyB = B->getType();
4214 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4215 WidthB = TyB->getPrimitiveSizeInBits();
4218 if (WidthB > WidthA) {
4226 // In order to replace the original mul with a narrower mul.with.overflow,
4227 // all uses must ignore upper bits of the product. The number of used low
4228 // bits must be not greater than the width of mul.with.overflow.
4229 if (MulVal->hasNUsesOrMore(2))
4230 for (User *U : MulVal->users()) {
4233 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4234 // Check if truncation ignores bits above MulWidth.
4235 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4236 if (TruncWidth > MulWidth)
4238 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4239 // Check if AND ignores bits above MulWidth.
4240 if (BO->getOpcode() != Instruction::And)
4242 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4243 const APInt &CVal = CI->getValue();
4244 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4247 // In this case we could have the operand of the binary operation
4248 // being defined in another block, and performing the replacement
4249 // could break the dominance relation.
4253 // Other uses prohibit this transformation.
4258 // Recognize patterns
4259 switch (I.getPredicate()) {
4260 case ICmpInst::ICMP_EQ:
4261 case ICmpInst::ICMP_NE:
4262 // Recognize pattern:
4263 // mulval = mul(zext A, zext B)
4264 // cmp eq/neq mulval, zext trunc mulval
4265 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
4266 if (Zext->hasOneUse()) {
4267 Value *ZextArg = Zext->getOperand(0);
4268 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
4269 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
4273 // Recognize pattern:
4274 // mulval = mul(zext A, zext B)
4275 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4278 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4279 if (ValToMask != MulVal)
4281 const APInt &CVal = CI->getValue() + 1;
4282 if (CVal.isPowerOf2()) {
4283 unsigned MaskWidth = CVal.logBase2();
4284 if (MaskWidth == MulWidth)
4285 break; // Recognized
4290 case ICmpInst::ICMP_UGT:
4291 // Recognize pattern:
4292 // mulval = mul(zext A, zext B)
4293 // cmp ugt mulval, max
4294 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4295 APInt MaxVal = APInt::getMaxValue(MulWidth);
4296 MaxVal = MaxVal.zext(CI->getBitWidth());
4297 if (MaxVal.eq(CI->getValue()))
4298 break; // Recognized
4302 case ICmpInst::ICMP_UGE:
4303 // Recognize pattern:
4304 // mulval = mul(zext A, zext B)
4305 // cmp uge mulval, max+1
4306 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4307 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4308 if (MaxVal.eq(CI->getValue()))
4309 break; // Recognized
4313 case ICmpInst::ICMP_ULE:
4314 // Recognize pattern:
4315 // mulval = mul(zext A, zext B)
4316 // cmp ule mulval, max
4317 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4318 APInt MaxVal = APInt::getMaxValue(MulWidth);
4319 MaxVal = MaxVal.zext(CI->getBitWidth());
4320 if (MaxVal.eq(CI->getValue()))
4321 break; // Recognized
4325 case ICmpInst::ICMP_ULT:
4326 // Recognize pattern:
4327 // mulval = mul(zext A, zext B)
4328 // cmp ule mulval, max + 1
4329 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4330 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4331 if (MaxVal.eq(CI->getValue()))
4332 break; // Recognized
4340 InstCombiner::BuilderTy &Builder = IC.Builder;
4341 Builder.SetInsertPoint(MulInstr);
4343 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4344 Value *MulA = A, *MulB = B;
4345 if (WidthA < MulWidth)
4346 MulA = Builder.CreateZExt(A, MulType);
4347 if (WidthB < MulWidth)
4348 MulB = Builder.CreateZExt(B, MulType);
4349 Function *F = Intrinsic::getDeclaration(
4350 I.getModule(), Intrinsic::umul_with_overflow, MulType);
4351 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4352 IC.Worklist.Add(MulInstr);
4354 // If there are uses of mul result other than the comparison, we know that
4355 // they are truncation or binary AND. Change them to use result of
4356 // mul.with.overflow and adjust properly mask/size.
4357 if (MulVal->hasNUsesOrMore(2)) {
4358 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4359 for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4361 if (U == &I || U == OtherVal)
4363 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4364 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4365 IC.replaceInstUsesWith(*TI, Mul);
4367 TI->setOperand(0, Mul);
4368 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4369 assert(BO->getOpcode() == Instruction::And);
4370 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4371 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4372 APInt ShortMask = CI->getValue().trunc(MulWidth);
4373 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4375 cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4376 IC.Worklist.Add(Zext);
4377 IC.replaceInstUsesWith(*BO, Zext);
4379 llvm_unreachable("Unexpected Binary operation");
4381 IC.Worklist.Add(cast<Instruction>(U));
4384 if (isa<Instruction>(OtherVal))
4385 IC.Worklist.Add(cast<Instruction>(OtherVal));
4387 // The original icmp gets replaced with the overflow value, maybe inverted
4388 // depending on predicate.
4389 bool Inverse = false;
4390 switch (I.getPredicate()) {
4391 case ICmpInst::ICMP_NE:
4393 case ICmpInst::ICMP_EQ:
4396 case ICmpInst::ICMP_UGT:
4397 case ICmpInst::ICMP_UGE:
4398 if (I.getOperand(0) == MulVal)
4402 case ICmpInst::ICMP_ULT:
4403 case ICmpInst::ICMP_ULE:
4404 if (I.getOperand(1) == MulVal)
4409 llvm_unreachable("Unexpected predicate");
4412 Value *Res = Builder.CreateExtractValue(Call, 1);
4413 return BinaryOperator::CreateNot(Res);
4416 return ExtractValueInst::Create(Call, 1);
4419 /// When performing a comparison against a constant, it is possible that not all
4420 /// the bits in the LHS are demanded. This helper method computes the mask that
4422 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4424 if (!match(I.getOperand(1), m_APInt(RHS)))
4425 return APInt::getAllOnesValue(BitWidth);
4427 // If this is a normal comparison, it demands all bits. If it is a sign bit
4428 // comparison, it only demands the sign bit.
