1 //===- InstCombineCompares.cpp --------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitICmp and visitFCmp functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/APSInt.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/ConstantFolding.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/TargetLibraryInfo.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/GetElementPtrTypeIterator.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/KnownBits.h"
30 using namespace PatternMatch;
32 #define DEBUG_TYPE "instcombine"
34 // How many times is a select replaced by one of its operands?
35 STATISTIC(NumSel, "Number of select opts");
38 /// Compute Result = In1+In2, returning true if the result overflowed for this
40 static bool addWithOverflow(APInt &Result, const APInt &In1,
41 const APInt &In2, bool IsSigned = false) {
44 Result = In1.sadd_ov(In2, Overflow);
46 Result = In1.uadd_ov(In2, Overflow);
51 /// Compute Result = In1-In2, returning true if the result overflowed for this
53 static bool subWithOverflow(APInt &Result, const APInt &In1,
54 const APInt &In2, bool IsSigned = false) {
57 Result = In1.ssub_ov(In2, Overflow);
59 Result = In1.usub_ov(In2, Overflow);
64 /// Given an icmp instruction, return true if any use of this comparison is a
65 /// branch on sign bit comparison.
66 static bool hasBranchUse(ICmpInst &I) {
67 for (auto *U : I.users())
68 if (isa<BranchInst>(U))
73 /// Given an exploded icmp instruction, return true if the comparison only
74 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
75 /// result of the comparison is true when the input value is signed.
76 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
79 case ICmpInst::ICMP_SLT: // True if LHS s< 0
81 return RHS.isNullValue();
82 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
84 return RHS.isAllOnesValue();
85 case ICmpInst::ICMP_SGT: // True if LHS s> -1
87 return RHS.isAllOnesValue();
88 case ICmpInst::ICMP_UGT:
89 // True if LHS u> RHS and RHS == high-bit-mask - 1
91 return RHS.isMaxSignedValue();
92 case ICmpInst::ICMP_UGE:
93 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
95 return RHS.isSignMask();
101 /// Returns true if the exploded icmp can be expressed as a signed comparison
102 /// to zero and updates the predicate accordingly.
103 /// The signedness of the comparison is preserved.
104 /// TODO: Refactor with decomposeBitTestICmp()?
105 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
106 if (!ICmpInst::isSigned(Pred))
110 return ICmpInst::isRelational(Pred);
112 if (C.isOneValue()) {
113 if (Pred == ICmpInst::ICMP_SLT) {
114 Pred = ICmpInst::ICMP_SLE;
117 } else if (C.isAllOnesValue()) {
118 if (Pred == ICmpInst::ICMP_SGT) {
119 Pred = ICmpInst::ICMP_SGE;
127 /// Given a signed integer type and a set of known zero and one bits, compute
128 /// the maximum and minimum values that could have the specified known zero and
129 /// known one bits, returning them in Min/Max.
130 /// TODO: Move to method on KnownBits struct?
131 static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known,
132 APInt &Min, APInt &Max) {
133 assert(Known.getBitWidth() == Min.getBitWidth() &&
134 Known.getBitWidth() == Max.getBitWidth() &&
135 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
136 APInt UnknownBits = ~(Known.Zero|Known.One);
138 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
139 // bit if it is unknown.
141 Max = Known.One|UnknownBits;
143 if (UnknownBits.isNegative()) { // Sign bit is unknown
149 /// Given an unsigned integer type and a set of known zero and one bits, compute
150 /// the maximum and minimum values that could have the specified known zero and
151 /// known one bits, returning them in Min/Max.
152 /// TODO: Move to method on KnownBits struct?
153 static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known,
154 APInt &Min, APInt &Max) {
155 assert(Known.getBitWidth() == Min.getBitWidth() &&
156 Known.getBitWidth() == Max.getBitWidth() &&
157 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
158 APInt UnknownBits = ~(Known.Zero|Known.One);
160 // The minimum value is when the unknown bits are all zeros.
162 // The maximum value is when the unknown bits are all ones.
163 Max = Known.One|UnknownBits;
166 /// This is called when we see this pattern:
167 /// cmp pred (load (gep GV, ...)), cmpcst
168 /// where GV is a global variable with a constant initializer. Try to simplify
169 /// this into some simple computation that does not need the load. For example
170 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
172 /// If AndCst is non-null, then the loaded value is masked with that constant
173 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
174 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
177 ConstantInt *AndCst) {
178 Constant *Init = GV->getInitializer();
179 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
182 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
183 // Don't blow up on huge arrays.
184 if (ArrayElementCount > MaxArraySizeForCombine)
187 // There are many forms of this optimization we can handle, for now, just do
188 // the simple index into a single-dimensional array.
190 // Require: GEP GV, 0, i {{, constant indices}}
191 if (GEP->getNumOperands() < 3 ||
192 !isa<ConstantInt>(GEP->getOperand(1)) ||
193 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
194 isa<Constant>(GEP->getOperand(2)))
197 // Check that indices after the variable are constants and in-range for the
198 // type they index. Collect the indices. This is typically for arrays of
200 SmallVector<unsigned, 4> LaterIndices;
202 Type *EltTy = Init->getType()->getArrayElementType();
203 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
204 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
205 if (!Idx) return nullptr; // Variable index.
207 uint64_t IdxVal = Idx->getZExtValue();
208 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
210 if (StructType *STy = dyn_cast<StructType>(EltTy))
211 EltTy = STy->getElementType(IdxVal);
212 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
213 if (IdxVal >= ATy->getNumElements()) return nullptr;
214 EltTy = ATy->getElementType();
216 return nullptr; // Unknown type.
219 LaterIndices.push_back(IdxVal);
222 enum { Overdefined = -3, Undefined = -2 };
224 // Variables for our state machines.
226 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
227 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
228 // and 87 is the second (and last) index. FirstTrueElement is -2 when
229 // undefined, otherwise set to the first true element. SecondTrueElement is
230 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
231 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
233 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
234 // form "i != 47 & i != 87". Same state transitions as for true elements.
235 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
237 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
238 /// define a state machine that triggers for ranges of values that the index
239 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
240 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
241 /// index in the range (inclusive). We use -2 for undefined here because we
242 /// use relative comparisons and don't want 0-1 to match -1.
243 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
245 // MagicBitvector - This is a magic bitvector where we set a bit if the
246 // comparison is true for element 'i'. If there are 64 elements or less in
247 // the array, this will fully represent all the comparison results.
248 uint64_t MagicBitvector = 0;
250 // Scan the array and see if one of our patterns matches.
251 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
252 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
253 Constant *Elt = Init->getAggregateElement(i);
254 if (!Elt) return nullptr;
256 // If this is indexing an array of structures, get the structure element.
257 if (!LaterIndices.empty())
258 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
260 // If the element is masked, handle it.
261 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
263 // Find out if the comparison would be true or false for the i'th element.
264 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
265 CompareRHS, DL, &TLI);
266 // If the result is undef for this element, ignore it.
267 if (isa<UndefValue>(C)) {
268 // Extend range state machines to cover this element in case there is an
269 // undef in the middle of the range.
270 if (TrueRangeEnd == (int)i-1)
272 if (FalseRangeEnd == (int)i-1)
277 // If we can't compute the result for any of the elements, we have to give
278 // up evaluating the entire conditional.
279 if (!isa<ConstantInt>(C)) return nullptr;
281 // Otherwise, we know if the comparison is true or false for this element,
282 // update our state machines.
283 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
285 // State machine for single/double/range index comparison.
287 // Update the TrueElement state machine.
288 if (FirstTrueElement == Undefined)
289 FirstTrueElement = TrueRangeEnd = i; // First true element.
291 // Update double-compare state machine.
292 if (SecondTrueElement == Undefined)
293 SecondTrueElement = i;
295 SecondTrueElement = Overdefined;
297 // Update range state machine.
298 if (TrueRangeEnd == (int)i-1)
301 TrueRangeEnd = Overdefined;
304 // Update the FalseElement state machine.
305 if (FirstFalseElement == Undefined)
306 FirstFalseElement = FalseRangeEnd = i; // First false element.
308 // Update double-compare state machine.
309 if (SecondFalseElement == Undefined)
310 SecondFalseElement = i;
312 SecondFalseElement = Overdefined;
314 // Update range state machine.
315 if (FalseRangeEnd == (int)i-1)
318 FalseRangeEnd = Overdefined;
322 // If this element is in range, update our magic bitvector.
323 if (i < 64 && IsTrueForElt)
324 MagicBitvector |= 1ULL << i;
326 // If all of our states become overdefined, bail out early. Since the
327 // predicate is expensive, only check it every 8 elements. This is only
328 // really useful for really huge arrays.
329 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
330 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
331 FalseRangeEnd == Overdefined)
335 // Now that we've scanned the entire array, emit our new comparison(s). We
336 // order the state machines in complexity of the generated code.
337 Value *Idx = GEP->getOperand(2);
339 // If the index is larger than the pointer size of the target, truncate the
340 // index down like the GEP would do implicitly. We don't have to do this for
341 // an inbounds GEP because the index can't be out of range.
342 if (!GEP->isInBounds()) {
343 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
344 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
345 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
346 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
349 // If the comparison is only true for one or two elements, emit direct
351 if (SecondTrueElement != Overdefined) {
352 // None true -> false.
353 if (FirstTrueElement == Undefined)
354 return replaceInstUsesWith(ICI, Builder.getFalse());
356 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
358 // True for one element -> 'i == 47'.
359 if (SecondTrueElement == Undefined)
360 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
362 // True for two elements -> 'i == 47 | i == 72'.
363 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
364 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
365 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
366 return BinaryOperator::CreateOr(C1, C2);
369 // If the comparison is only false for one or two elements, emit direct
371 if (SecondFalseElement != Overdefined) {
372 // None false -> true.
373 if (FirstFalseElement == Undefined)
374 return replaceInstUsesWith(ICI, Builder.getTrue());
376 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
378 // False for one element -> 'i != 47'.
379 if (SecondFalseElement == Undefined)
380 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
382 // False for two elements -> 'i != 47 & i != 72'.
383 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
384 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
385 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
386 return BinaryOperator::CreateAnd(C1, C2);
389 // If the comparison can be replaced with a range comparison for the elements
390 // where it is true, emit the range check.
391 if (TrueRangeEnd != Overdefined) {
392 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
394 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
395 if (FirstTrueElement) {
396 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
397 Idx = Builder.CreateAdd(Idx, Offs);
400 Value *End = ConstantInt::get(Idx->getType(),
401 TrueRangeEnd-FirstTrueElement+1);
402 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
405 // False range check.
406 if (FalseRangeEnd != Overdefined) {
407 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
408 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
409 if (FirstFalseElement) {
410 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
411 Idx = Builder.CreateAdd(Idx, Offs);
414 Value *End = ConstantInt::get(Idx->getType(),
415 FalseRangeEnd-FirstFalseElement);
416 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
419 // If a magic bitvector captures the entire comparison state
420 // of this load, replace it with computation that does:
421 // ((magic_cst >> i) & 1) != 0
425 // Look for an appropriate type:
426 // - The type of Idx if the magic fits
427 // - The smallest fitting legal type
428 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
431 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
434 Value *V = Builder.CreateIntCast(Idx, Ty, false);
435 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
436 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
437 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
444 /// Return a value that can be used to compare the *offset* implied by a GEP to
445 /// zero. For example, if we have &A[i], we want to return 'i' for
446 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
447 /// are involved. The above expression would also be legal to codegen as
448 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
449 /// This latter form is less amenable to optimization though, and we are allowed
450 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
452 /// If we can't emit an optimized form for this expression, this returns null.
454 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC,
455 const DataLayout &DL) {
456 gep_type_iterator GTI = gep_type_begin(GEP);
458 // Check to see if this gep only has a single variable index. If so, and if
459 // any constant indices are a multiple of its scale, then we can compute this
460 // in terms of the scale of the variable index. For example, if the GEP
461 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
462 // because the expression will cross zero at the same point.
463 unsigned i, e = GEP->getNumOperands();
465 for (i = 1; i != e; ++i, ++GTI) {
466 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
467 // Compute the aggregate offset of constant indices.
468 if (CI->isZero()) continue;
470 // Handle a struct index, which adds its field offset to the pointer.
471 if (StructType *STy = GTI.getStructTypeOrNull()) {
472 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
474 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
475 Offset += Size*CI->getSExtValue();
478 // Found our variable index.
483 // If there are no variable indices, we must have a constant offset, just
484 // evaluate it the general way.
485 if (i == e) return nullptr;
487 Value *VariableIdx = GEP->getOperand(i);
488 // Determine the scale factor of the variable element. For example, this is
489 // 4 if the variable index is into an array of i32.
490 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
492 // Verify that there are no other variable indices. If so, emit the hard way.
493 for (++i, ++GTI; i != e; ++i, ++GTI) {
494 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
495 if (!CI) return nullptr;
497 // Compute the aggregate offset of constant indices.
498 if (CI->isZero()) continue;
500 // Handle a struct index, which adds its field offset to the pointer.
501 if (StructType *STy = GTI.getStructTypeOrNull()) {
502 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
504 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
505 Offset += Size*CI->getSExtValue();
509 // Okay, we know we have a single variable index, which must be a
510 // pointer/array/vector index. If there is no offset, life is simple, return
512 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
513 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
515 // Cast to intptrty in case a truncation occurs. If an extension is needed,
516 // we don't need to bother extending: the extension won't affect where the
517 // computation crosses zero.
518 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
519 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
524 // Otherwise, there is an index. The computation we will do will be modulo
526 Offset = SignExtend64(Offset, IntPtrWidth);
527 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
529 // To do this transformation, any constant index must be a multiple of the
530 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
531 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
532 // multiple of the variable scale.
533 int64_t NewOffs = Offset / (int64_t)VariableScale;
534 if (Offset != NewOffs*(int64_t)VariableScale)
537 // Okay, we can do this evaluation. Start by converting the index to intptr.
538 if (VariableIdx->getType() != IntPtrTy)
539 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
541 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
542 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
545 /// Returns true if we can rewrite Start as a GEP with pointer Base
546 /// and some integer offset. The nodes that need to be re-written
547 /// for this transformation will be added to Explored.
548 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
549 const DataLayout &DL,
550 SetVector<Value *> &Explored) {
551 SmallVector<Value *, 16> WorkList(1, Start);
552 Explored.insert(Base);
554 // The following traversal gives us an order which can be used
555 // when doing the final transformation. Since in the final
556 // transformation we create the PHI replacement instructions first,
557 // we don't have to get them in any particular order.
559 // However, for other instructions we will have to traverse the
560 // operands of an instruction first, which means that we have to
561 // do a post-order traversal.
562 while (!WorkList.empty()) {
563 SetVector<PHINode *> PHIs;
565 while (!WorkList.empty()) {
566 if (Explored.size() >= 100)
569 Value *V = WorkList.back();
571 if (Explored.count(V) != 0) {
576 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
577 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
578 // We've found some value that we can't explore which is different from
579 // the base. Therefore we can't do this transformation.
582 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
583 auto *CI = dyn_cast<CastInst>(V);
584 if (!CI->isNoopCast(DL))
587 if (Explored.count(CI->getOperand(0)) == 0)
588 WorkList.push_back(CI->getOperand(0));
591 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
592 // We're limiting the GEP to having one index. This will preserve
593 // the original pointer type. We could handle more cases in the
595 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
596 GEP->getType() != Start->getType())
599 if (Explored.count(GEP->getOperand(0)) == 0)
600 WorkList.push_back(GEP->getOperand(0));
603 if (WorkList.back() == V) {
605 // We've finished visiting this node, mark it as such.
609 if (auto *PN = dyn_cast<PHINode>(V)) {
610 // We cannot transform PHIs on unsplittable basic blocks.
611 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
618 // Explore the PHI nodes further.
619 for (auto *PN : PHIs)
620 for (Value *Op : PN->incoming_values())
621 if (Explored.count(Op) == 0)
622 WorkList.push_back(Op);
625 // Make sure that we can do this. Since we can't insert GEPs in a basic
626 // block before a PHI node, we can't easily do this transformation if
627 // we have PHI node users of transformed instructions.
628 for (Value *Val : Explored) {
629 for (Value *Use : Val->uses()) {
631 auto *PHI = dyn_cast<PHINode>(Use);
632 auto *Inst = dyn_cast<Instruction>(Val);
634 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
635 Explored.count(PHI) == 0)
638 if (PHI->getParent() == Inst->getParent())
645 // Sets the appropriate insert point on Builder where we can add
646 // a replacement Instruction for V (if that is possible).
647 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
648 bool Before = true) {
649 if (auto *PHI = dyn_cast<PHINode>(V)) {
650 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
653 if (auto *I = dyn_cast<Instruction>(V)) {
655 I = &*std::next(I->getIterator());
656 Builder.SetInsertPoint(I);
659 if (auto *A = dyn_cast<Argument>(V)) {
660 // Set the insertion point in the entry block.
661 BasicBlock &Entry = A->getParent()->getEntryBlock();
662 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
665 // Otherwise, this is a constant and we don't need to set a new
667 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
670 /// Returns a re-written value of Start as an indexed GEP using Base as a
672 static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
673 const DataLayout &DL,
674 SetVector<Value *> &Explored) {
675 // Perform all the substitutions. This is a bit tricky because we can
676 // have cycles in our use-def chains.
677 // 1. Create the PHI nodes without any incoming values.
678 // 2. Create all the other values.
679 // 3. Add the edges for the PHI nodes.
680 // 4. Emit GEPs to get the original pointers.
681 // 5. Remove the original instructions.
682 Type *IndexType = IntegerType::get(
683 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
685 DenseMap<Value *, Value *> NewInsts;
686 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
688 // Create the new PHI nodes, without adding any incoming values.
689 for (Value *Val : Explored) {
692 // Create empty phi nodes. This avoids cyclic dependencies when creating
693 // the remaining instructions.
694 if (auto *PHI = dyn_cast<PHINode>(Val))
695 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
696 PHI->getName() + ".idx", PHI);
698 IRBuilder<> Builder(Base->getContext());
700 // Create all the other instructions.
701 for (Value *Val : Explored) {
703 if (NewInsts.find(Val) != NewInsts.end())
706 if (auto *CI = dyn_cast<CastInst>(Val)) {
707 NewInsts[CI] = NewInsts[CI->getOperand(0)];
710 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
711 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
712 : GEP->getOperand(1);
713 setInsertionPoint(Builder, GEP);
714 // Indices might need to be sign extended. GEPs will magically do
715 // this, but we need to do it ourselves here.
716 if (Index->getType()->getScalarSizeInBits() !=
717 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
718 Index = Builder.CreateSExtOrTrunc(
719 Index, NewInsts[GEP->getOperand(0)]->getType(),
720 GEP->getOperand(0)->getName() + ".sext");
723 auto *Op = NewInsts[GEP->getOperand(0)];
724 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
725 NewInsts[GEP] = Index;
727 NewInsts[GEP] = Builder.CreateNSWAdd(
728 Op, Index, GEP->getOperand(0)->getName() + ".add");
731 if (isa<PHINode>(Val))
734 llvm_unreachable("Unexpected instruction type");
737 // Add the incoming values to the PHI nodes.
738 for (Value *Val : Explored) {
741 // All the instructions have been created, we can now add edges to the
743 if (auto *PHI = dyn_cast<PHINode>(Val)) {
744 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
745 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
746 Value *NewIncoming = PHI->getIncomingValue(I);
748 if (NewInsts.find(NewIncoming) != NewInsts.end())
749 NewIncoming = NewInsts[NewIncoming];
751 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
756 for (Value *Val : Explored) {
760 // Depending on the type, for external users we have to emit
761 // a GEP or a GEP + ptrtoint.
762 setInsertionPoint(Builder, Val, false);
764 // If required, create an inttoptr instruction for Base.
765 Value *NewBase = Base;
766 if (!Base->getType()->isPointerTy())
767 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
768 Start->getName() + "to.ptr");
770 Value *GEP = Builder.CreateInBoundsGEP(
771 Start->getType()->getPointerElementType(), NewBase,
772 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
774 if (!Val->getType()->isPointerTy()) {
775 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
776 Val->getName() + ".conv");
779 Val->replaceAllUsesWith(GEP);
782 return NewInsts[Start];
785 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
786 /// the input Value as a constant indexed GEP. Returns a pair containing
787 /// the GEPs Pointer and Index.
788 static std::pair<Value *, Value *>
789 getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
790 Type *IndexType = IntegerType::get(V->getContext(),
791 DL.getIndexTypeSizeInBits(V->getType()));
793 Constant *Index = ConstantInt::getNullValue(IndexType);
795 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
796 // We accept only inbouds GEPs here to exclude the possibility of
798 if (!GEP->isInBounds())
800 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
801 GEP->getType() == V->getType()) {
802 V = GEP->getOperand(0);
803 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
804 Index = ConstantExpr::getAdd(
805 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
810 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
811 if (!CI->isNoopCast(DL))
813 V = CI->getOperand(0);
816 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
817 if (!CI->isNoopCast(DL))
819 V = CI->getOperand(0);
827 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
828 /// We can look through PHIs, GEPs and casts in order to determine a common base
829 /// between GEPLHS and RHS.
830 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
831 ICmpInst::Predicate Cond,
832 const DataLayout &DL) {
833 if (!GEPLHS->hasAllConstantIndices())
836 // Make sure the pointers have the same type.
837 if (GEPLHS->getType() != RHS->getType())
840 Value *PtrBase, *Index;
841 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
843 // The set of nodes that will take part in this transformation.
844 SetVector<Value *> Nodes;
846 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
849 // We know we can re-write this as
850 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
851 // Since we've only looked through inbouds GEPs we know that we
852 // can't have overflow on either side. We can therefore re-write
854 // OFFSET1 cmp OFFSET2
855 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
857 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
858 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
859 // offset. Since Index is the offset of LHS to the base pointer, we will now
860 // compare the offsets instead of comparing the pointers.
861 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
864 /// Fold comparisons between a GEP instruction and something else. At this point
865 /// we know that the GEP is on the LHS of the comparison.
866 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
867 ICmpInst::Predicate Cond,
869 // Don't transform signed compares of GEPs into index compares. Even if the
870 // GEP is inbounds, the final add of the base pointer can have signed overflow
871 // and would change the result of the icmp.
872 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
873 // the maximum signed value for the pointer type.
874 if (ICmpInst::isSigned(Cond))
877 // Look through bitcasts and addrspacecasts. We do not however want to remove
879 if (!isa<GetElementPtrInst>(RHS))
880 RHS = RHS->stripPointerCasts();
882 Value *PtrBase = GEPLHS->getOperand(0);
883 if (PtrBase == RHS && GEPLHS->isInBounds()) {
884 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
885 // This transformation (ignoring the base and scales) is valid because we
886 // know pointers can't overflow since the gep is inbounds. See if we can
887 // output an optimized form.
888 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
890 // If not, synthesize the offset the hard way.
892 Offset = EmitGEPOffset(GEPLHS);
893 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
894 Constant::getNullValue(Offset->getType()));
895 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
896 // If the base pointers are different, but the indices are the same, just
897 // compare the base pointer.
898 if (PtrBase != GEPRHS->getOperand(0)) {
899 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
900 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
901 GEPRHS->getOperand(0)->getType();
903 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
904 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
905 IndicesTheSame = false;
909 // If all indices are the same, just compare the base pointers.
910 Type *BaseType = GEPLHS->getOperand(0)->getType();
911 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
912 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
914 // If we're comparing GEPs with two base pointers that only differ in type
915 // and both GEPs have only constant indices or just one use, then fold
916 // the compare with the adjusted indices.
917 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
918 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
919 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
920 PtrBase->stripPointerCasts() ==
921 GEPRHS->getOperand(0)->stripPointerCasts()) {
922 Value *LOffset = EmitGEPOffset(GEPLHS);
923 Value *ROffset = EmitGEPOffset(GEPRHS);
925 // If we looked through an addrspacecast between different sized address
926 // spaces, the LHS and RHS pointers are different sized
927 // integers. Truncate to the smaller one.
928 Type *LHSIndexTy = LOffset->getType();
929 Type *RHSIndexTy = ROffset->getType();
930 if (LHSIndexTy != RHSIndexTy) {
931 if (LHSIndexTy->getPrimitiveSizeInBits() <
932 RHSIndexTy->getPrimitiveSizeInBits()) {
933 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
935 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
938 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
940 return replaceInstUsesWith(I, Cmp);
943 // Otherwise, the base pointers are different and the indices are
944 // different. Try convert this to an indexed compare by looking through
946 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
949 // If one of the GEPs has all zero indices, recurse.
950 if (GEPLHS->hasAllZeroIndices())
951 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
952 ICmpInst::getSwappedPredicate(Cond), I);
954 // If the other GEP has all zero indices, recurse.
955 if (GEPRHS->hasAllZeroIndices())
956 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
958 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
959 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
960 // If the GEPs only differ by one index, compare it.
961 unsigned NumDifferences = 0; // Keep track of # differences.
962 unsigned DiffOperand = 0; // The operand that differs.
963 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
964 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
965 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
966 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
967 // Irreconcilable differences.
971 if (NumDifferences++) break;
976 if (NumDifferences == 0) // SAME GEP?
977 return replaceInstUsesWith(I, // No comparison is needed here.
978 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
980 else if (NumDifferences == 1 && GEPsInBounds) {
981 Value *LHSV = GEPLHS->getOperand(DiffOperand);
982 Value *RHSV = GEPRHS->getOperand(DiffOperand);
983 // Make sure we do a signed comparison here.
984 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
988 // Only lower this if the icmp is the only user of the GEP or if we expect
989 // the result to fold to a constant!
990 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
991 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
992 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
993 Value *L = EmitGEPOffset(GEPLHS);
994 Value *R = EmitGEPOffset(GEPRHS);
995 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
999 // Try convert this to an indexed compare by looking through PHIs/casts as a
1001 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1004 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1005 const AllocaInst *Alloca,
1006 const Value *Other) {
1007 assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1009 // It would be tempting to fold away comparisons between allocas and any
1010 // pointer not based on that alloca (e.g. an argument). However, even
1011 // though such pointers cannot alias, they can still compare equal.
1013 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1014 // doesn't escape we can argue that it's impossible to guess its value, and we
1015 // can therefore act as if any such guesses are wrong.
1017 // The code below checks that the alloca doesn't escape, and that it's only
1018 // used in a comparison once (the current instruction). The
1019 // single-comparison-use condition ensures that we're trivially folding all
1020 // comparisons against the alloca consistently, and avoids the risk of
1021 // erroneously folding a comparison of the pointer with itself.
1023 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1025 SmallVector<const Use *, 32> Worklist;
1026 for (const Use &U : Alloca->uses()) {
1027 if (Worklist.size() >= MaxIter)
1029 Worklist.push_back(&U);
1032 unsigned NumCmps = 0;
1033 while (!Worklist.empty()) {
1034 assert(Worklist.size() <= MaxIter);
1035 const Use *U = Worklist.pop_back_val();
1036 const Value *V = U->getUser();
1039 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1040 isa<SelectInst>(V)) {
1042 } else if (isa<LoadInst>(V)) {
1043 // Loading from the pointer doesn't escape it.
1045 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1046 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1047 if (SI->getValueOperand() == U->get())
1050 } else if (isa<ICmpInst>(V)) {
1052 return nullptr; // Found more than one cmp.
1054 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1055 switch (Intrin->getIntrinsicID()) {
1056 // These intrinsics don't escape or compare the pointer. Memset is safe
1057 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1058 // we don't allow stores, so src cannot point to V.
1059 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1060 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1068 for (const Use &U : V->uses()) {
1069 if (Worklist.size() >= MaxIter)
1071 Worklist.push_back(&U);
1075 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1076 return replaceInstUsesWith(
1078 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1081 /// Fold "icmp pred (X+C), X".
1082 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1083 ICmpInst::Predicate Pred) {
1084 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1085 // so the values can never be equal. Similarly for all other "or equals"
1087 assert(!!C && "C should not be zero!");
1089 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1090 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1091 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1092 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1093 Constant *R = ConstantInt::get(X->getType(),
1094 APInt::getMaxValue(C.getBitWidth()) - C);
1095 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1098 // (X+1) >u X --> X <u (0-1) --> X != 255
1099 // (X+2) >u X --> X <u (0-2) --> X <u 254
1100 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1101 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1102 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1103 ConstantInt::get(X->getType(), -C));
1105 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1107 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1108 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1109 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1110 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1111 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1112 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1113 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1114 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1115 ConstantInt::get(X->getType(), SMax - C));
1117 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1118 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1119 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1120 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1121 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1122 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1124 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1125 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1126 ConstantInt::get(X->getType(), SMax - (C - 1)));
1129 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1130 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1131 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1132 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1135 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1137 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1138 if (I.getPredicate() == I.ICMP_NE)
1139 Pred = CmpInst::getInversePredicate(Pred);
1140 return new ICmpInst(Pred, LHS, RHS);
1143 // Don't bother doing any work for cases which InstSimplify handles.
1144 if (AP2.isNullValue())
1147 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1149 if (AP2.isAllOnesValue())
1151 if (AP2.isNegative() != AP1.isNegative())
1158 // 'A' must be large enough to shift out the highest set bit.
1159 return getICmp(I.ICMP_UGT, A,
1160 ConstantInt::get(A->getType(), AP2.logBase2()));
1163 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1166 if (IsAShr && AP1.isNegative())
1167 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1169 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1172 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1173 // There are multiple solutions if we are comparing against -1 and the LHS
1174 // of the ashr is not a power of two.
1175 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1176 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1177 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1178 } else if (AP1 == AP2.lshr(Shift)) {
1179 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1183 // Shifting const2 will never be equal to const1.
1184 // FIXME: This should always be handled by InstSimplify?
1185 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1186 return replaceInstUsesWith(I, TorF);
1189 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1190 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1191 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1194 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1196 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1197 if (I.getPredicate() == I.ICMP_NE)
1198 Pred = CmpInst::getInversePredicate(Pred);
1199 return new ICmpInst(Pred, LHS, RHS);
1202 // Don't bother doing any work for cases which InstSimplify handles.
1203 if (AP2.isNullValue())
1206 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1208 if (!AP1 && AP2TrailingZeros != 0)
1211 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1214 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1216 // Get the distance between the lowest bits that are set.
1217 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1219 if (Shift > 0 && AP2.shl(Shift) == AP1)
1220 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1222 // Shifting const2 will never be equal to const1.
1223 // FIXME: This should always be handled by InstSimplify?
1224 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1225 return replaceInstUsesWith(I, TorF);
1228 /// The caller has matched a pattern of the form:
1229 /// I = icmp ugt (add (add A, B), CI2), CI1
1230 /// If this is of the form:
1232 /// if (sum+128 >u 255)
1233 /// Then replace it with llvm.sadd.with.overflow.i8.