4430 if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4431 return APInt::getSignMask(BitWidth);
4433 switch (I.getPredicate()) {
4434 // For a UGT comparison, we don't care about any bits that
4435 // correspond to the trailing ones of the comparand. The value of these
4436 // bits doesn't impact the outcome of the comparison, because any value
4437 // greater than the RHS must differ in a bit higher than these due to carry.
4438 case ICmpInst::ICMP_UGT:
4439 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4441 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4442 // Any value less than the RHS must differ in a higher bit because of carries.
4443 case ICmpInst::ICMP_ULT:
4444 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4447 return APInt::getAllOnesValue(BitWidth);
4451 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4452 /// should be swapped.
4453 /// The decision is based on how many times these two operands are reused
4454 /// as subtract operands and their positions in those instructions.
4455 /// The rationale is that several architectures use the same instruction for
4456 /// both subtract and cmp. Thus, it is better if the order of those operands
4458 /// \return true if Op0 and Op1 should be swapped.
4459 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4460 // Filter out pointer values as those cannot appear directly in subtract.
4461 // FIXME: we may want to go through inttoptrs or bitcasts.
4462 if (Op0->getType()->isPointerTy())
4464 // If a subtract already has the same operands as a compare, swapping would be
4465 // bad. If a subtract has the same operands as a compare but in reverse order,
4466 // then swapping is good.
4468 for (const User *U : Op0->users()) {
4469 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4471 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4474 return GoodToSwap > 0;
4477 /// Check that one use is in the same block as the definition and all
4478 /// other uses are in blocks dominated by a given block.
4480 /// \param DI Definition
4482 /// \param DB Block that must dominate all uses of \p DI outside
4483 /// the parent block
4484 /// \return true when \p UI is the only use of \p DI in the parent block
4485 /// and all other uses of \p DI are in blocks dominated by \p DB.
4487 bool InstCombiner::dominatesAllUses(const Instruction *DI,
4488 const Instruction *UI,
4489 const BasicBlock *DB) const {
4490 assert(DI && UI && "Instruction not defined\n");
4491 // Ignore incomplete definitions.
4492 if (!DI->getParent())
4494 // DI and UI must be in the same block.
4495 if (DI->getParent() != UI->getParent())
4497 // Protect from self-referencing blocks.
4498 if (DI->getParent() == DB)
4500 for (const User *U : DI->users()) {
4501 auto *Usr = cast<Instruction>(U);
4502 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4508 /// Return true when the instruction sequence within a block is select-cmp-br.
4509 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4510 const BasicBlock *BB = SI->getParent();
4513 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4514 if (!BI || BI->getNumSuccessors() != 2)
4516 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4517 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4522 /// True when a select result is replaced by one of its operands
4523 /// in select-icmp sequence. This will eventually result in the elimination
4526 /// \param SI Select instruction
4527 /// \param Icmp Compare instruction
4528 /// \param SIOpd Operand that replaces the select
4531 /// - The replacement is global and requires dominator information
4532 /// - The caller is responsible for the actual replacement
4537 /// %4 = select i1 %3, %C* %0, %C* null
4538 /// %5 = icmp eq %C* %4, null
4539 /// br i1 %5, label %9, label %7
4541 /// ; <label>:7 ; preds = %entry
4542 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4545 /// can be transformed to
4547 /// %5 = icmp eq %C* %0, null
4548 /// %6 = select i1 %3, i1 %5, i1 true
4549 /// br i1 %6, label %9, label %7
4551 /// ; <label>:7 ; preds = %entry
4552 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4554 /// Similar when the first operand of the select is a constant or/and
4555 /// the compare is for not equal rather than equal.
4557 /// NOTE: The function is only called when the select and compare constants
4558 /// are equal, the optimization can work only for EQ predicates. This is not a
4559 /// major restriction since a NE compare should be 'normalized' to an equal
4560 /// compare, which usually happens in the combiner and test case
4561 /// select-cmp-br.ll checks for it.
4562 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
4563 const ICmpInst *Icmp,
4564 const unsigned SIOpd) {
4565 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4566 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4567 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4568 // The check for the single predecessor is not the best that can be
4569 // done. But it protects efficiently against cases like when SI's
4570 // home block has two successors, Succ and Succ1, and Succ1 predecessor
4571 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4572 // replaced can be reached on either path. So the uniqueness check
4573 // guarantees that the path all uses of SI (outside SI's parent) are on
4574 // is disjoint from all other paths out of SI. But that information
4575 // is more expensive to compute, and the trade-off here is in favor
4576 // of compile-time. It should also be noticed that we check for a single
4577 // predecessor and not only uniqueness. This to handle the situation when
4578 // Succ and Succ1 points to the same basic block.
4579 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4581 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4588 /// Try to fold the comparison based on range information we can get by checking
4589 /// whether bits are known to be zero or one in the inputs.
4590 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4591 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4592 Type *Ty = Op0->getType();
4593 ICmpInst::Predicate Pred = I.getPredicate();
4595 // Get scalar or pointer size.
4596 unsigned BitWidth = Ty->isIntOrIntVectorTy()
4597 ? Ty->getScalarSizeInBits()
4598 : DL.getIndexTypeSizeInBits(Ty->getScalarType());
4603 KnownBits Op0Known(BitWidth);
4604 KnownBits Op1Known(BitWidth);
4606 if (SimplifyDemandedBits(&I, 0,
4607 getDemandedBitsLHSMask(I, BitWidth),
4611 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4615 // Given the known and unknown bits, compute a range that the LHS could be
4616 // in. Compute the Min, Max and RHS values based on the known bits. For the
4617 // EQ and NE we use unsigned values.