1235 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1236 ConstantInt *CI2, ConstantInt *CI1,
1238 // The transformation we're trying to do here is to transform this into an
1239 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1240 // with a narrower add, and discard the add-with-constant that is part of the
1241 // range check (if we can't eliminate it, this isn't profitable).
1243 // In order to eliminate the add-with-constant, the compare can be its only
1245 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1246 if (!AddWithCst->hasOneUse())
1249 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1250 if (!CI2->getValue().isPowerOf2())
1252 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1253 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1256 // The width of the new add formed is 1 more than the bias.
1259 // Check to see that CI1 is an all-ones value with NewWidth bits.
1260 if (CI1->getBitWidth() == NewWidth ||
1261 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1264 // This is only really a signed overflow check if the inputs have been
1265 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1266 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1267 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1268 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1269 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1272 // In order to replace the original add with a narrower
1273 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1274 // and truncates that discard the high bits of the add. Verify that this is
1276 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1277 for (User *U : OrigAdd->users()) {
1278 if (U == AddWithCst)
1281 // Only accept truncates for now. We would really like a nice recursive
1282 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1283 // chain to see which bits of a value are actually demanded. If the
1284 // original add had another add which was then immediately truncated, we
1285 // could still do the transformation.
1286 TruncInst *TI = dyn_cast<TruncInst>(U);
1287 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1291 // If the pattern matches, truncate the inputs to the narrower type and
1292 // use the sadd_with_overflow intrinsic to efficiently compute both the
1293 // result and the overflow bit.
1294 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1295 Value *F = Intrinsic::getDeclaration(I.getModule(),
1296 Intrinsic::sadd_with_overflow, NewType);
1298 InstCombiner::BuilderTy &Builder = IC.Builder;
1300 // Put the new code above the original add, in case there are any uses of the
1301 // add between the add and the compare.
1302 Builder.SetInsertPoint(OrigAdd);
1304 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1305 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1306 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1307 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1308 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1310 // The inner add was the result of the narrow add, zero extended to the
1311 // wider type. Replace it with the result computed by the intrinsic.
1312 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1314 // The original icmp gets replaced with the overflow value.
1315 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1318 // Handle (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1319 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1320 CmpInst::Predicate Pred = Cmp.getPredicate();
1321 Value *X = Cmp.getOperand(0);
1323 if (match(Cmp.getOperand(1), m_Zero()) && Pred == ICmpInst::ICMP_SGT) {
1325 SelectPatternResult SPR = matchSelectPattern(X, A, B);
1326 if (SPR.Flavor == SPF_SMIN) {
1327 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1328 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1329 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1330 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1336 /// Fold icmp Pred X, C.
1337 /// TODO: This code structure does not make sense. The saturating add fold
1338 /// should be moved to some other helper and extended as noted below (it is also
1339 /// possible that code has been made unnecessary - do we canonicalize IR to
1340 /// overflow/saturating intrinsics or not?).
1341 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1342 // Match the following pattern, which is a common idiom when writing
1343 // overflow-safe integer arithmetic functions. The source performs an addition
1344 // in wider type and explicitly checks for overflow using comparisons against
1345 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1347 // TODO: This could probably be generalized to handle other overflow-safe
1348 // operations if we worked out the formulas to compute the appropriate magic
1352 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1353 CmpInst::Predicate Pred = Cmp.getPredicate();
1354 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1356 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1357 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1358 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1359 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1365 /// Canonicalize icmp instructions based on dominating conditions.
1366 Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1367 // This is a cheap/incomplete check for dominance - just match a single
1368 // predecessor with a conditional branch.
1369 BasicBlock *CmpBB = Cmp.getParent();
1370 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1375 BasicBlock *TrueBB, *FalseBB;
1376 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1379 assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1380 "Predecessor block does not point to successor?");
1382 // The branch should get simplified. Don't bother simplifying this condition.
1383 if (TrueBB == FalseBB)
1386 // Try to simplify this compare to T/F based on the dominating condition.
1387 Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1389 return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1391 CmpInst::Predicate Pred = Cmp.getPredicate();
1392 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1393 ICmpInst::Predicate DomPred;
1394 const APInt *C, *DomC;
1395 if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1396 match(Y, m_APInt(C))) {
1397 // We have 2 compares of a variable with constants. Calculate the constant
1398 // ranges of those compares to see if we can transform the 2nd compare:
1400 // DomCond = icmp DomPred X, DomC
1401 // br DomCond, CmpBB, FalseBB
1403 // Cmp = icmp Pred X, C
1404 ConstantRange CR = ConstantRange::makeAllowedICmpRegion(Pred, *C);
1405 ConstantRange DominatingCR =
1406 (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1407 : ConstantRange::makeExactICmpRegion(
1408 CmpInst::getInversePredicate(DomPred), *DomC);
1409 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1410 ConstantRange Difference = DominatingCR.difference(CR);
1411 if (Intersection.isEmptySet())
1412 return replaceInstUsesWith(Cmp, Builder.getFalse());
1413 if (Difference.isEmptySet())
1414 return replaceInstUsesWith(Cmp, Builder.getTrue());
1416 // Canonicalizing a sign bit comparison that gets used in a branch,
1417 // pessimizes codegen by generating branch on zero instruction instead
1418 // of a test and branch. So we avoid canonicalizing in such situations
1419 // because test and branch instruction has better branch displacement
1420 // than compare and branch instruction.
1422 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1423 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1426 if (const APInt *EqC = Intersection.getSingleElement())
1427 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1428 if (const APInt *NeC = Difference.getSingleElement())
1429 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1435 /// Fold icmp (trunc X, Y), C.
1436 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1439 ICmpInst::Predicate Pred = Cmp.getPredicate();
1440 Value *X = Trunc->getOperand(0);
1441 if (C.isOneValue() && C.getBitWidth() > 1) {
1442 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1444 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1445 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1446 ConstantInt::get(V->getType(), 1));
1449 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1450 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1451 // of the high bits truncated out of x are known.
1452 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1453 SrcBits = X->getType()->getScalarSizeInBits();
1454 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1456 // If all the high bits are known, we can do this xform.
1457 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1458 // Pull in the high bits from known-ones set.
1459 APInt NewRHS = C.zext(SrcBits);
1460 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1461 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1468 /// Fold icmp (xor X, Y), C.
1469 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1470 BinaryOperator *Xor,
1472 Value *X = Xor->getOperand(0);
1473 Value *Y = Xor->getOperand(1);
1475 if (!match(Y, m_APInt(XorC)))
1478 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1480 ICmpInst::Predicate Pred = Cmp.getPredicate();
1481 bool TrueIfSigned = false;
1482 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1484 // If the sign bit of the XorCst is not set, there is no change to
1485 // the operation, just stop using the Xor.
1486 if (!XorC->isNegative()) {
1487 Cmp.setOperand(0, X);
1492 // Emit the opposite comparison.
1494 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1495 ConstantInt::getAllOnesValue(X->getType()));
1497 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1498 ConstantInt::getNullValue(X->getType()));
1501 if (Xor->hasOneUse()) {
1502 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1503 if (!Cmp.isEquality() && XorC->isSignMask()) {
1504 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1505 : Cmp.getSignedPredicate();
1506 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1509 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1510 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1511 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1512 : Cmp.getSignedPredicate();
1513 Pred = Cmp.getSwappedPredicate(Pred);
1514 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1518 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1519 if (Pred == ICmpInst::ICMP_UGT) {
1520 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1521 if (*XorC == ~C && (C + 1).isPowerOf2())
1522 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1523 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1524 if (*XorC == C && (C + 1).isPowerOf2())
1525 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1527 if (Pred == ICmpInst::ICMP_ULT) {
1528 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1529 if (*XorC == -C && C.isPowerOf2())
1530 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1531 ConstantInt::get(X->getType(), ~C));
1532 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1533 if (*XorC == C && (-C).isPowerOf2())
1534 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1535 ConstantInt::get(X->getType(), ~C));
1540 /// Fold icmp (and (sh X, Y), C2), C1.
1541 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1542 const APInt &C1, const APInt &C2) {
1543 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1544 if (!Shift || !Shift->isShift())
1547 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1548 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1549 // code produced by the clang front-end, for bitfield access.
1550 // This seemingly simple opportunity to fold away a shift turns out to be
1551 // rather complicated. See PR17827 for details.
1552 unsigned ShiftOpcode = Shift->getOpcode();
1553 bool IsShl = ShiftOpcode == Instruction::Shl;
1555 if (match(Shift->getOperand(1), m_APInt(C3))) {
1556 bool CanFold = false;
1557 if (ShiftOpcode == Instruction::Shl) {
1558 // For a left shift, we can fold if the comparison is not signed. We can
1559 // also fold a signed comparison if the mask value and comparison value
1560 // are not negative. These constraints may not be obvious, but we can
1561 // prove that they are correct using an SMT solver.
1562 if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1565 bool IsAshr = ShiftOpcode == Instruction::AShr;
1566 // For a logical right shift, we can fold if the comparison is not signed.
1567 // We can also fold a signed comparison if the shifted mask value and the
1568 // shifted comparison value are not negative. These constraints may not be
1569 // obvious, but we can prove that they are correct using an SMT solver.
1570 // For an arithmetic shift right we can do the same, if we ensure
1571 // the And doesn't use any bits being shifted in. Normally these would
1572 // be turned into lshr by SimplifyDemandedBits, but not if there is an
1574 if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1575 if (!Cmp.isSigned() ||
1576 (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1582 APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1583 APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1584 // Check to see if we are shifting out any of the bits being compared.
1585 if (SameAsC1 != C1) {
1586 // If we shifted bits out, the fold is not going to work out. As a
1587 // special case, check to see if this means that the result is always
1588 // true or false now.
1589 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1590 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1591 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1592 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1594 Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1595 APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1596 And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1597 And->setOperand(0, Shift->getOperand(0));
1598 Worklist.Add(Shift); // Shift is dead.
1604 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1605 // preferable because it allows the C2 << Y expression to be hoisted out of a
1606 // loop if Y is invariant and X is not.
1607 if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1608 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1611 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1612 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1614 // Compute X & (C2 << Y).
1615 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1616 Cmp.setOperand(0, NewAnd);
1623 /// Fold icmp (and X, C2), C1.
1624 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1625 BinaryOperator *And,
1627 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1628 // TODO: We canonicalize to the longer form for scalars because we have
1629 // better analysis/folds for icmp, and codegen may be better with icmp.
1630 if (Cmp.getPredicate() == CmpInst::ICMP_NE && Cmp.getType()->isVectorTy() &&
1631 C1.isNullValue() && match(And->getOperand(1), m_One()))
1632 return new TruncInst(And->getOperand(0), Cmp.getType());
1635 if (!match(And->getOperand(1), m_APInt(C2)))
1638 if (!And->hasOneUse())
1641 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1642 // the input width without changing the value produced, eliminate the cast:
1644 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1646 // We can do this transformation if the constants do not have their sign bits
1647 // set or if it is an equality comparison. Extending a relational comparison
1648 // when we're checking the sign bit would not work.
1650 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1651 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1652 // TODO: Is this a good transform for vectors? Wider types may reduce
1653 // throughput. Should this transform be limited (even for scalars) by using
1654 // shouldChangeType()?
1655 if (!Cmp.getType()->isVectorTy()) {
1656 Type *WideType = W->getType();
1657 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1658 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1659 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1660 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1661 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1665 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1668 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1669 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1671 // iff pred isn't signed
1672 if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1673 match(And->getOperand(1), m_One())) {
1674 Constant *One = cast<Constant>(And->getOperand(1));
1675 Value *Or = And->getOperand(0);
1676 Value *A, *B, *LShr;
1677 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1678 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1679 unsigned UsesRemoved = 0;
1680 if (And->hasOneUse())
1682 if (Or->hasOneUse())
1684 if (LShr->hasOneUse())
1687 // Compute A & ((1 << B) | 1)
1688 Value *NewOr = nullptr;
1689 if (auto *C = dyn_cast<Constant>(B)) {
1690 if (UsesRemoved >= 1)
1691 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1693 if (UsesRemoved >= 3)
1694 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1696 One, Or->getName());
1699 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1700 Cmp.setOperand(0, NewAnd);
1709 /// Fold icmp (and X, Y), C.
1710 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1711 BinaryOperator *And,
1713 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1716 // TODO: These all require that Y is constant too, so refactor with the above.
1718 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1719 Value *X = And->getOperand(0);
1720 Value *Y = And->getOperand(1);
1721 if (auto *LI = dyn_cast<LoadInst>(X))
1722 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1723 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1724 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1725 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1726 ConstantInt *C2 = cast<ConstantInt>(Y);
1727 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1731 if (!Cmp.isEquality())
1734 // X & -C == -C -> X > u ~C
1735 // X & -C != -C -> X <= u ~C
1736 // iff C is a power of 2
1737 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1738 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1739 : CmpInst::ICMP_ULE;
1740 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1743 // (X & C2) == 0 -> (trunc X) >= 0
1744 // (X & C2) != 0 -> (trunc X) < 0
1745 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1747 if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1748 int32_t ExactLogBase2 = C2->exactLogBase2();
1749 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1750 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1751 if (And->getType()->isVectorTy())
1752 NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1753 Value *Trunc = Builder.CreateTrunc(X, NTy);
1754 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1755 : CmpInst::ICMP_SLT;
1756 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1763 /// Fold icmp (or X, Y), C.
1764 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1766 ICmpInst::Predicate Pred = Cmp.getPredicate();
1767 if (C.isOneValue()) {
1768 // icmp slt signum(V) 1 --> icmp slt V, 1
1770 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1771 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1772 ConstantInt::get(V->getType(), 1));
1775 // X | C == C --> X <=u C
1776 // X | C != C --> X >u C
1777 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1778 if (Cmp.isEquality() && Cmp.getOperand(1) == Or->getOperand(1) &&
1779 (C + 1).isPowerOf2()) {
1780 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1781 return new ICmpInst(Pred, Or->getOperand(0), Or->getOperand(1));
1784 if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1788 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1789 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1790 // -> and (icmp eq P, null), (icmp eq Q, null).
1792 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1794 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1795 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1796 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1799 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1800 // a shorter form that has more potential to be folded even further.
1801 Value *X1, *X2, *X3, *X4;
1802 if (match(Or->getOperand(0), m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1803 match(Or->getOperand(1), m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1804 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1805 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1806 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1807 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1808 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1809 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1815 /// Fold icmp (mul X, Y), C.
1816 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1817 BinaryOperator *Mul,
1820 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1823 // If this is a test of the sign bit and the multiply is sign-preserving with
1824 // a constant operand, use the multiply LHS operand instead.
1825 ICmpInst::Predicate Pred = Cmp.getPredicate();
1826 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1827 if (MulC->isNegative())
1828 Pred = ICmpInst::getSwappedPredicate(Pred);
1829 return new ICmpInst(Pred, Mul->getOperand(0),
1830 Constant::getNullValue(Mul->getType()));
1836 /// Fold icmp (shl 1, Y), C.
1837 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
1840 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1843 Type *ShiftType = Shl->getType();
1844 unsigned TypeBits = C.getBitWidth();
1845 bool CIsPowerOf2 = C.isPowerOf2();
1846 ICmpInst::Predicate Pred = Cmp.getPredicate();
1847 if (Cmp.isUnsigned()) {
1848 // (1 << Y) pred C -> Y pred Log2(C)
1850 // (1 << Y) < 30 -> Y <= 4
1851 // (1 << Y) <= 30 -> Y <= 4
1852 // (1 << Y) >= 30 -> Y > 4
1853 // (1 << Y) > 30 -> Y > 4
1854 if (Pred == ICmpInst::ICMP_ULT)
1855 Pred = ICmpInst::ICMP_ULE;
1856 else if (Pred == ICmpInst::ICMP_UGE)
1857 Pred = ICmpInst::ICMP_UGT;
1860 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1861 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1862 unsigned CLog2 = C.logBase2();
1863 if (CLog2 == TypeBits - 1) {
1864 if (Pred == ICmpInst::ICMP_UGE)
1865 Pred = ICmpInst::ICMP_EQ;
1866 else if (Pred == ICmpInst::ICMP_ULT)
1867 Pred = ICmpInst::ICMP_NE;
1869 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1870 } else if (Cmp.isSigned()) {
1871 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1872 if (C.isAllOnesValue()) {
1873 // (1 << Y) <= -1 -> Y == 31
1874 if (Pred == ICmpInst::ICMP_SLE)
1875 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1877 // (1 << Y) > -1 -> Y != 31
1878 if (Pred == ICmpInst::ICMP_SGT)
1879 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1881 // (1 << Y) < 0 -> Y == 31
1882 // (1 << Y) <= 0 -> Y == 31
1883 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1884 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1886 // (1 << Y) >= 0 -> Y != 31
1887 // (1 << Y) > 0 -> Y != 31
1888 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1889 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1891 } else if (Cmp.isEquality() && CIsPowerOf2) {
1892 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
1898 /// Fold icmp (shl X, Y), C.
1899 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1900 BinaryOperator *Shl,
1902 const APInt *ShiftVal;
1903 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1904 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
1906 const APInt *ShiftAmt;
1907 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1908 return foldICmpShlOne(Cmp, Shl, C);
1910 // Check that the shift amount is in range. If not, don't perform undefined
1911 // shifts. When the shift is visited, it will be simplified.
1912 unsigned TypeBits = C.getBitWidth();
1913 if (ShiftAmt->uge(TypeBits))
1916 ICmpInst::Predicate Pred = Cmp.getPredicate();
1917 Value *X = Shl->getOperand(0);
1918 Type *ShType = Shl->getType();
1920 // NSW guarantees that we are only shifting out sign bits from the high bits,
1921 // so we can ASHR the compare constant without needing a mask and eliminate
1923 if (Shl->hasNoSignedWrap()) {
1924 if (Pred == ICmpInst::ICMP_SGT) {
1925 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
1926 APInt ShiftedC = C.ashr(*ShiftAmt);
1927 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1929 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1930 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
1931 APInt ShiftedC = C.ashr(*ShiftAmt);
1932 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1934 if (Pred == ICmpInst::ICMP_SLT) {
1935 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
1936 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
1937 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
1938 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
1939 assert(!C.isMinSignedValue() && "Unexpected icmp slt");
1940 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
1941 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1943 // If this is a signed comparison to 0 and the shift is sign preserving,
1944 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
1945 // do that if we're sure to not continue on in this function.
1946 if (isSignTest(Pred, C))
1947 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
1950 // NUW guarantees that we are only shifting out zero bits from the high bits,
1951 // so we can LSHR the compare constant without needing a mask and eliminate
1953 if (Shl->hasNoUnsignedWrap()) {
1954 if (Pred == ICmpInst::ICMP_UGT) {
1955 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
1956 APInt ShiftedC = C.lshr(*ShiftAmt);
1957 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1959 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1960 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
1961 APInt ShiftedC = C.lshr(*ShiftAmt);
1962 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1964 if (Pred == ICmpInst::ICMP_ULT) {
1965 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
1966 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
1967 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
1968 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
1969 assert(C.ugt(0) && "ult 0 should have been eliminated");
1970 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
1971 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1975 if (Cmp.isEquality() && Shl->hasOneUse()) {
1976 // Strength-reduce the shift into an 'and'.
1977 Constant *Mask = ConstantInt::get(
1979 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
1980 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
1981 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
1982 return new ICmpInst(Pred, And, LShrC);
1985 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1986 bool TrueIfSigned = false;
1987 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
1988 // (X << 31) <s 0 --> (X & 1) != 0
1989 Constant *Mask = ConstantInt::get(
1991 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
1992 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
1993 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1994 And, Constant::getNullValue(ShType));
1997 // Transform (icmp pred iM (shl iM %v, N), C)
1998 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
1999 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2000 // This enables us to get rid of the shift in favor of a trunc that may be
2001 // free on the target. It has the additional benefit of comparing to a
2002 // smaller constant that may be more target-friendly.
2003 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2004 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2005 DL.isLegalInteger(TypeBits - Amt)) {
2006 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2007 if (ShType->isVectorTy())
2008 TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2010 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2011 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2017 /// Fold icmp ({al}shr X, Y), C.
2018 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2019 BinaryOperator *Shr,
2021 // An exact shr only shifts out zero bits, so:
2022 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2023 Value *X = Shr->getOperand(0);
2024 CmpInst::Predicate Pred = Cmp.getPredicate();
2025 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2027 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2029 const APInt *ShiftVal;
2030 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2031 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2033 const APInt *ShiftAmt;
2034 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2037 // Check that the shift amount is in range. If not, don't perform undefined
2038 // shifts. When the shift is visited it will be simplified.
2039 unsigned TypeBits = C.getBitWidth();
2040 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2041 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2044 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2045 bool IsExact = Shr->isExact();
2046 Type *ShrTy = Shr->getType();
2047 // TODO: If we could guarantee that InstSimplify would handle all of the
2048 // constant-value-based preconditions in the folds below, then we could assert
2049 // those conditions rather than checking them. This is difficult because of
2050 // undef/poison (PR34838).
2052 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2053 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2054 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2055 APInt ShiftedC = C.shl(ShAmtVal);
2056 if (ShiftedC.ashr(ShAmtVal) == C)
2057 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2059 if (Pred == CmpInst::ICMP_SGT) {
2060 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2061 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2062 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2063 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2064 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2067 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2068 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2069 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2070 APInt ShiftedC = C.shl(ShAmtVal);
2071 if (ShiftedC.lshr(ShAmtVal) == C)
2072 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2074 if (Pred == CmpInst::ICMP_UGT) {
2075 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2076 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2077 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2078 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2082 if (!Cmp.isEquality())
2085 // Handle equality comparisons of shift-by-constant.
2087 // If the comparison constant changes with the shift, the comparison cannot
2088 // succeed (bits of the comparison constant cannot match the shifted value).
2089 // This should be known by InstSimplify and already be folded to true/false.
2090 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2091 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2092 "Expected icmp+shr simplify did not occur.");
2094 // If the bits shifted out are known zero, compare the unshifted value:
2095 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2097 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2099 if (Shr->hasOneUse()) {
2100 // Canonicalize the shift into an 'and':
2101 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2102 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2103 Constant *Mask = ConstantInt::get(ShrTy, Val);
2104 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2105 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2111 /// Fold icmp (udiv X, Y), C.
2112 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2113 BinaryOperator *UDiv,
2116 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2119 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2121 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2122 Value *Y = UDiv->getOperand(1);
2123 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2124 assert(!C.isMaxValue() &&
2125 "icmp ugt X, UINT_MAX should have been simplified already.");
2126 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2127 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2130 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2131 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2132 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2133 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2134 ConstantInt::get(Y->getType(), C2->udiv(C)));
2140 /// Fold icmp ({su}div X, Y), C.
2141 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2142 BinaryOperator *Div,
2144 // Fold: icmp pred ([us]div X, C2), C -> range test
2145 // Fold this div into the comparison, producing a range check.
2146 // Determine, based on the divide type, what the range is being
2147 // checked. If there is an overflow on the low or high side, remember
2148 // it, otherwise compute the range [low, hi) bounding the new value.
2149 // See: InsertRangeTest above for the kinds of replacements possible.
2151 if (!match(Div->getOperand(1), m_APInt(C2)))
2154 // FIXME: If the operand types don't match the type of the divide
2155 // then don't attempt this transform. The code below doesn't have the
2156 // logic to deal with a signed divide and an unsigned compare (and
2157 // vice versa). This is because (x /s C2) <s C produces different
2158 // results than (x /s C2) <u C or (x /u C2) <s C or even
2159 // (x /u C2) <u C. Simply casting the operands and result won't
2160 // work. :( The if statement below tests that condition and bails
2162 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2163 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2166 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2167 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2168 // division-by-constant cases should be present, we can not assert that they
2169 // have happened before we reach this icmp instruction.
2170 if (C2->isNullValue() || C2->isOneValue() ||
2171 (DivIsSigned && C2->isAllOnesValue()))
2174 // Compute Prod = C * C2. We are essentially solving an equation of
2175 // form X / C2 = C. We solve for X by multiplying C2 and C.
2176 // By solving for X, we can turn this into a range check instead of computing
2178 APInt Prod = C * *C2;
2180 // Determine if the product overflows by seeing if the product is not equal to
2181 // the divide. Make sure we do the same kind of divide as in the LHS
2182 // instruction that we're folding.
2183 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2185 ICmpInst::Predicate Pred = Cmp.getPredicate();
2187 // If the division is known to be exact, then there is no remainder from the
2188 // divide, so the covered range size is unit, otherwise it is the divisor.
2189 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2191 // Figure out the interval that is being checked. For example, a comparison
2192 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2193 // Compute this interval based on the constants involved and the signedness of
2194 // the compare/divide. This computes a half-open interval, keeping track of
2195 // whether either value in the interval overflows. After analysis each
2196 // overflow variable is set to 0 if it's corresponding bound variable is valid
2197 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2198 int LoOverflow = 0, HiOverflow = 0;
2199 APInt LoBound, HiBound;
2201 if (!DivIsSigned) { // udiv
2202 // e.g. X/5 op 3 --> [15, 20)
2204 HiOverflow = LoOverflow = ProdOV;
2206 // If this is not an exact divide, then many values in the range collapse
2207 // to the same result value.
2208 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2210 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2211 if (C.isNullValue()) { // (X / pos) op 0
2212 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2213 LoBound = -(RangeSize - 1);
2214 HiBound = RangeSize;
2215 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2216 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2217 HiOverflow = LoOverflow = ProdOV;
2219 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2220 } else { // (X / pos) op neg
2221 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2223 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2225 APInt DivNeg = -RangeSize;
2226 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2229 } else if (C2->isNegative()) { // Divisor is < 0.
2232 if (C.isNullValue()) { // (X / neg) op 0
2233 // e.g. X/-5 op 0 --> [-4, 5)
2234 LoBound = RangeSize + 1;
2235 HiBound = -RangeSize;
2236 if (HiBound == *C2) { // -INTMIN = INTMIN
2237 HiOverflow = 1; // [INTMIN+1, overflow)
2238 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2240 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2241 // e.g. X/-5 op 3 --> [-19, -14)
2243 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2245 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2246 } else { // (X / neg) op neg
2247 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2248 LoOverflow = HiOverflow = ProdOV;
2250 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2253 // Dividing by a negative swaps the condition. LT <-> GT
2254 Pred = ICmpInst::getSwappedPredicate(Pred);
2257 Value *X = Div->getOperand(0);
2259 default: llvm_unreachable("Unhandled icmp opcode!");
2260 case ICmpInst::ICMP_EQ:
2261 if (LoOverflow && HiOverflow)
2262 return replaceInstUsesWith(Cmp, Builder.getFalse());
2264 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2265 ICmpInst::ICMP_UGE, X,
2266 ConstantInt::get(Div->getType(), LoBound));
2268 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2269 ICmpInst::ICMP_ULT, X,
2270 ConstantInt::get(Div->getType(), HiBound));
2271 return replaceInstUsesWith(
2272 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2273 case ICmpInst::ICMP_NE:
2274 if (LoOverflow && HiOverflow)
2275 return replaceInstUsesWith(Cmp, Builder.getTrue());
2277 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2278 ICmpInst::ICMP_ULT, X,
2279 ConstantInt::get(Div->getType(), LoBound));
2281 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2282 ICmpInst::ICMP_UGE, X,
2283 ConstantInt::get(Div->getType(), HiBound));
2284 return replaceInstUsesWith(Cmp,
2285 insertRangeTest(X, LoBound, HiBound,
2286 DivIsSigned, false));
2287 case ICmpInst::ICMP_ULT:
2288 case ICmpInst::ICMP_SLT:
2289 if (LoOverflow == +1) // Low bound is greater than input range.
2290 return replaceInstUsesWith(Cmp, Builder.getTrue());
2291 if (LoOverflow == -1) // Low bound is less than input range.
2292 return replaceInstUsesWith(Cmp, Builder.getFalse());
2293 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2294 case ICmpInst::ICMP_UGT:
2295 case ICmpInst::ICMP_SGT:
2296 if (HiOverflow == +1) // High bound greater than input range.
2297 return replaceInstUsesWith(Cmp, Builder.getFalse());
2298 if (HiOverflow == -1) // High bound less than input range.
2299 return replaceInstUsesWith(Cmp, Builder.getTrue());
2300 if (Pred == ICmpInst::ICMP_UGT)
2301 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2302 ConstantInt::get(Div->getType(), HiBound));
2303 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2304 ConstantInt::get(Div->getType(), HiBound));
2310 /// Fold icmp (sub X, Y), C.
2311 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2312 BinaryOperator *Sub,
2314 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2315 ICmpInst::Predicate Pred = Cmp.getPredicate();
2317 // The following transforms are only worth it if the only user of the subtract
2319 if (!Sub->hasOneUse())
2322 if (Sub->hasNoSignedWrap()) {
2323 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2324 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2325 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2327 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2328 if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2329 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2331 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2332 if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2333 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2335 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2336 if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2337 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2341 if (!match(X, m_APInt(C2)))
2344 // C2 - Y <u C -> (Y | (C - 1)) == C2
2345 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2346 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2347 (*C2 & (C - 1)) == (C - 1))
2348 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2350 // C2 - Y >u C -> (Y | C) != C2
2351 // iff C2 & C == C and C + 1 is a power of 2
2352 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2353 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2358 /// Fold icmp (add X, Y), C.
2359 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2360 BinaryOperator *Add,
2362 Value *Y = Add->getOperand(1);
2364 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2367 // Fold icmp pred (add X, C2), C.
2368 Value *X = Add->getOperand(0);
2369 Type *Ty = Add->getType();
2370 CmpInst::Predicate Pred = Cmp.getPredicate();
2372 if (!Add->hasOneUse())
2375 // If the add does not wrap, we can always adjust the compare by subtracting
2376 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2377 // are canonicalized to SGT/SLT/UGT/ULT.
2378 if ((Add->hasNoSignedWrap() &&
2379 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2380 (Add->hasNoUnsignedWrap() &&
2381 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2384 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2385 // If there is overflow, the result must be true or false.
2386 // TODO: Can we assert there is no overflow because InstSimplify always
2387 // handles those cases?
2389 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2390 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2393 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2394 const APInt &Upper = CR.getUpper();
2395 const APInt &Lower = CR.getLower();
2396 if (Cmp.isSigned()) {
2397 if (Lower.isSignMask())
2398 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2399 if (Upper.isSignMask())
2400 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2402 if (Lower.isMinValue())
2403 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2404 if (Upper.isMinValue())
2405 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2408 // X+C <u C2 -> (X & -C2) == C
2409 // iff C & (C2-1) == 0
2410 // C2 is a power of 2
2411 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2412 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2413 ConstantExpr::getNeg(cast<Constant>(Y)));
2415 // X+C >u C2 -> (X & ~C2) != C
2417 // C2+1 is a power of 2
2418 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2419 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2420 ConstantExpr::getNeg(cast<Constant>(Y)));
2425 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2426 Value *&RHS, ConstantInt *&Less,
2427 ConstantInt *&Equal,
2428 ConstantInt *&Greater) {
2429 // TODO: Generalize this to work with other comparison idioms or ensure
2430 // they get canonicalized into this form.