4618 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4619 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4621 computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4622 computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4624 computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4625 computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4628 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4629 // out that the LHS or RHS is a constant. Constant fold this now, so that
4630 // code below can assume that Min != Max.
4631 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4632 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4633 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4634 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4636 // Based on the range information we know about the LHS, see if we can
4637 // simplify this comparison. For example, (x&4) < 8 is always true.
4640 llvm_unreachable("Unknown icmp opcode!");
4641 case ICmpInst::ICMP_EQ:
4642 case ICmpInst::ICMP_NE: {
4643 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4644 return Pred == CmpInst::ICMP_EQ
4645 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4646 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4649 // If all bits are known zero except for one, then we know at most one bit
4650 // is set. If the comparison is against zero, then this is a check to see if
4651 // *that* bit is set.
4652 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4653 if (Op1Known.isZero()) {
4654 // If the LHS is an AND with the same constant, look through it.
4655 Value *LHS = nullptr;
4657 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4658 *LHSC != Op0KnownZeroInverted)
4662 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4663 APInt ValToCheck = Op0KnownZeroInverted;
4664 Type *XTy = X->getType();
4665 if (ValToCheck.isPowerOf2()) {
4666 // ((1 << X) & 8) == 0 -> X != 3
4667 // ((1 << X) & 8) != 0 -> X == 3
4668 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4669 auto NewPred = ICmpInst::getInversePredicate(Pred);
4670 return new ICmpInst(NewPred, X, CmpC);
4671 } else if ((++ValToCheck).isPowerOf2()) {
4672 // ((1 << X) & 7) == 0 -> X >= 3
4673 // ((1 << X) & 7) != 0 -> X < 3
4674 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4676 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
4677 return new ICmpInst(NewPred, X, CmpC);
4681 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4683 if (Op0KnownZeroInverted.isOneValue() &&
4684 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4685 // ((8 >>u X) & 1) == 0 -> X != 3
4686 // ((8 >>u X) & 1) != 0 -> X == 3
4687 unsigned CmpVal = CI->countTrailingZeros();
4688 auto NewPred = ICmpInst::getInversePredicate(Pred);
4689 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4694 case ICmpInst::ICMP_ULT: {
4695 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4696 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4697 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4698 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4699 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4700 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4703 if (match(Op1, m_APInt(CmpC))) {
4704 // A <u C -> A == C-1 if min(A)+1 == C
4705 if (*CmpC == Op0Min + 1)
4706 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4707 ConstantInt::get(Op1->getType(), *CmpC - 1));
4708 // X <u C --> X == 0, if the number of zero bits in the bottom of X
4709 // exceeds the log2 of C.
4710 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
4711 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4712 Constant::getNullValue(Op1->getType()));
4716 case ICmpInst::ICMP_UGT: {
4717 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4718 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4719 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4720 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4721 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4722 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4725 if (match(Op1, m_APInt(CmpC))) {
4726 // A >u C -> A == C+1 if max(a)-1 == C
4727 if (*CmpC == Op0Max - 1)
4728 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4729 ConstantInt::get(Op1->getType(), *CmpC + 1));
4730 // X >u C --> X != 0, if the number of zero bits in the bottom of X
4731 // exceeds the log2 of C.
4732 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
4733 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
4734 Constant::getNullValue(Op1->getType()));
4738 case ICmpInst::ICMP_SLT: {
4739 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4740 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4741 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4742 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4743 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4744 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4746 if (match(Op1, m_APInt(CmpC))) {
4747 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4748 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4749 ConstantInt::get(Op1->getType(), *CmpC - 1));
4753 case ICmpInst::ICMP_SGT: {
4754 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4755 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4756 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4757 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4758 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4759 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4761 if (match(Op1, m_APInt(CmpC))) {
4762 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4763 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4764 ConstantInt::get(Op1->getType(), *CmpC + 1));
4768 case ICmpInst::ICMP_SGE:
4769 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4770 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4771 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4772 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4773 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4774 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
4775 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4777 case ICmpInst::ICMP_SLE:
4778 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4779 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4780 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4781 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4782 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4783 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
4784 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4786 case ICmpInst::ICMP_UGE:
4787 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4788 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4789 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4790 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4791 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4792 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
4793 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4795 case ICmpInst::ICMP_ULE:
4796 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4797 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4798 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4799 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4800 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4801 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
4802 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4806 // Turn a signed comparison into an unsigned one if both operands are known to
4807 // have the same sign.
4809 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
4810 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
4811 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4816 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4817 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4818 /// allows them to be folded in visitICmpInst.
4819 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
4820 ICmpInst::Predicate Pred = I.getPredicate();
4821 if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4822 Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4825 Value *Op0 = I.getOperand(0);
4826 Value *Op1 = I.getOperand(1);
4827 auto *Op1C = dyn_cast<Constant>(Op1);
4831 // Check if the constant operand can be safely incremented/decremented without
4832 // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4833 // the edge cases for us, so we just assert on them. For vectors, we must
4834 // handle the edge cases.
4835 Type *Op1Type = Op1->getType();
4836 bool IsSigned = I.isSigned();
4837 bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4838 auto *CI = dyn_cast<ConstantInt>(Op1C);
4840 // A <= MAX -> TRUE ; A >= MIN -> TRUE
4841 assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4842 } else if (Op1Type->isVectorTy()) {
4843 // TODO? If the edge cases for vectors were guaranteed to be handled as they
4844 // are for scalar, we could remove the min/max checks. However, to do that,
4845 // we would have to use insertelement/shufflevector to replace edge values.