2432 // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32
2433 // Greater), where Equal, Less and Greater are placeholders for any three
2435 ICmpInst::Predicate PredA, PredB;
2436 if (match(SI->getTrueValue(), m_ConstantInt(Equal)) &&
2437 match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) &&
2438 PredA == ICmpInst::ICMP_EQ &&
2439 match(SI->getFalseValue(),
2440 m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)),
2441 m_ConstantInt(Less), m_ConstantInt(Greater))) &&
2442 PredB == ICmpInst::ICMP_SLT) {
2448 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2452 assert(C && "Cmp RHS should be a constant int!");
2453 // If we're testing a constant value against the result of a three way
2454 // comparison, the result can be expressed directly in terms of the
2455 // original values being compared. Note: We could possibly be more
2456 // aggressive here and remove the hasOneUse test. The original select is
2457 // really likely to simplify or sink when we remove a test of the result.
2458 Value *OrigLHS, *OrigRHS;
2459 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2460 if (Cmp.hasOneUse() &&
2461 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2463 assert(C1LessThan && C2Equal && C3GreaterThan);
2465 bool TrueWhenLessThan =
2466 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2468 bool TrueWhenEqual =
2469 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2471 bool TrueWhenGreaterThan =
2472 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2475 // This generates the new instruction that will replace the original Cmp
2476 // Instruction. Instead of enumerating the various combinations when
2477 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2478 // false, we rely on chaining of ORs and future passes of InstCombine to
2479 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2481 // When none of the three constants satisfy the predicate for the RHS (C),
2482 // the entire original Cmp can be simplified to a false.
2483 Value *Cond = Builder.getFalse();
2484 if (TrueWhenLessThan)
2485 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT, OrigLHS, OrigRHS));
2487 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ, OrigLHS, OrigRHS));
2488 if (TrueWhenGreaterThan)
2489 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT, OrigLHS, OrigRHS));
2491 return replaceInstUsesWith(Cmp, Cond);
2496 Instruction *InstCombiner::foldICmpBitCastConstant(ICmpInst &Cmp,
2497 BitCastInst *Bitcast,
2499 // Folding: icmp <pred> iN X, C
2500 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2501 // and C is a splat of a K-bit pattern
2502 // and SC is a constant vector = <C', C', C', ..., C'>
2504 // %E = extractelement <M x iK> %vec, i32 C'
2505 // icmp <pred> iK %E, trunc(C)
2506 if (!Bitcast->getType()->isIntegerTy() ||
2507 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2510 Value *BCIOp = Bitcast->getOperand(0);
2511 Value *Vec = nullptr; // 1st vector arg of the shufflevector
2512 Constant *Mask = nullptr; // Mask arg of the shufflevector
2514 m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2515 // Check whether every element of Mask is the same constant
2516 if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2517 auto *VecTy = cast<VectorType>(BCIOp->getType());
2518 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2519 auto Pred = Cmp.getPredicate();
2520 if (C.isSplat(EltTy->getBitWidth())) {
2521 // Fold the icmp based on the value of C
2522 // If C is M copies of an iK sized bit pattern,
2524 // => %E = extractelement <N x iK> %vec, i32 Elem
2525 // icmp <pred> iK %SplatVal, <pattern>
2526 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2527 Value *NewC = ConstantInt::get(EltTy, C.trunc(EltTy->getBitWidth()));
2528 return new ICmpInst(Pred, Extract, NewC);
2535 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2536 /// where X is some kind of instruction.
2537 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2539 if (!match(Cmp.getOperand(1), m_APInt(C)))
2542 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2543 switch (BO->getOpcode()) {
2544 case Instruction::Xor:
2545 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2548 case Instruction::And:
2549 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2552 case Instruction::Or:
2553 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2556 case Instruction::Mul:
2557 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2560 case Instruction::Shl:
2561 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2564 case Instruction::LShr:
2565 case Instruction::AShr:
2566 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2569 case Instruction::UDiv:
2570 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2573 case Instruction::SDiv:
2574 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2577 case Instruction::Sub:
2578 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2581 case Instruction::Add:
2582 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2588 // TODO: These folds could be refactored to be part of the above calls.
2589 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2593 // Match against CmpInst LHS being instructions other than binary operators.
2595 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2596 // For now, we only support constant integers while folding the
2597 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2598 // similar to the cases handled by binary ops above.
2599 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2600 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2604 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2605 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2609 if (auto *BCI = dyn_cast<BitCastInst>(Cmp.getOperand(0))) {
2610 if (Instruction *I = foldICmpBitCastConstant(Cmp, BCI, *C))
2614 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, *C))
2620 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2621 /// icmp eq/ne BO, C.
2622 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2625 // TODO: Some of these folds could work with arbitrary constants, but this
2626 // function is limited to scalar and vector splat constants.
2627 if (!Cmp.isEquality())
2630 ICmpInst::Predicate Pred = Cmp.getPredicate();
2631 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2632 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2633 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2635 switch (BO->getOpcode()) {
2636 case Instruction::SRem:
2637 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2638 if (C.isNullValue() && BO->hasOneUse()) {
2640 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2641 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2642 return new ICmpInst(Pred, NewRem,
2643 Constant::getNullValue(BO->getType()));
2647 case Instruction::Add: {
2648 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2650 if (match(BOp1, m_APInt(BOC))) {
2651 if (BO->hasOneUse()) {
2652 Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2653 return new ICmpInst(Pred, BOp0, SubC);
2655 } else if (C.isNullValue()) {
2656 // Replace ((add A, B) != 0) with (A != -B) if A or B is
2657 // efficiently invertible, or if the add has just this one use.
2658 if (Value *NegVal = dyn_castNegVal(BOp1))
2659 return new ICmpInst(Pred, BOp0, NegVal);
2660 if (Value *NegVal = dyn_castNegVal(BOp0))
2661 return new ICmpInst(Pred, NegVal, BOp1);
2662 if (BO->hasOneUse()) {
2663 Value *Neg = Builder.CreateNeg(BOp1);
2665 return new ICmpInst(Pred, BOp0, Neg);
2670 case Instruction::Xor:
2671 if (BO->hasOneUse()) {
2672 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2673 // For the xor case, we can xor two constants together, eliminating
2674 // the explicit xor.
2675 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2676 } else if (C.isNullValue()) {
2677 // Replace ((xor A, B) != 0) with (A != B)
2678 return new ICmpInst(Pred, BOp0, BOp1);
2682 case Instruction::Sub:
2683 if (BO->hasOneUse()) {
2685 if (match(BOp0, m_APInt(BOC))) {
2686 // Replace ((sub BOC, B) != C) with (B != BOC-C).
2687 Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2688 return new ICmpInst(Pred, BOp1, SubC);
2689 } else if (C.isNullValue()) {
2690 // Replace ((sub A, B) != 0) with (A != B).
2691 return new ICmpInst(Pred, BOp0, BOp1);
2695 case Instruction::Or: {
2697 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2698 // Comparing if all bits outside of a constant mask are set?
2699 // Replace (X | C) == -1 with (X & ~C) == ~C.
2700 // This removes the -1 constant.
2701 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2702 Value *And = Builder.CreateAnd(BOp0, NotBOC);
2703 return new ICmpInst(Pred, And, NotBOC);
2707 case Instruction::And: {
2709 if (match(BOp1, m_APInt(BOC))) {
2710 // If we have ((X & C) == C), turn it into ((X & C) != 0).
2711 if (C == *BOC && C.isPowerOf2())
2712 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2713 BO, Constant::getNullValue(RHS->getType()));
2715 // Don't perform the following transforms if the AND has multiple uses
2716 if (!BO->hasOneUse())
2719 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
2720 if (BOC->isSignMask()) {
2721 Constant *Zero = Constant::getNullValue(BOp0->getType());
2722 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
2723 return new ICmpInst(NewPred, BOp0, Zero);
2726 // ((X & ~7) == 0) --> X < 8
2727 if (C.isNullValue() && (~(*BOC) + 1).isPowerOf2()) {
2728 Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1));
2729 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
2730 return new ICmpInst(NewPred, BOp0, NegBOC);
2735 case Instruction::Mul:
2736 if (C.isNullValue() && BO->hasNoSignedWrap()) {
2738 if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2739 // The trivial case (mul X, 0) is handled by InstSimplify.
2740 // General case : (mul X, C) != 0 iff X != 0
2741 // (mul X, C) == 0 iff X == 0
2742 return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2746 case Instruction::UDiv:
2747 if (C.isNullValue()) {
2748 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2749 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2750 return new ICmpInst(NewPred, BOp1, BOp0);
2759 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
2760 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2762 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0));
2763 if (!II || !Cmp.isEquality())
2766 // Handle icmp {eq|ne} <intrinsic>, Constant.
2767 Type *Ty = II->getType();
2768 unsigned BitWidth = C.getBitWidth();
2769 switch (II->getIntrinsicID()) {
2770 case Intrinsic::bswap:
2772 Cmp.setOperand(0, II->getArgOperand(0));
2773 Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
2776 case Intrinsic::ctlz:
2777 case Intrinsic::cttz: {
2778 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
2779 if (C == BitWidth) {
2781 Cmp.setOperand(0, II->getArgOperand(0));
2782 Cmp.setOperand(1, ConstantInt::getNullValue(Ty));
2786 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
2787 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
2788 // Limit to one use to ensure we don't increase instruction count.
2789 unsigned Num = C.getLimitedValue(BitWidth);
2790 if (Num != BitWidth && II->hasOneUse()) {
2791 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
2792 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
2793 : APInt::getHighBitsSet(BitWidth, Num + 1);
2794 APInt Mask2 = IsTrailing
2795 ? APInt::getOneBitSet(BitWidth, Num)
2796 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
2797 Cmp.setOperand(0, Builder.CreateAnd(II->getArgOperand(0), Mask1));
2798 Cmp.setOperand(1, ConstantInt::get(Ty, Mask2));
2805 case Intrinsic::ctpop: {
2806 // popcount(A) == 0 -> A == 0 and likewise for !=
2807 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
2808 bool IsZero = C.isNullValue();
2809 if (IsZero || C == BitWidth) {
2811 Cmp.setOperand(0, II->getArgOperand(0));
2813 IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty);
2814 Cmp.setOperand(1, NewOp);
2826 /// Handle icmp with constant (but not simple integer constant) RHS.
2827 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
2828 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2829 Constant *RHSC = dyn_cast<Constant>(Op1);
2830 Instruction *LHSI = dyn_cast<Instruction>(Op0);
2834 switch (LHSI->getOpcode()) {
2835 case Instruction::GetElementPtr:
2836 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2837 if (RHSC->isNullValue() &&
2838 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2839 return new ICmpInst(
2840 I.getPredicate(), LHSI->getOperand(0),
2841 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2843 case Instruction::PHI:
2844 // Only fold icmp into the PHI if the phi and icmp are in the same
2845 // block. If in the same block, we're encouraging jump threading. If
2846 // not, we are just pessimizing the code by making an i1 phi.
2847 if (LHSI->getParent() == I.getParent())
2848 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
2851 case Instruction::Select: {
2852 // If either operand of the select is a constant, we can fold the
2853 // comparison into the select arms, which will cause one to be
2854 // constant folded and the select turned into a bitwise or.
2855 Value *Op1 = nullptr, *Op2 = nullptr;
2856 ConstantInt *CI = nullptr;
2857 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2858 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2859 CI = dyn_cast<ConstantInt>(Op1);
2861 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2862 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2863 CI = dyn_cast<ConstantInt>(Op2);
2866 // We only want to perform this transformation if it will not lead to
2867 // additional code. This is true if either both sides of the select
2868 // fold to a constant (in which case the icmp is replaced with a select
2869 // which will usually simplify) or this is the only user of the
2870 // select (in which case we are trading a select+icmp for a simpler
2871 // select+icmp) or all uses of the select can be replaced based on
2872 // dominance information ("Global cases").
2873 bool Transform = false;
2876 else if (Op1 || Op2) {
2878 if (LHSI->hasOneUse())
2881 else if (CI && !CI->isZero())
2882 // When Op1 is constant try replacing select with second operand.
2883 // Otherwise Op2 is constant and try replacing select with first
2886 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
2890 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
2893 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
2895 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2899 case Instruction::IntToPtr:
2900 // icmp pred inttoptr(X), null -> icmp pred X, 0
2901 if (RHSC->isNullValue() &&
2902 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
2903 return new ICmpInst(
2904 I.getPredicate(), LHSI->getOperand(0),
2905 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2908 case Instruction::Load:
2909 // Try to optimize things like "A[i] > 4" to index computations.
2910 if (GetElementPtrInst *GEP =
2911 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2912 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2913 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2914 !cast<LoadInst>(LHSI)->isVolatile())
2915 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
2924 /// Some comparisons can be simplified.
2925 /// In this case, we are looking for comparisons that look like
2926 /// a check for a lossy truncation.
2928 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
2929 /// Where Mask is some pattern that produces all-ones in low bits:
2931 /// ((-1 << y) >> y) <- non-canonical, has extra uses
2933 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
2934 /// The Mask can be a constant, too.
2935 /// For some predicates, the operands are commutative.
2936 /// For others, x can only be on a specific side.
2937 static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
2938 InstCombiner::BuilderTy &Builder) {
2939 ICmpInst::Predicate SrcPred;
2941 auto m_VariableMask = m_CombineOr(
2942 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
2943 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
2944 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
2945 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
2946 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
2947 if (!match(&I, m_c_ICmp(SrcPred,
2948 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
2952 ICmpInst::Predicate DstPred;
2954 case ICmpInst::Predicate::ICMP_EQ:
2955 // x & (-1 >> y) == x -> x u<= (-1 >> y)
2956 DstPred = ICmpInst::Predicate::ICMP_ULE;
2958 case ICmpInst::Predicate::ICMP_NE:
2959 // x & (-1 >> y) != x -> x u> (-1 >> y)
2960 DstPred = ICmpInst::Predicate::ICMP_UGT;
2962 case ICmpInst::Predicate::ICMP_UGT:
2963 // x u> x & (-1 >> y) -> x u> (-1 >> y)
2964 assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
2965 DstPred = ICmpInst::Predicate::ICMP_UGT;
2967 case ICmpInst::Predicate::ICMP_UGE:
2968 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
2969 assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
2970 DstPred = ICmpInst::Predicate::ICMP_ULE;
2972 case ICmpInst::Predicate::ICMP_ULT:
2973 // x & (-1 >> y) u< x -> x u> (-1 >> y)
2974 assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
2975 DstPred = ICmpInst::Predicate::ICMP_UGT;
2977 case ICmpInst::Predicate::ICMP_ULE:
2978 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
2979 assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
2980 DstPred = ICmpInst::Predicate::ICMP_ULE;
2982 case ICmpInst::Predicate::ICMP_SGT:
2983 // x s> x & (-1 >> y) -> x s> (-1 >> y)
2984 if (X != I.getOperand(0)) // X must be on LHS of comparison!
2985 return nullptr; // Ignore the other case.