4846 unsigned NumElts = Op1Type->getVectorNumElements();
4847 for (unsigned i = 0; i != NumElts; ++i) {
4848 Constant *Elt = Op1C->getAggregateElement(i);
4852 if (isa<UndefValue>(Elt))
4855 // Bail out if we can't determine if this constant is min/max or if we
4856 // know that this constant is min/max.
4857 auto *CI = dyn_cast<ConstantInt>(Elt);
4858 if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4866 // Increment or decrement the constant and set the new comparison predicate:
4867 // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4868 Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4869 CmpInst::Predicate NewPred = IsLE ? ICmpInst::ICMP_ULT: ICmpInst::ICMP_UGT;
4870 NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4871 return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4874 /// Integer compare with boolean values can always be turned into bitwise ops.
4875 static Instruction *canonicalizeICmpBool(ICmpInst &I,
4876 InstCombiner::BuilderTy &Builder) {
4877 Value *A = I.getOperand(0), *B = I.getOperand(1);
4878 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
4880 // A boolean compared to true/false can be simplified to Op0/true/false in
4881 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
4882 // Cases not handled by InstSimplify are always 'not' of Op0.
4883 if (match(B, m_Zero())) {
4884 switch (I.getPredicate()) {
4885 case CmpInst::ICMP_EQ: // A == 0 -> !A
4886 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
4887 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
4888 return BinaryOperator::CreateNot(A);
4890 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4892 } else if (match(B, m_One())) {
4893 switch (I.getPredicate()) {
4894 case CmpInst::ICMP_NE: // A != 1 -> !A
4895 case CmpInst::ICMP_ULT: // A <u 1 -> !A
4896 case CmpInst::ICMP_SGT: // A >s -1 -> !A
4897 return BinaryOperator::CreateNot(A);
4899 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4903 switch (I.getPredicate()) {
4905 llvm_unreachable("Invalid icmp instruction!");
4906 case ICmpInst::ICMP_EQ:
4907 // icmp eq i1 A, B -> ~(A ^ B)
4908 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
4910 case ICmpInst::ICMP_NE:
4911 // icmp ne i1 A, B -> A ^ B
4912 return BinaryOperator::CreateXor(A, B);
4914 case ICmpInst::ICMP_UGT:
4915 // icmp ugt -> icmp ult
4918 case ICmpInst::ICMP_ULT:
4919 // icmp ult i1 A, B -> ~A & B
4920 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
4922 case ICmpInst::ICMP_SGT:
4923 // icmp sgt -> icmp slt
4926 case ICmpInst::ICMP_SLT:
4927 // icmp slt i1 A, B -> A & ~B
4928 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
4930 case ICmpInst::ICMP_UGE:
4931 // icmp uge -> icmp ule
4934 case ICmpInst::ICMP_ULE:
4935 // icmp ule i1 A, B -> ~A | B
4936 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
4938 case ICmpInst::ICMP_SGE:
4939 // icmp sge -> icmp sle
4942 case ICmpInst::ICMP_SLE:
4943 // icmp sle i1 A, B -> A | ~B
4944 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
4948 // Transform pattern like:
4949 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
4950 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
4954 static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
4955 InstCombiner::BuilderTy &Builder) {
4956 ICmpInst::Predicate Pred, NewPred;
4959 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
4960 // We want X to be the icmp's second operand, so swap predicate if it isn't.
4961 if (Cmp.getOperand(0) == X)
4962 Pred = Cmp.getSwappedPredicate();
4965 case ICmpInst::ICMP_ULE:
4966 NewPred = ICmpInst::ICMP_NE;
4968 case ICmpInst::ICMP_UGT:
4969 NewPred = ICmpInst::ICMP_EQ;
4974 } else if (match(&Cmp, m_c_ICmp(Pred,
4975 m_OneUse(m_CombineOr(
4976 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
4977 m_Add(m_Shl(m_One(), m_Value(Y)),
4980 // The variant with 'add' is not canonical, (the variant with 'not' is)
4981 // we only get it because it has extra uses, and can't be canonicalized,
4983 // We want X to be the icmp's second operand, so swap predicate if it isn't.
4984 if (Cmp.getOperand(0) == X)
4985 Pred = Cmp.getSwappedPredicate();
4988 case ICmpInst::ICMP_ULT:
4989 NewPred = ICmpInst::ICMP_NE;
4991 case ICmpInst::ICMP_UGE:
4992 NewPred = ICmpInst::ICMP_EQ;
5000 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
5001 Constant *Zero = Constant::getNullValue(NewX->getType());
5002 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
5005 static Instruction *foldVectorCmp(CmpInst &Cmp,
5006 InstCombiner::BuilderTy &Builder) {
5007 // If both arguments of the cmp are shuffles that use the same mask and
5008 // shuffle within a single vector, move the shuffle after the cmp.
5009 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
5012 if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(M))) &&
5013 match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(M))) &&
5014 V1->getType() == V2->getType() &&
5015 (LHS->hasOneUse() || RHS->hasOneUse())) {
5016 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
5017 CmpInst::Predicate P = Cmp.getPredicate();
5018 Value *NewCmp = isa<ICmpInst>(Cmp) ? Builder.CreateICmp(P, V1, V2)
5019 : Builder.CreateFCmp(P, V1, V2);
5020 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
5025 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5026 bool Changed = false;
5027 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5028 unsigned Op0Cplxity = getComplexity(Op0);
5029 unsigned Op1Cplxity = getComplexity(Op1);
5031 /// Orders the operands of the compare so that they are listed from most
5032 /// complex to least complex. This puts constants before unary operators,
5033 /// before binary operators.