2986 DstPred = ICmpInst::Predicate::ICMP_SGT;
2988 case ICmpInst::Predicate::ICMP_SGE:
2989 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
2990 if (X != I.getOperand(1)) // X must be on RHS of comparison!
2991 return nullptr; // Ignore the other case.
2992 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
2994 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
2996 DstPred = ICmpInst::Predicate::ICMP_SLE;
2998 case ICmpInst::Predicate::ICMP_SLT:
2999 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3000 if (X != I.getOperand(1)) // X must be on RHS of comparison!
3001 return nullptr; // Ignore the other case.
3002 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3004 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3006 DstPred = ICmpInst::Predicate::ICMP_SGT;
3008 case ICmpInst::Predicate::ICMP_SLE:
3009 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3010 if (X != I.getOperand(0)) // X must be on LHS of comparison!
3011 return nullptr; // Ignore the other case.
3012 DstPred = ICmpInst::Predicate::ICMP_SLE;
3015 llvm_unreachable("All possible folds are handled.");
3018 return Builder.CreateICmp(DstPred, X, M);
3021 /// Some comparisons can be simplified.
3022 /// In this case, we are looking for comparisons that look like
3023 /// a check for a lossy signed truncation.
3024 /// Folds: (MaskedBits is a constant.)
3025 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3027 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3028 /// Where KeptBits = bitwidth(%x) - MaskedBits
3030 foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3031 InstCombiner::BuilderTy &Builder) {
3032 ICmpInst::Predicate SrcPred;
3034 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3035 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3036 if (!match(&I, m_c_ICmp(SrcPred,
3037 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3042 // Potential handling of non-splats: for each element:
3043 // * if both are undef, replace with constant 0.
3044 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3045 // * if both are not undef, and are different, bailout.
3046 // * else, only one is undef, then pick the non-undef one.
3048 // The shift amount must be equal.
3051 const APInt &MaskedBits = *C0;
3052 assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3054 ICmpInst::Predicate DstPred;
3056 case ICmpInst::Predicate::ICMP_EQ:
3057 // ((%x << MaskedBits) a>> MaskedBits) == %x
3059 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3060 DstPred = ICmpInst::Predicate::ICMP_ULT;
3062 case ICmpInst::Predicate::ICMP_NE:
3063 // ((%x << MaskedBits) a>> MaskedBits) != %x
3065 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3066 DstPred = ICmpInst::Predicate::ICMP_UGE;
3068 // FIXME: are more folds possible?
3073 auto *XType = X->getType();
3074 const unsigned XBitWidth = XType->getScalarSizeInBits();
3075 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3076 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3078 // KeptBits = bitwidth(%x) - MaskedBits
3079 const APInt KeptBits = BitWidth - MaskedBits;
3080 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3081 // ICmpCst = (1 << KeptBits)
3082 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3083 assert(ICmpCst.isPowerOf2());
3084 // AddCst = (1 << (KeptBits-1))
3085 const APInt AddCst = ICmpCst.lshr(1);
3086 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3088 // T0 = add %x, AddCst
3089 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3090 // T1 = T0 DstPred ICmpCst
3091 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3096 /// Try to fold icmp (binop), X or icmp X, (binop).
3097 /// TODO: A large part of this logic is duplicated in InstSimplify's
3098 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3100 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
3101 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3103 // Special logic for binary operators.
3104 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3105 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3109 const CmpInst::Predicate Pred = I.getPredicate();
3112 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3113 // (Op1 + X) <u Op1 --> ~Op1 <u X
3114 // Op0 >u (Op0 + X) --> X >u ~Op0
3115 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3116 Pred == ICmpInst::ICMP_ULT)
3117 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3118 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3119 Pred == ICmpInst::ICMP_UGT)
3120 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3122 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3123 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3125 ICmpInst::isEquality(Pred) ||
3126 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3127 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3128 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3130 ICmpInst::isEquality(Pred) ||
3131 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3132 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3134 // Analyze the case when either Op0 or Op1 is an add instruction.
3135 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3136 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3137 if (BO0 && BO0->getOpcode() == Instruction::Add) {
3138 A = BO0->getOperand(0);
3139 B = BO0->getOperand(1);
3141 if (BO1 && BO1->getOpcode() == Instruction::Add) {
3142 C = BO1->getOperand(0);
3143 D = BO1->getOperand(1);
3146 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3147 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3148 return new ICmpInst(Pred, A == Op1 ? B : A,
3149 Constant::getNullValue(Op1->getType()));
3151 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3152 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3153 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3156 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3157 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3159 // Try not to increase register pressure.
3160 BO0->hasOneUse() && BO1->hasOneUse()) {
3161 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3164 // C + B == C + D -> B == D
3167 } else if (A == D) {
3168 // D + B == C + D -> B == C
3171 } else if (B == C) {
3172 // A + C == C + D -> A == D
3177 // A + D == C + D -> A == C
3181 return new ICmpInst(Pred, Y, Z);
3184 // icmp slt (X + -1), Y -> icmp sle X, Y
3185 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3186 match(B, m_AllOnes()))
3187 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3189 // icmp sge (X + -1), Y -> icmp sgt X, Y
3190 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3191 match(B, m_AllOnes()))
3192 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3194 // icmp sle (X + 1), Y -> icmp slt X, Y
3195 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3196 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3198 // icmp sgt (X + 1), Y -> icmp sge X, Y
3199 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3200 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3202 // icmp sgt X, (Y + -1) -> icmp sge X, Y
3203 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3204 match(D, m_AllOnes()))
3205 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3207 // icmp sle X, (Y + -1) -> icmp slt X, Y
3208 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3209 match(D, m_AllOnes()))
3210 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3212 // icmp sge X, (Y + 1) -> icmp sgt X, Y
3213 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3214 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3216 // icmp slt X, (Y + 1) -> icmp sle X, Y
3217 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3218 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3220 // TODO: The subtraction-related identities shown below also hold, but
3221 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3222 // wouldn't happen even if they were implemented.
3224 // icmp ult (X - 1), Y -> icmp ule X, Y
3225 // icmp uge (X - 1), Y -> icmp ugt X, Y
3226 // icmp ugt X, (Y - 1) -> icmp uge X, Y
3227 // icmp ule X, (Y - 1) -> icmp ult X, Y
3229 // icmp ule (X + 1), Y -> icmp ult X, Y
3230 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3231 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3233 // icmp ugt (X + 1), Y -> icmp uge X, Y
3234 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3235 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3237 // icmp uge X, (Y + 1) -> icmp ugt X, Y
3238 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3239 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3241 // icmp ult X, (Y + 1) -> icmp ule X, Y
3242 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3243 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3245 // if C1 has greater magnitude than C2:
3246 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3247 // s.t. C3 = C1 - C2
3249 // if C2 has greater magnitude than C1:
3250 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3251 // s.t. C3 = C2 - C1
3252 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3253 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3254 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3255 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3256 const APInt &AP1 = C1->getValue();
3257 const APInt &AP2 = C2->getValue();
3258 if (AP1.isNegative() == AP2.isNegative()) {
3259 APInt AP1Abs = C1->getValue().abs();
3260 APInt AP2Abs = C2->getValue().abs();
3261 if (AP1Abs.uge(AP2Abs)) {
3262 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3263 Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3264 return new ICmpInst(Pred, NewAdd, C);
3266 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3267 Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3268 return new ICmpInst(Pred, A, NewAdd);
3273 // Analyze the case when either Op0 or Op1 is a sub instruction.
3274 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3279 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3280 A = BO0->getOperand(0);
3281 B = BO0->getOperand(1);
3283 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3284 C = BO1->getOperand(0);
3285 D = BO1->getOperand(1);
3288 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3289 if (A == Op1 && NoOp0WrapProblem)
3290 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3291 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3292 if (C == Op0 && NoOp1WrapProblem)
3293 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3295 // (A - B) >u A --> A <u B
3296 if (A == Op1 && Pred == ICmpInst::ICMP_UGT)
3297 return new ICmpInst(ICmpInst::ICMP_ULT, A, B);
3298 // C <u (C - D) --> C <u D
3299 if (C == Op0 && Pred == ICmpInst::ICMP_ULT)
3300 return new ICmpInst(ICmpInst::ICMP_ULT, C, D);
3302 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3303 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3304 // Try not to increase register pressure.
3305 BO0->hasOneUse() && BO1->hasOneUse())
3306 return new ICmpInst(Pred, A, C);
3307 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3308 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3309 // Try not to increase register pressure.
3310 BO0->hasOneUse() && BO1->hasOneUse())
3311 return new ICmpInst(Pred, D, B);
3313 // icmp (0-X) < cst --> x > -cst
3314 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3316 if (match(BO0, m_Neg(m_Value(X))))
3317 if (Constant *RHSC = dyn_cast<Constant>(Op1))
3318 if (RHSC->isNotMinSignedValue())
3319 return new ICmpInst(I.getSwappedPredicate(), X,
3320 ConstantExpr::getNeg(RHSC));
3323 BinaryOperator *SRem = nullptr;
3324 // icmp (srem X, Y), Y
3325 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3327 // icmp Y, (srem X, Y)
3328 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3329 Op0 == BO1->getOperand(1))
3332 // We don't check hasOneUse to avoid increasing register pressure because
3333 // the value we use is the same value this instruction was already using.
3334 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3337 case ICmpInst::ICMP_EQ:
3338 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3339 case ICmpInst::ICMP_NE:
3340 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3341 case ICmpInst::ICMP_SGT:
3342 case ICmpInst::ICMP_SGE:
3343 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3344 Constant::getAllOnesValue(SRem->getType()));
3345 case ICmpInst::ICMP_SLT:
3346 case ICmpInst::ICMP_SLE:
3347 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3348 Constant::getNullValue(SRem->getType()));
3352 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3353 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3354 switch (BO0->getOpcode()) {
3357 case Instruction::Add:
3358 case Instruction::Sub:
3359 case Instruction::Xor: {
3360 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3361 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3364 if (match(BO0->getOperand(1), m_APInt(C))) {
3365 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3366 if (C->isSignMask()) {
3367 ICmpInst::Predicate NewPred =
3368 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3369 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3372 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3373 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3374 ICmpInst::Predicate NewPred =
3375 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3376 NewPred = I.getSwappedPredicate(NewPred);
3377 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3382 case Instruction::Mul: {
3383 if (!I.isEquality())
3387 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3389 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3390 // Mask = -1 >> count-trailing-zeros(C).
3391 if (unsigned TZs = C->countTrailingZeros()) {
3392 Constant *Mask = ConstantInt::get(
3394 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3395 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3396 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3397 return new ICmpInst(Pred, And1, And2);
3399 // If there are no trailing zeros in the multiplier, just eliminate
3400 // the multiplies (no masking is needed):
3401 // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3402 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3406 case Instruction::UDiv:
3407 case Instruction::LShr:
3408 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3410 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3412 case Instruction::SDiv:
3413 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3415 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3417 case Instruction::AShr:
3418 if (!BO0->isExact() || !BO1->isExact())
3420 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3422 case Instruction::Shl: {
3423 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3424 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3427 if (!NSW && I.isSigned())
3429 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3435 // Transform A & (L - 1) `ult` L --> L != 0
3436 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3437 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3439 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3440 auto *Zero = Constant::getNullValue(BO0->getType());
3441 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3445 if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
3446 return replaceInstUsesWith(I, V);
3448 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
3449 return replaceInstUsesWith(I, V);
3454 /// Fold icmp Pred min|max(X, Y), X.
3455 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
3456 ICmpInst::Predicate Pred = Cmp.getPredicate();
3457 Value *Op0 = Cmp.getOperand(0);
3458 Value *X = Cmp.getOperand(1);
3460 // Canonicalize minimum or maximum operand to LHS of the icmp.
3461 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3462 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3463 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3464 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3466 Pred = Cmp.getSwappedPredicate();
3470 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3471 // smin(X, Y) == X --> X s<= Y
3472 // smin(X, Y) s>= X --> X s<= Y
3473 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3474 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3476 // smin(X, Y) != X --> X s> Y
3477 // smin(X, Y) s< X --> X s> Y
3478 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3479 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3481 // These cases should be handled in InstSimplify:
3482 // smin(X, Y) s<= X --> true
3483 // smin(X, Y) s> X --> false
3487 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3488 // smax(X, Y) == X --> X s>= Y
3489 // smax(X, Y) s<= X --> X s>= Y
3490 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3491 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3493 // smax(X, Y) != X --> X s< Y
3494 // smax(X, Y) s> X --> X s< Y
3495 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3496 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3498 // These cases should be handled in InstSimplify:
3499 // smax(X, Y) s>= X --> true
3500 // smax(X, Y) s< X --> false
3504 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3505 // umin(X, Y) == X --> X u<= Y
3506 // umin(X, Y) u>= X --> X u<= Y
3507 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3508 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3510 // umin(X, Y) != X --> X u> Y
3511 // umin(X, Y) u< X --> X u> Y
3512 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3513 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3515 // These cases should be handled in InstSimplify:
3516 // umin(X, Y) u<= X --> true
3517 // umin(X, Y) u> X --> false
3521 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3522 // umax(X, Y) == X --> X u>= Y
3523 // umax(X, Y) u<= X --> X u>= Y
3524 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3525 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3527 // umax(X, Y) != X --> X u< Y
3528 // umax(X, Y) u> X --> X u< Y
3529 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3530 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3532 // These cases should be handled in InstSimplify:
3533 // umax(X, Y) u>= X --> true
3534 // umax(X, Y) u< X --> false
3541 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3542 if (!I.isEquality())
3545 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3546 const CmpInst::Predicate Pred = I.getPredicate();
3547 Value *A, *B, *C, *D;
3548 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3549 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3550 Value *OtherVal = A == Op1 ? B : A;
3551 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3554 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3555 // A^c1 == C^c2 --> A == C^(c1^c2)
3556 ConstantInt *C1, *C2;
3557 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3559 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
3560 Value *Xor = Builder.CreateXor(C, NC);
3561 return new ICmpInst(Pred, A, Xor);
3564 // A^B == A^D -> B == D
3566 return new ICmpInst(Pred, B, D);
3568 return new ICmpInst(Pred, B, C);
3570 return new ICmpInst(Pred, A, D);
3572 return new ICmpInst(Pred, A, C);
3576 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3577 // A == (A^B) -> B == 0
3578 Value *OtherVal = A == Op0 ? B : A;
3579 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3582 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3583 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3584 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3585 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3591 } else if (A == D) {
3595 } else if (B == C) {
3599 } else if (B == D) {
3605 if (X) { // Build (X^Y) & Z
3606 Op1 = Builder.CreateXor(X, Y);
3607 Op1 = Builder.CreateAnd(Op1, Z);
3608 I.setOperand(0, Op1);
3609 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3614 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3615 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3617 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3618 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3619 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3620 match(Op1, m_ZExt(m_Value(A))))) {
3621 APInt Pow2 = Cst1->getValue() + 1;
3622 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3623 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3624 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
3627 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3628 // For lshr and ashr pairs.
3629 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3630 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3631 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3632 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3633 unsigned TypeBits = Cst1->getBitWidth();
3634 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3635 if (ShAmt < TypeBits && ShAmt != 0) {
3636 ICmpInst::Predicate NewPred =
3637 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
3638 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3639 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3640 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
3644 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3645 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3646 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3647 unsigned TypeBits = Cst1->getBitWidth();
3648 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3649 if (ShAmt < TypeBits && ShAmt != 0) {
3650 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3651 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3652 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
3653 I.getName() + ".mask");
3654 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
3658 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3659 // "icmp (and X, mask), cst"
3661 if (Op0->hasOneUse() &&
3662 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3663 match(Op1, m_ConstantInt(Cst1)) &&
3664 // Only do this when A has multiple uses. This is most important to do
3665 // when it exposes other optimizations.