5034 if (Op0Cplxity < Op1Cplxity ||
5035 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
5037 std::swap(Op0, Op1);
5041 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1,
5042 SQ.getWithInstruction(&I)))
5043 return replaceInstUsesWith(I, V);
5045 // Comparing -val or val with non-zero is the same as just comparing val
5046 // ie, abs(val) != 0 -> val != 0
5047 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
5048 Value *Cond, *SelectTrue, *SelectFalse;
5049 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
5050 m_Value(SelectFalse)))) {
5051 if (Value *V = dyn_castNegVal(SelectTrue)) {
5052 if (V == SelectFalse)
5053 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5055 else if (Value *V = dyn_castNegVal(SelectFalse)) {
5056 if (V == SelectTrue)
5057 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5062 if (Op0->getType()->isIntOrIntVectorTy(1))
5063 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
5066 if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
5069 if (Instruction *Res = foldICmpWithConstant(I))
5072 if (Instruction *Res = foldICmpWithDominatingICmp(I))
5075 if (Instruction *Res = foldICmpUsingKnownBits(I))
5078 // Test if the ICmpInst instruction is used exclusively by a select as
5079 // part of a minimum or maximum operation. If so, refrain from doing
5080 // any other folding. This helps out other analyses which understand
5081 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5082 // and CodeGen. And in this case, at least one of the comparison
5083 // operands has at least one user besides the compare (the select),
5084 // which would often largely negate the benefit of folding anyway.
5086 // Do the same for the other patterns recognized by matchSelectPattern.
5088 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
5090 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
5091 if (SPR.Flavor != SPF_UNKNOWN)
5095 // Do this after checking for min/max to prevent infinite looping.
5096 if (Instruction *Res = foldICmpWithZero(I))
5099 // FIXME: We only do this after checking for min/max to prevent infinite
5100 // looping caused by a reverse canonicalization of these patterns for min/max.
5101 // FIXME: The organization of folds is a mess. These would naturally go into
5102 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
5103 // down here after the min/max restriction.
5104 ICmpInst::Predicate Pred = I.getPredicate();
5106 if (match(Op1, m_APInt(C))) {
5107 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
5108 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
5109 Constant *Zero = Constant::getNullValue(Op0->getType());
5110 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
5113 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
5114 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
5115 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
5116 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
5120 if (Instruction *Res = foldICmpInstWithConstant(I))
5123 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
5126 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5127 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
5128 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
5130 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
5131 if (Instruction *NI = foldGEPICmp(GEP, Op0,
5132 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5135 // Try to optimize equality comparisons against alloca-based pointers.
5136 if (Op0->getType()->isPointerTy() && I.isEquality()) {
5137 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
5138 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
5139 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
5141 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
5142 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
5146 if (Instruction *Res = foldICmpBitCast(I, Builder))
5149 if (isa<CastInst>(Op0)) {
5150 // Handle the special case of: icmp (cast bool to X), <cst>
5151 // This comes up when you have code like
5154 // For generality, we handle any zero-extension of any operand comparison
5155 // with a constant or another cast from the same type.
5156 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
5157 if (Instruction *R = foldICmpWithCastAndCast(I))
5161 if (Instruction *Res = foldICmpBinOp(I))
5164 if (Instruction *Res = foldICmpWithMinMax(I))
5169 // Transform (A & ~B) == 0 --> (A & B) != 0
5170 // and (A & ~B) != 0 --> (A & B) == 0
5171 // if A is a power of 2.
5172 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
5173 match(Op1, m_Zero()) &&
5174 isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
5175 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
5178 // ~X < ~Y --> Y < X
5179 // ~X < C --> X > ~C
5180 if (match(Op0, m_Not(m_Value(A)))) {
5181 if (match(Op1, m_Not(m_Value(B))))
5182 return new ICmpInst(I.getPredicate(), B, A);
5185 if (match(Op1, m_APInt(C)))
5186 return new ICmpInst(I.getSwappedPredicate(), A,
5187 ConstantInt::get(Op1->getType(), ~(*C)));
5190 Instruction *AddI = nullptr;
5191 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
5192 m_Instruction(AddI))) &&
5193 isa<IntegerType>(A->getType())) {
5196 if (OptimizeOverflowCheck(Instruction::Add, /*Signed*/false, A, B,
5197 *AddI, Result, Overflow)) {
5198 replaceInstUsesWith(*AddI, Result);
5199 return replaceInstUsesWith(I, Overflow);
5203 // (zext a) * (zext b) --> llvm.umul.with.overflow.
5204 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5205 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
5208 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5209 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
5214 if (Instruction *Res = foldICmpEquality(I))
5217 // The 'cmpxchg' instruction returns an aggregate containing the old value and
5218 // an i1 which indicates whether or not we successfully did the swap.
5220 // Replace comparisons between the old value and the expected value with the
5221 // indicator that 'cmpxchg' returns.
5223 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
5224 // spuriously fail. In those cases, the old value may equal the expected
5225 // value but it is possible for the swap to not occur.
5226 if (I.getPredicate() == ICmpInst::ICMP_EQ)
5227 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
5228 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
5229 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
5231 return ExtractValueInst::Create(ACXI, 1);
5237 if (match(Op0, m_Add(m_Value(X), m_APInt(C))) && Op1 == X)
5238 return foldICmpAddOpConst(X, *C, I.getPredicate());
5241 if (match(Op1, m_Add(m_Value(X), m_APInt(C))) && Op0 == X)
5242 return foldICmpAddOpConst(X, *C, I.getSwappedPredicate());
5245 if (Instruction *Res = foldICmpWithHighBitMask(I, Builder))
5248 if (I.getType()->isVectorTy())
5249 if (Instruction *Res = foldVectorCmp(I, Builder))
5252 return Changed ? &I : nullptr;
5255 /// Fold fcmp ([us]itofp x, cst) if possible.
5256 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
5258 if (!isa<ConstantFP>(RHSC)) return nullptr;
5259 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5261 // Get the width of the mantissa. We don't want to hack on conversions that
5262 // might lose information from the integer, e.g. "i64 -> float"
5263 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5264 if (MantissaWidth == -1) return nullptr; // Unknown.
5266 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5268 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5270 if (I.isEquality()) {
5271 FCmpInst::Predicate P = I.getPredicate();
5272 bool IsExact = false;
5273 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
5274 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
5276 // If the floating point constant isn't an integer value, we know if we will
5277 // ever compare equal / not equal to it.
5279 // TODO: Can never be -0.0 and other non-representable values
5280 APFloat RHSRoundInt(RHS);
5281 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
5282 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
5283 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
5284 return replaceInstUsesWith(I, Builder.getFalse());
5286 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
5287 return replaceInstUsesWith(I, Builder.getTrue());
5291 // TODO: If the constant is exactly representable, is it always OK to do
5292 // equality compares as integer?
5295 // Check to see that the input is converted from an integer type that is small
5296 // enough that preserves all bits. TODO: check here for "known" sign bits.
5297 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5298 unsigned InputSize = IntTy->getScalarSizeInBits();
5300 // Following test does NOT adjust InputSize downwards for signed inputs,
5301 // because the most negative value still requires all the mantissa bits
5302 // to distinguish it from one less than that value.
5303 if ((int)InputSize > MantissaWidth) {
5304 // Conversion would lose accuracy. Check if loss can impact comparison.
5305 int Exp = ilogb(RHS);
5306 if (Exp == APFloat::IEK_Inf) {
5307 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
5308 if (MaxExponent < (int)InputSize - !LHSUnsigned)
5309 // Conversion could create infinity.
5312 // Note that if RHS is zero or NaN, then Exp is negative
5313 // and first condition is trivially false.
5314 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
5315 // Conversion could affect comparison.
5320 // Otherwise, we can potentially simplify the comparison. We know that it
5321 // will always come through as an integer value and we know the constant is
5322 // not a NAN (it would have been previously simplified).
5323 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5325 ICmpInst::Predicate Pred;
5326 switch (I.getPredicate()) {
5327 default: llvm_unreachable("Unexpected predicate!");
5328 case FCmpInst::FCMP_UEQ:
5329 case FCmpInst::FCMP_OEQ:
5330 Pred = ICmpInst::ICMP_EQ;
5332 case FCmpInst::FCMP_UGT:
5333 case FCmpInst::FCMP_OGT:
5334 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5336 case FCmpInst::FCMP_UGE:
5337 case FCmpInst::FCMP_OGE:
5338 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5340 case FCmpInst::FCMP_ULT:
5341 case FCmpInst::FCMP_OLT:
5342 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5344 case FCmpInst::FCMP_ULE:
5345 case FCmpInst::FCMP_OLE:
5346 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5348 case FCmpInst::FCMP_UNE:
5349 case FCmpInst::FCMP_ONE:
5350 Pred = ICmpInst::ICMP_NE;
5352 case FCmpInst::FCMP_ORD:
5353 return replaceInstUsesWith(I, Builder.getTrue());
5354 case FCmpInst::FCMP_UNO:
5355 return replaceInstUsesWith(I, Builder.getFalse());
5358 // Now we know that the APFloat is a normal number, zero or inf.
5360 // See if the FP constant is too large for the integer. For example,
5361 // comparing an i8 to 300.0.
5362 unsigned IntWidth = IntTy->getScalarSizeInBits();
5365 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5366 // and large values.
5367 APFloat SMax(RHS.getSemantics());
5368 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5369 APFloat::rmNearestTiesToEven);
5370 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5371 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5372 Pred == ICmpInst::ICMP_SLE)
5373 return replaceInstUsesWith(I, Builder.getTrue());
5374 return replaceInstUsesWith(I, Builder.getFalse());
5377 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5378 // +INF and large values.
5379 APFloat UMax(RHS.getSemantics());
5380 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5381 APFloat::rmNearestTiesToEven);
5382 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5383 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5384 Pred == ICmpInst::ICMP_ULE)
5385 return replaceInstUsesWith(I, Builder.getTrue());
5386 return replaceInstUsesWith(I, Builder.getFalse());
5391 // See if the RHS value is < SignedMin.
5392 APFloat SMin(RHS.getSemantics());
5393 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5394 APFloat::rmNearestTiesToEven);
5395 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5396 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5397 Pred == ICmpInst::ICMP_SGE)
5398 return replaceInstUsesWith(I, Builder.getTrue());
5399 return replaceInstUsesWith(I, Builder.getFalse());
5402 // See if the RHS value is < UnsignedMin.
5403 APFloat SMin(RHS.getSemantics());
5404 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
5405 APFloat::rmNearestTiesToEven);
5406 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
5407 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
5408 Pred == ICmpInst::ICMP_UGE)
5409 return replaceInstUsesWith(I, Builder.getTrue());
5410 return replaceInstUsesWith(I, Builder.getFalse());
5414 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5415 // [0, UMAX], but it may still be fractional. See if it is fractional by
5416 // casting the FP value to the integer value and back, checking for equality.
5417 // Don't do this for zero, because -0.0 is not fractional.