3667 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3669 if (ShAmt < ASize) {
3671 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3674 APInt CmpV = Cst1->getValue().zext(ASize);
3677 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
3678 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
3682 // If both operands are byte-swapped or bit-reversed, just compare the
3684 // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
3685 // and handle more intrinsics.
3686 if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
3687 (match(Op0, m_BitReverse(m_Value(A))) &&
3688 match(Op1, m_BitReverse(m_Value(B)))))
3689 return new ICmpInst(Pred, A, B);
3694 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3696 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3697 const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3698 Value *LHSCIOp = LHSCI->getOperand(0);
3699 Type *SrcTy = LHSCIOp->getType();
3700 Type *DestTy = LHSCI->getType();
3703 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3704 // integer type is the same size as the pointer type.
3705 const auto& CompatibleSizes = [&](Type* SrcTy, Type* DestTy) -> bool {
3706 if (isa<VectorType>(SrcTy)) {
3707 SrcTy = cast<VectorType>(SrcTy)->getElementType();
3708 DestTy = cast<VectorType>(DestTy)->getElementType();
3710 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
3712 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3713 CompatibleSizes(SrcTy, DestTy)) {
3714 Value *RHSOp = nullptr;
3715 if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3716 Value *RHSCIOp = RHSC->getOperand(0);
3717 if (RHSCIOp->getType()->getPointerAddressSpace() ==
3718 LHSCIOp->getType()->getPointerAddressSpace()) {
3719 RHSOp = RHSC->getOperand(0);
3720 // If the pointer types don't match, insert a bitcast.
3721 if (LHSCIOp->getType() != RHSOp->getType())
3722 RHSOp = Builder.CreateBitCast(RHSOp, LHSCIOp->getType());
3724 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
3725 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
3729 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
3732 // The code below only handles extension cast instructions, so far.
3734 if (LHSCI->getOpcode() != Instruction::ZExt &&
3735 LHSCI->getOpcode() != Instruction::SExt)
3738 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
3739 bool isSignedCmp = ICmp.isSigned();
3741 if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
3742 // Not an extension from the same type?
3743 RHSCIOp = CI->getOperand(0);
3744 if (RHSCIOp->getType() != LHSCIOp->getType())
3747 // If the signedness of the two casts doesn't agree (i.e. one is a sext
3748 // and the other is a zext), then we can't handle this.
3749 if (CI->getOpcode() != LHSCI->getOpcode())
3752 // Deal with equality cases early.
3753 if (ICmp.isEquality())
3754 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3756 // A signed comparison of sign extended values simplifies into a
3757 // signed comparison.
3758 if (isSignedCmp && isSignedExt)
3759 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3761 // The other three cases all fold into an unsigned comparison.
3762 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
3765 // If we aren't dealing with a constant on the RHS, exit early.
3766 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
3770 // Compute the constant that would happen if we truncated to SrcTy then
3771 // re-extended to DestTy.
3772 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
3773 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
3775 // If the re-extended constant didn't change...
3777 // Deal with equality cases early.
3778 if (ICmp.isEquality())
3779 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3781 // A signed comparison of sign extended values simplifies into a
3782 // signed comparison.
3783 if (isSignedExt && isSignedCmp)
3784 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3786 // The other three cases all fold into an unsigned comparison.
3787 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
3790 // The re-extended constant changed, partly changed (in the case of a vector),
3791 // or could not be determined to be equal (in the case of a constant
3792 // expression), so the constant cannot be represented in the shorter type.
3793 // Consequently, we cannot emit a simple comparison.
3794 // All the cases that fold to true or false will have already been handled
3795 // by SimplifyICmpInst, so only deal with the tricky case.
3797 if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
3800 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
3801 // should have been folded away previously and not enter in here.
3803 // We're performing an unsigned comp with a sign extended value.
3804 // This is true if the input is >= 0. [aka >s -1]
3805 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
3806 Value *Result = Builder.CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
3808 // Finally, return the value computed.
3809 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
3810 return replaceInstUsesWith(ICmp, Result);
3812 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
3813 return BinaryOperator::CreateNot(Result);
3816 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
3817 Value *RHS, Instruction &OrigI,
3818 Value *&Result, Constant *&Overflow) {
3819 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
3820 std::swap(LHS, RHS);
3822 auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
3824 Overflow = OverflowVal;
3826 Result->takeName(&OrigI);
3830 // If the overflow check was an add followed by a compare, the insertion point
3831 // may be pointing to the compare. We want to insert the new instructions
3832 // before the add in case there are uses of the add between the add and the
3834 Builder.SetInsertPoint(&OrigI);
3838 llvm_unreachable("bad overflow check kind!");
3840 case OCF_UNSIGNED_ADD: {
3841 OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
3842 if (OR == OverflowResult::NeverOverflows)
3843 return SetResult(Builder.CreateNUWAdd(LHS, RHS), Builder.getFalse(),
3846 if (OR == OverflowResult::AlwaysOverflows)
3847 return SetResult(Builder.CreateAdd(LHS, RHS), Builder.getTrue(), true);
3849 // Fall through uadd into sadd
3852 case OCF_SIGNED_ADD: {
3853 // X + 0 -> {X, false}
3854 if (match(RHS, m_Zero()))
3855 return SetResult(LHS, Builder.getFalse(), false);
3857 // We can strength reduce this signed add into a regular add if we can prove
3858 // that it will never overflow.
3859 if (OCF == OCF_SIGNED_ADD)
3860 if (willNotOverflowSignedAdd(LHS, RHS, OrigI))
3861 return SetResult(Builder.CreateNSWAdd(LHS, RHS), Builder.getFalse(),
3866 case OCF_UNSIGNED_SUB:
3867 case OCF_SIGNED_SUB: {
3868 // X - 0 -> {X, false}
3869 if (match(RHS, m_Zero()))
3870 return SetResult(LHS, Builder.getFalse(), false);
3872 if (OCF == OCF_SIGNED_SUB) {
3873 if (willNotOverflowSignedSub(LHS, RHS, OrigI))
3874 return SetResult(Builder.CreateNSWSub(LHS, RHS), Builder.getFalse(),
3877 if (willNotOverflowUnsignedSub(LHS, RHS, OrigI))
3878 return SetResult(Builder.CreateNUWSub(LHS, RHS), Builder.getFalse(),
3884 case OCF_UNSIGNED_MUL: {
3885 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
3886 if (OR == OverflowResult::NeverOverflows)
3887 return SetResult(Builder.CreateNUWMul(LHS, RHS), Builder.getFalse(),
3889 if (OR == OverflowResult::AlwaysOverflows)
3890 return SetResult(Builder.CreateMul(LHS, RHS), Builder.getTrue(), true);
3893 case OCF_SIGNED_MUL:
3894 // X * undef -> undef
3895 if (isa<UndefValue>(RHS))
3896 return SetResult(RHS, UndefValue::get(Builder.getInt1Ty()), false);
3898 // X * 0 -> {0, false}
3899 if (match(RHS, m_Zero()))
3900 return SetResult(RHS, Builder.getFalse(), false);
3902 // X * 1 -> {X, false}
3903 if (match(RHS, m_One()))
3904 return SetResult(LHS, Builder.getFalse(), false);
3906 if (OCF == OCF_SIGNED_MUL)
3907 if (willNotOverflowSignedMul(LHS, RHS, OrigI))
3908 return SetResult(Builder.CreateNSWMul(LHS, RHS), Builder.getFalse(),
3916 /// Recognize and process idiom involving test for multiplication
3919 /// The caller has matched a pattern of the form:
3920 /// I = cmp u (mul(zext A, zext B), V
3921 /// The function checks if this is a test for overflow and if so replaces
3922 /// multiplication with call to 'mul.with.overflow' intrinsic.
3924 /// \param I Compare instruction.
3925 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
3926 /// the compare instruction. Must be of integer type.
3927 /// \param OtherVal The other argument of compare instruction.
3928 /// \returns Instruction which must replace the compare instruction, NULL if no
3929 /// replacement required.
3930 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
3931 Value *OtherVal, InstCombiner &IC) {
3932 // Don't bother doing this transformation for pointers, don't do it for
3934 if (!isa<IntegerType>(MulVal->getType()))
3937 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
3938 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
3939 auto *MulInstr = dyn_cast<Instruction>(MulVal);
3942 assert(MulInstr->getOpcode() == Instruction::Mul);
3944 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
3945 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
3946 assert(LHS->getOpcode() == Instruction::ZExt);
3947 assert(RHS->getOpcode() == Instruction::ZExt);
3948 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
3950 // Calculate type and width of the result produced by mul.with.overflow.
3951 Type *TyA = A->getType(), *TyB = B->getType();
3952 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
3953 WidthB = TyB->getPrimitiveSizeInBits();
3956 if (WidthB > WidthA) {
3964 // In order to replace the original mul with a narrower mul.with.overflow,
3965 // all uses must ignore upper bits of the product. The number of used low
3966 // bits must be not greater than the width of mul.with.overflow.
3967 if (MulVal->hasNUsesOrMore(2))
3968 for (User *U : MulVal->users()) {
3971 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3972 // Check if truncation ignores bits above MulWidth.
3973 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
3974 if (TruncWidth > MulWidth)
3976 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3977 // Check if AND ignores bits above MulWidth.
3978 if (BO->getOpcode() != Instruction::And)
3980 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3981 const APInt &CVal = CI->getValue();
3982 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
3985 // In this case we could have the operand of the binary operation
3986 // being defined in another block, and performing the replacement
3987 // could break the dominance relation.
3991 // Other uses prohibit this transformation.
3996 // Recognize patterns
3997 switch (I.getPredicate()) {
3998 case ICmpInst::ICMP_EQ:
3999 case ICmpInst::ICMP_NE:
4000 // Recognize pattern:
4001 // mulval = mul(zext A, zext B)
4002 // cmp eq/neq mulval, zext trunc mulval
4003 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
4004 if (Zext->hasOneUse()) {
4005 Value *ZextArg = Zext->getOperand(0);
4006 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
4007 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
4011 // Recognize pattern:
4012 // mulval = mul(zext A, zext B)
4013 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4016 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4017 if (ValToMask != MulVal)
4019 const APInt &CVal = CI->getValue() + 1;
4020 if (CVal.isPowerOf2()) {
4021 unsigned MaskWidth = CVal.logBase2();
4022 if (MaskWidth == MulWidth)
4023 break; // Recognized
4028 case ICmpInst::ICMP_UGT:
4029 // Recognize pattern:
4030 // mulval = mul(zext A, zext B)
4031 // cmp ugt mulval, max
4032 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4033 APInt MaxVal = APInt::getMaxValue(MulWidth);
4034 MaxVal = MaxVal.zext(CI->getBitWidth());
4035 if (MaxVal.eq(CI->getValue()))
4036 break; // Recognized
4040 case ICmpInst::ICMP_UGE:
4041 // Recognize pattern:
4042 // mulval = mul(zext A, zext B)
4043 // cmp uge mulval, max+1
4044 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4045 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4046 if (MaxVal.eq(CI->getValue()))
4047 break; // Recognized
4051 case ICmpInst::ICMP_ULE:
4052 // Recognize pattern:
4053 // mulval = mul(zext A, zext B)
4054 // cmp ule mulval, max
4055 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4056 APInt MaxVal = APInt::getMaxValue(MulWidth);
4057 MaxVal = MaxVal.zext(CI->getBitWidth());
4058 if (MaxVal.eq(CI->getValue()))
4059 break; // Recognized
4063 case ICmpInst::ICMP_ULT:
4064 // Recognize pattern:
4065 // mulval = mul(zext A, zext B)
4066 // cmp ule mulval, max + 1
4067 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4068 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4069 if (MaxVal.eq(CI->getValue()))
4070 break; // Recognized
4078 InstCombiner::BuilderTy &Builder = IC.Builder;
4079 Builder.SetInsertPoint(MulInstr);
4081 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4082 Value *MulA = A, *MulB = B;
4083 if (WidthA < MulWidth)
4084 MulA = Builder.CreateZExt(A, MulType);
4085 if (WidthB < MulWidth)
4086 MulB = Builder.CreateZExt(B, MulType);
4087 Value *F = Intrinsic::getDeclaration(I.getModule(),
4088 Intrinsic::umul_with_overflow, MulType);
4089 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4090 IC.Worklist.Add(MulInstr);
4092 // If there are uses of mul result other than the comparison, we know that
4093 // they are truncation or binary AND. Change them to use result of
4094 // mul.with.overflow and adjust properly mask/size.
4095 if (MulVal->hasNUsesOrMore(2)) {
4096 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4097 for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4099 if (U == &I || U == OtherVal)
4101 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4102 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4103 IC.replaceInstUsesWith(*TI, Mul);
4105 TI->setOperand(0, Mul);
4106 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4107 assert(BO->getOpcode() == Instruction::And);
4108 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4109 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4110 APInt ShortMask = CI->getValue().trunc(MulWidth);
4111 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4113 cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4114 IC.Worklist.Add(Zext);
4115 IC.replaceInstUsesWith(*BO, Zext);
4117 llvm_unreachable("Unexpected Binary operation");
4119 IC.Worklist.Add(cast<Instruction>(U));
4122 if (isa<Instruction>(OtherVal))
4123 IC.Worklist.Add(cast<Instruction>(OtherVal));
4125 // The original icmp gets replaced with the overflow value, maybe inverted
4126 // depending on predicate.
4127 bool Inverse = false;
4128 switch (I.getPredicate()) {
4129 case ICmpInst::ICMP_NE:
4131 case ICmpInst::ICMP_EQ:
4134 case ICmpInst::ICMP_UGT:
4135 case ICmpInst::ICMP_UGE:
4136 if (I.getOperand(0) == MulVal)
4140 case ICmpInst::ICMP_ULT:
4141 case ICmpInst::ICMP_ULE:
4142 if (I.getOperand(1) == MulVal)
4147 llvm_unreachable("Unexpected predicate");
4150 Value *Res = Builder.CreateExtractValue(Call, 1);
4151 return BinaryOperator::CreateNot(Res);
4154 return ExtractValueInst::Create(Call, 1);
4157 /// When performing a comparison against a constant, it is possible that not all
4158 /// the bits in the LHS are demanded. This helper method computes the mask that
4160 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4162 if (!match(I.getOperand(1), m_APInt(RHS)))
4163 return APInt::getAllOnesValue(BitWidth);
4165 // If this is a normal comparison, it demands all bits. If it is a sign bit
4166 // comparison, it only demands the sign bit.
4168 if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4169 return APInt::getSignMask(BitWidth);
4171 switch (I.getPredicate()) {
4172 // For a UGT comparison, we don't care about any bits that
4173 // correspond to the trailing ones of the comparand. The value of these
4174 // bits doesn't impact the outcome of the comparison, because any value
4175 // greater than the RHS must differ in a bit higher than these due to carry.
4176 case ICmpInst::ICMP_UGT:
4177 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4179 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4180 // Any value less than the RHS must differ in a higher bit because of carries.
4181 case ICmpInst::ICMP_ULT:
4182 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4185 return APInt::getAllOnesValue(BitWidth);
4189 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4190 /// should be swapped.
4191 /// The decision is based on how many times these two operands are reused
4192 /// as subtract operands and their positions in those instructions.