5418 Constant *RHSInt = LHSUnsigned
5419 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5420 : ConstantExpr::getFPToSI(RHSC, IntTy);
5421 if (!RHS.isZero()) {
5422 bool Equal = LHSUnsigned
5423 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5424 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5426 // If we had a comparison against a fractional value, we have to adjust
5427 // the compare predicate and sometimes the value. RHSC is rounded towards
5428 // zero at this point.
5430 default: llvm_unreachable("Unexpected integer comparison!");
5431 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5432 return replaceInstUsesWith(I, Builder.getTrue());
5433 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5434 return replaceInstUsesWith(I, Builder.getFalse());
5435 case ICmpInst::ICMP_ULE:
5436 // (float)int <= 4.4 --> int <= 4
5437 // (float)int <= -4.4 --> false
5438 if (RHS.isNegative())
5439 return replaceInstUsesWith(I, Builder.getFalse());
5441 case ICmpInst::ICMP_SLE:
5442 // (float)int <= 4.4 --> int <= 4
5443 // (float)int <= -4.4 --> int < -4
5444 if (RHS.isNegative())
5445 Pred = ICmpInst::ICMP_SLT;
5447 case ICmpInst::ICMP_ULT:
5448 // (float)int < -4.4 --> false
5449 // (float)int < 4.4 --> int <= 4
5450 if (RHS.isNegative())
5451 return replaceInstUsesWith(I, Builder.getFalse());
5452 Pred = ICmpInst::ICMP_ULE;
5454 case ICmpInst::ICMP_SLT:
5455 // (float)int < -4.4 --> int < -4
5456 // (float)int < 4.4 --> int <= 4
5457 if (!RHS.isNegative())
5458 Pred = ICmpInst::ICMP_SLE;
5460 case ICmpInst::ICMP_UGT:
5461 // (float)int > 4.4 --> int > 4
5462 // (float)int > -4.4 --> true
5463 if (RHS.isNegative())
5464 return replaceInstUsesWith(I, Builder.getTrue());
5466 case ICmpInst::ICMP_SGT:
5467 // (float)int > 4.4 --> int > 4
5468 // (float)int > -4.4 --> int >= -4
5469 if (RHS.isNegative())
5470 Pred = ICmpInst::ICMP_SGE;
5472 case ICmpInst::ICMP_UGE:
5473 // (float)int >= -4.4 --> true
5474 // (float)int >= 4.4 --> int > 4
5475 if (RHS.isNegative())
5476 return replaceInstUsesWith(I, Builder.getTrue());
5477 Pred = ICmpInst::ICMP_UGT;
5479 case ICmpInst::ICMP_SGE:
5480 // (float)int >= -4.4 --> int >= -4
5481 // (float)int >= 4.4 --> int > 4
5482 if (!RHS.isNegative())
5483 Pred = ICmpInst::ICMP_SGT;
5489 // Lower this FP comparison into an appropriate integer version of the
5491 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5494 /// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary.
5495 static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI,
5497 // When C is not 0.0 and infinities are not allowed:
5498 // (C / X) < 0.0 is a sign-bit test of X
5499 // (C / X) < 0.0 --> X < 0.0 (if C is positive)
5500 // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate)
5503 // Multiply (C / X) < 0.0 by X * X / C.
5504 // - X is non zero, if it is the flag 'ninf' is violated.
5505 // - C defines the sign of X * X * C. Thus it also defines whether to swap
5506 // the predicate. C is also non zero by definition.
5508 // Thus X * X / C is non zero and the transformation is valid. [qed]
5510 FCmpInst::Predicate Pred = I.getPredicate();
5512 // Check that predicates are valid.
5513 if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) &&
5514 (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE))
5517 // Check that RHS operand is zero.
5518 if (!match(RHSC, m_AnyZeroFP()))
5521 // Check fastmath flags ('ninf').
5522 if (!LHSI->hasNoInfs() || !I.hasNoInfs())
5525 // Check the properties of the dividend. It must not be zero to avoid a
5526 // division by zero (see Proof).
5528 if (!match(LHSI->getOperand(0), m_APFloat(C)))
5534 // Get swapped predicate if necessary.
5535 if (C->isNegative())
5536 Pred = I.getSwappedPredicate();
5538 return new FCmpInst(Pred, LHSI->getOperand(1), RHSC, "", &I);
5541 /// Optimize fabs(X) compared with zero.
5542 static Instruction *foldFabsWithFcmpZero(FCmpInst &I) {
5544 if (!match(I.getOperand(0), m_Intrinsic<Intrinsic::fabs>(m_Value(X))) ||
5545 !match(I.getOperand(1), m_PosZeroFP()))
5548 auto replacePredAndOp0 = [](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
5550 I->setOperand(0, X);
5554 switch (I.getPredicate()) {
5555 case FCmpInst::FCMP_UGE:
5556 case FCmpInst::FCMP_OLT:
5557 // fabs(X) >= 0.0 --> true
5558 // fabs(X) < 0.0 --> false
5559 llvm_unreachable("fcmp should have simplified");
5561 case FCmpInst::FCMP_OGT:
5562 // fabs(X) > 0.0 --> X != 0.0
5563 return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
5565 case FCmpInst::FCMP_UGT:
5566 // fabs(X) u> 0.0 --> X u!= 0.0
5567 return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
5569 case FCmpInst::FCMP_OLE:
5570 // fabs(X) <= 0.0 --> X == 0.0
5571 return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
5573 case FCmpInst::FCMP_ULE:
5574 // fabs(X) u<= 0.0 --> X u== 0.0
5575 return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
5577 case FCmpInst::FCMP_OGE:
5578 // fabs(X) >= 0.0 --> !isnan(X)
5579 assert(!I.hasNoNaNs() && "fcmp should have simplified");
5580 return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
5582 case FCmpInst::FCMP_ULT:
5583 // fabs(X) u< 0.0 --> isnan(X)
5584 assert(!I.hasNoNaNs() && "fcmp should have simplified");
5585 return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
5587 case FCmpInst::FCMP_OEQ:
5588 case FCmpInst::FCMP_UEQ:
5589 case FCmpInst::FCMP_ONE:
5590 case FCmpInst::FCMP_UNE:
5591 case FCmpInst::FCMP_ORD:
5592 case FCmpInst::FCMP_UNO:
5593 // Look through the fabs() because it doesn't change anything but the sign.
5594 // fabs(X) == 0.0 --> X == 0.0,
5595 // fabs(X) != 0.0 --> X != 0.0
5596 // isnan(fabs(X)) --> isnan(X)
5597 // !isnan(fabs(X) --> !isnan(X)
5598 return replacePredAndOp0(&I, I.getPredicate(), X);
5605 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5606 bool Changed = false;
5608 /// Orders the operands of the compare so that they are listed from most
5609 /// complex to least complex. This puts constants before unary operators,
5610 /// before binary operators.
5611 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
5616 const CmpInst::Predicate Pred = I.getPredicate();
5617 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5618 if (Value *V = SimplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(),
5619 SQ.getWithInstruction(&I)))
5620 return replaceInstUsesWith(I, V);
5622 // Simplify 'fcmp pred X, X'
5623 Type *OpType = Op0->getType();
5624 assert(OpType == Op1->getType() && "fcmp with different-typed operands?");
5628 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5629 case FCmpInst::FCMP_ULT: // True if unordered or less than
5630 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5631 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5632 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5633 I.setPredicate(FCmpInst::FCMP_UNO);
5634 I.setOperand(1, Constant::getNullValue(OpType));
5637 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5638 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5639 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5640 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5641 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5642 I.setPredicate(FCmpInst::FCMP_ORD);
5643 I.setOperand(1, Constant::getNullValue(OpType));
5648 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
5649 // then canonicalize the operand to 0.0.
5650 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
5651 if (!match(Op0, m_PosZeroFP()) && isKnownNeverNaN(Op0, &TLI)) {
5652 I.setOperand(0, ConstantFP::getNullValue(OpType));
5655 if (!match(Op1, m_PosZeroFP()) && isKnownNeverNaN(Op1, &TLI)) {
5656 I.setOperand(1, ConstantFP::getNullValue(OpType));
5661 // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y
5663 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
5664 return new FCmpInst(I.getSwappedPredicate(), X, Y, "", &I);
5666 // Test if the FCmpInst instruction is used exclusively by a select as
5667 // part of a minimum or maximum operation. If so, refrain from doing
5668 // any other folding. This helps out other analyses which understand
5669 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5670 // and CodeGen. And in this case, at least one of the comparison
5671 // operands has at least one user besides the compare (the select),
5672 // which would often largely negate the benefit of folding anyway.
5674 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
5676 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
5677 if (SPR.Flavor != SPF_UNKNOWN)
5681 // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0:
5682 // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0
5683 if (match(Op1, m_AnyZeroFP()) && !match(Op1, m_PosZeroFP())) {
5684 I.setOperand(1, ConstantFP::getNullValue(OpType));
5688 // Handle fcmp with instruction LHS and constant RHS.
5691 if (match(Op0, m_Instruction(LHSI)) && match(Op1, m_Constant(RHSC))) {
5692 switch (LHSI->getOpcode()) {
5693 case Instruction::PHI:
5694 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5695 // block. If in the same block, we're encouraging jump threading. If
5696 // not, we are just pessimizing the code by making an i1 phi.
5697 if (LHSI->getParent() == I.getParent())
5698 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
5701 case Instruction::SIToFP:
5702 case Instruction::UIToFP:
5703 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
5706 case Instruction::FDiv:
5707 if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC))
5710 case Instruction::Load:
5711 if (auto *GEP = dyn_cast<GetElementPtrInst>(LHSI->getOperand(0)))
5712 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
5713 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
5714 !cast<LoadInst>(LHSI)->isVolatile())
5715 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
5721 if (Instruction *R = foldFabsWithFcmpZero(I))
5724 if (match(Op0, m_FNeg(m_Value(X)))) {
5725 // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
5727 if (match(Op1, m_Constant(C))) {
5728 Constant *NegC = ConstantExpr::getFNeg(C);
5729 return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
5733 if (match(Op0, m_FPExt(m_Value(X)))) {
5734 // fcmp (fpext X), (fpext Y) -> fcmp X, Y
5735 if (match(Op1, m_FPExt(m_Value(Y))) && X->getType() == Y->getType())
5736 return new FCmpInst(Pred, X, Y, "", &I);
5738 // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless
5740 if (match(Op1, m_APFloat(C))) {
5741 const fltSemantics &FPSem =
5742 X->getType()->getScalarType()->getFltSemantics();
5744 APFloat TruncC = *C;
5745 TruncC.convert(FPSem, APFloat::rmNearestTiesToEven, &Lossy);
5747 // Avoid lossy conversions and denormals.
5748 // Zero is a special case that's OK to convert.
5749 APFloat Fabs = TruncC;
5752 ((Fabs.compare(APFloat::getSmallestNormalized(FPSem)) !=
5753 APFloat::cmpLessThan) || Fabs.isZero())) {
5754 Constant *NewC = ConstantFP::get(X->getType(), TruncC);
5755 return new FCmpInst(Pred, X, NewC, "", &I);
5760 if (I.getType()->isVectorTy())
5761 if (Instruction *Res = foldVectorCmp(I, Builder))
5764 return Changed ? &I : nullptr;