4193 /// The rationale is that several architectures use the same instruction for
4194 /// both subtract and cmp. Thus, it is better if the order of those operands
4196 /// \return true if Op0 and Op1 should be swapped.
4197 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4198 // Filter out pointer values as those cannot appear directly in subtract.
4199 // FIXME: we may want to go through inttoptrs or bitcasts.
4200 if (Op0->getType()->isPointerTy())
4202 // If a subtract already has the same operands as a compare, swapping would be
4203 // bad. If a subtract has the same operands as a compare but in reverse order,
4204 // then swapping is good.
4206 for (const User *U : Op0->users()) {
4207 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4209 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4212 return GoodToSwap > 0;
4215 /// Check that one use is in the same block as the definition and all
4216 /// other uses are in blocks dominated by a given block.
4218 /// \param DI Definition
4220 /// \param DB Block that must dominate all uses of \p DI outside
4221 /// the parent block
4222 /// \return true when \p UI is the only use of \p DI in the parent block
4223 /// and all other uses of \p DI are in blocks dominated by \p DB.
4225 bool InstCombiner::dominatesAllUses(const Instruction *DI,
4226 const Instruction *UI,
4227 const BasicBlock *DB) const {
4228 assert(DI && UI && "Instruction not defined\n");
4229 // Ignore incomplete definitions.
4230 if (!DI->getParent())
4232 // DI and UI must be in the same block.
4233 if (DI->getParent() != UI->getParent())
4235 // Protect from self-referencing blocks.
4236 if (DI->getParent() == DB)
4238 for (const User *U : DI->users()) {
4239 auto *Usr = cast<Instruction>(U);
4240 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4246 /// Return true when the instruction sequence within a block is select-cmp-br.
4247 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4248 const BasicBlock *BB = SI->getParent();
4251 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4252 if (!BI || BI->getNumSuccessors() != 2)
4254 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4255 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4260 /// True when a select result is replaced by one of its operands
4261 /// in select-icmp sequence. This will eventually result in the elimination
4264 /// \param SI Select instruction
4265 /// \param Icmp Compare instruction
4266 /// \param SIOpd Operand that replaces the select
4269 /// - The replacement is global and requires dominator information
4270 /// - The caller is responsible for the actual replacement
4275 /// %4 = select i1 %3, %C* %0, %C* null
4276 /// %5 = icmp eq %C* %4, null
4277 /// br i1 %5, label %9, label %7
4279 /// ; <label>:7 ; preds = %entry
4280 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4283 /// can be transformed to
4285 /// %5 = icmp eq %C* %0, null
4286 /// %6 = select i1 %3, i1 %5, i1 true
4287 /// br i1 %6, label %9, label %7
4289 /// ; <label>:7 ; preds = %entry
4290 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4292 /// Similar when the first operand of the select is a constant or/and
4293 /// the compare is for not equal rather than equal.
4295 /// NOTE: The function is only called when the select and compare constants
4296 /// are equal, the optimization can work only for EQ predicates. This is not a
4297 /// major restriction since a NE compare should be 'normalized' to an equal
4298 /// compare, which usually happens in the combiner and test case
4299 /// select-cmp-br.ll checks for it.
4300 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
4301 const ICmpInst *Icmp,
4302 const unsigned SIOpd) {
4303 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4304 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4305 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4306 // The check for the single predecessor is not the best that can be
4307 // done. But it protects efficiently against cases like when SI's
4308 // home block has two successors, Succ and Succ1, and Succ1 predecessor
4309 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4310 // replaced can be reached on either path. So the uniqueness check
4311 // guarantees that the path all uses of SI (outside SI's parent) are on
4312 // is disjoint from all other paths out of SI. But that information
4313 // is more expensive to compute, and the trade-off here is in favor
4314 // of compile-time. It should also be noticed that we check for a single
4315 // predecessor and not only uniqueness. This to handle the situation when
4316 // Succ and Succ1 points to the same basic block.
4317 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4319 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4326 /// Try to fold the comparison based on range information we can get by checking
4327 /// whether bits are known to be zero or one in the inputs.
4328 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4329 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4330 Type *Ty = Op0->getType();
4331 ICmpInst::Predicate Pred = I.getPredicate();
4333 // Get scalar or pointer size.
4334 unsigned BitWidth = Ty->isIntOrIntVectorTy()
4335 ? Ty->getScalarSizeInBits()
4336 : DL.getIndexTypeSizeInBits(Ty->getScalarType());
4341 KnownBits Op0Known(BitWidth);
4342 KnownBits Op1Known(BitWidth);
4344 if (SimplifyDemandedBits(&I, 0,
4345 getDemandedBitsLHSMask(I, BitWidth),
4349 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4353 // Given the known and unknown bits, compute a range that the LHS could be
4354 // in. Compute the Min, Max and RHS values based on the known bits. For the
4355 // EQ and NE we use unsigned values.
4356 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4357 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4359 computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4360 computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4362 computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4363 computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4366 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4367 // out that the LHS or RHS is a constant. Constant fold this now, so that
4368 // code below can assume that Min != Max.
4369 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4370 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4371 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4372 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4374 // Based on the range information we know about the LHS, see if we can
4375 // simplify this comparison. For example, (x&4) < 8 is always true.
4378 llvm_unreachable("Unknown icmp opcode!");
4379 case ICmpInst::ICMP_EQ:
4380 case ICmpInst::ICMP_NE: {
4381 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4382 return Pred == CmpInst::ICMP_EQ
4383 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4384 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4387 // If all bits are known zero except for one, then we know at most one bit
4388 // is set. If the comparison is against zero, then this is a check to see if
4389 // *that* bit is set.
4390 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4391 if (Op1Known.isZero()) {
4392 // If the LHS is an AND with the same constant, look through it.
4393 Value *LHS = nullptr;
4395 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4396 *LHSC != Op0KnownZeroInverted)
4400 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4401 APInt ValToCheck = Op0KnownZeroInverted;
4402 Type *XTy = X->getType();
4403 if (ValToCheck.isPowerOf2()) {
4404 // ((1 << X) & 8) == 0 -> X != 3
4405 // ((1 << X) & 8) != 0 -> X == 3
4406 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4407 auto NewPred = ICmpInst::getInversePredicate(Pred);
4408 return new ICmpInst(NewPred, X, CmpC);
4409 } else if ((++ValToCheck).isPowerOf2()) {
4410 // ((1 << X) & 7) == 0 -> X >= 3
4411 // ((1 << X) & 7) != 0 -> X < 3
4412 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4414 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
4415 return new ICmpInst(NewPred, X, CmpC);
4419 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4421 if (Op0KnownZeroInverted.isOneValue() &&
4422 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4423 // ((8 >>u X) & 1) == 0 -> X != 3
4424 // ((8 >>u X) & 1) != 0 -> X == 3
4425 unsigned CmpVal = CI->countTrailingZeros();
4426 auto NewPred = ICmpInst::getInversePredicate(Pred);
4427 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4432 case ICmpInst::ICMP_ULT: {
4433 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4434 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4435 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4436 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4437 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4438 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4441 if (match(Op1, m_APInt(CmpC))) {
4442 // A <u C -> A == C-1 if min(A)+1 == C
4443 if (*CmpC == Op0Min + 1)
4444 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4445 ConstantInt::get(Op1->getType(), *CmpC - 1));
4446 // X <u C --> X == 0, if the number of zero bits in the bottom of X
4447 // exceeds the log2 of C.
4448 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
4449 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4450 Constant::getNullValue(Op1->getType()));
4454 case ICmpInst::ICMP_UGT: {
4455 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4456 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4457 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4458 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4459 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4460 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4463 if (match(Op1, m_APInt(CmpC))) {
4464 // A >u C -> A == C+1 if max(a)-1 == C
4465 if (*CmpC == Op0Max - 1)
4466 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4467 ConstantInt::get(Op1->getType(), *CmpC + 1));
4468 // X >u C --> X != 0, if the number of zero bits in the bottom of X
4469 // exceeds the log2 of C.
4470 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
4471 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
4472 Constant::getNullValue(Op1->getType()));
4476 case ICmpInst::ICMP_SLT: {
4477 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4478 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4479 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4480 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4481 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4482 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4484 if (match(Op1, m_APInt(CmpC))) {
4485 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4486 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4487 ConstantInt::get(Op1->getType(), *CmpC - 1));
4491 case ICmpInst::ICMP_SGT: {
4492 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4493 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4494 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4495 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4496 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4497 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4499 if (match(Op1, m_APInt(CmpC))) {
4500 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4501 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4502 ConstantInt::get(Op1->getType(), *CmpC + 1));
4506 case ICmpInst::ICMP_SGE:
4507 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4508 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4509 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4510 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4511 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4512 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
4513 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4515 case ICmpInst::ICMP_SLE:
4516 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4517 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4518 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4519 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4520 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4521 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
4522 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4524 case ICmpInst::ICMP_UGE:
4525 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4526 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4527 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4528 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4529 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4530 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
4531 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4533 case ICmpInst::ICMP_ULE:
4534 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4535 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4536 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4537 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4538 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4539 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
4540 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4544 // Turn a signed comparison into an unsigned one if both operands are known to
4545 // have the same sign.
4547 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
4548 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
4549 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4554 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4555 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4556 /// allows them to be folded in visitICmpInst.
4557 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
4558 ICmpInst::Predicate Pred = I.getPredicate();
4559 if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4560 Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4563 Value *Op0 = I.getOperand(0);
4564 Value *Op1 = I.getOperand(1);
4565 auto *Op1C = dyn_cast<Constant>(Op1);
4569 // Check if the constant operand can be safely incremented/decremented without
4570 // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4571 // the edge cases for us, so we just assert on them. For vectors, we must
4572 // handle the edge cases.
4573 Type *Op1Type = Op1->getType();
4574 bool IsSigned = I.isSigned();
4575 bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4576 auto *CI = dyn_cast<ConstantInt>(Op1C);
4578 // A <= MAX -> TRUE ; A >= MIN -> TRUE
4579 assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4580 } else if (Op1Type->isVectorTy()) {
4581 // TODO? If the edge cases for vectors were guaranteed to be handled as they
4582 // are for scalar, we could remove the min/max checks. However, to do that,
4583 // we would have to use insertelement/shufflevector to replace edge values.
4584 unsigned NumElts = Op1Type->getVectorNumElements();
4585 for (unsigned i = 0; i != NumElts; ++i) {
4586 Constant *Elt = Op1C->getAggregateElement(i);
4590 if (isa<UndefValue>(Elt))
4593 // Bail out if we can't determine if this constant is min/max or if we
4594 // know that this constant is min/max.
4595 auto *CI = dyn_cast<ConstantInt>(Elt);
4596 if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4604 // Increment or decrement the constant and set the new comparison predicate:
4605 // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4606 Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4607 CmpInst::Predicate NewPred = IsLE ? ICmpInst::ICMP_ULT: ICmpInst::ICMP_UGT;
4608 NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4609 return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4612 /// Integer compare with boolean values can always be turned into bitwise ops.
4613 static Instruction *canonicalizeICmpBool(ICmpInst &I,
4614 InstCombiner::BuilderTy &Builder) {
4615 Value *A = I.getOperand(0), *B = I.getOperand(1);
4616 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
4618 // A boolean compared to true/false can be simplified to Op0/true/false in
4619 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
4620 // Cases not handled by InstSimplify are always 'not' of Op0.
4621 if (match(B, m_Zero())) {
4622 switch (I.getPredicate()) {
4623 case CmpInst::ICMP_EQ: // A == 0 -> !A
4624 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
4625 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
4626 return BinaryOperator::CreateNot(A);
4628 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4630 } else if (match(B, m_One())) {
4631 switch (I.getPredicate()) {
4632 case CmpInst::ICMP_NE: // A != 1 -> !A
4633 case CmpInst::ICMP_ULT: // A <u 1 -> !A
4634 case CmpInst::ICMP_SGT: // A >s -1 -> !A
4635 return BinaryOperator::CreateNot(A);
4637 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4641 switch (I.getPredicate()) {
4643 llvm_unreachable("Invalid icmp instruction!");
4644 case ICmpInst::ICMP_EQ:
4645 // icmp eq i1 A, B -> ~(A ^ B)
4646 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
4648 case ICmpInst::ICMP_NE:
4649 // icmp ne i1 A, B -> A ^ B
4650 return BinaryOperator::CreateXor(A, B);
4652 case ICmpInst::ICMP_UGT:
4653 // icmp ugt -> icmp ult
4656 case ICmpInst::ICMP_ULT:
4657 // icmp ult i1 A, B -> ~A & B
4658 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
4660 case ICmpInst::ICMP_SGT:
4661 // icmp sgt -> icmp slt
4664 case ICmpInst::ICMP_SLT:
4665 // icmp slt i1 A, B -> A & ~B
4666 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
4668 case ICmpInst::ICMP_UGE:
4669 // icmp uge -> icmp ule
4672 case ICmpInst::ICMP_ULE:
4673 // icmp ule i1 A, B -> ~A | B
4674 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
4676 case ICmpInst::ICMP_SGE:
4677 // icmp sge -> icmp sle
4680 case ICmpInst::ICMP_SLE:
4681 // icmp sle i1 A, B -> A | ~B
4682 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
4686 // Transform pattern like:
4687 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
4688 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
4692 static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
4693 InstCombiner::BuilderTy &Builder) {
4694 ICmpInst::Predicate Pred, NewPred;
4697 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
4698 // We want X to be the icmp's second operand, so swap predicate if it isn't.
4699 if (Cmp.getOperand(0) == X)
4700 Pred = Cmp.getSwappedPredicate();
4703 case ICmpInst::ICMP_ULE:
4704 NewPred = ICmpInst::ICMP_NE;
4706 case ICmpInst::ICMP_UGT:
4707 NewPred = ICmpInst::ICMP_EQ;
4712 } else if (match(&Cmp, m_c_ICmp(Pred,
4713 m_OneUse(m_CombineOr(
4714 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
4715 m_Add(m_Shl(m_One(), m_Value(Y)),
4718 // The variant with 'add' is not canonical, (the variant with 'not' is)
4719 // we only get it because it has extra uses, and can't be canonicalized,
4721 // We want X to be the icmp's second operand, so swap predicate if it isn't.
4722 if (Cmp.getOperand(0) == X)
4723 Pred = Cmp.getSwappedPredicate();
4726 case ICmpInst::ICMP_ULT:
4727 NewPred = ICmpInst::ICMP_NE;
4729 case ICmpInst::ICMP_UGE:
4730 NewPred = ICmpInst::ICMP_EQ;
4738 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
4739 Constant *Zero = Constant::getNullValue(NewX->getType());
4740 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
4743 static Instruction *foldVectorCmp(CmpInst &Cmp,
4744 InstCombiner::BuilderTy &Builder) {
4745 // If both arguments of the cmp are shuffles that use the same mask and
4746 // shuffle within a single vector, move the shuffle after the cmp.
4747 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
4750 if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(M))) &&
4751 match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(M))) &&
4752 V1->getType() == V2->getType() &&
4753 (LHS->hasOneUse() || RHS->hasOneUse())) {
4754 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
4755 CmpInst::Predicate P = Cmp.getPredicate();
4756 Value *NewCmp = isa<ICmpInst>(Cmp) ? Builder.CreateICmp(P, V1, V2)
4757 : Builder.CreateFCmp(P, V1, V2);
4758 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
4763 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4764 bool Changed = false;
4765 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4766 unsigned Op0Cplxity = getComplexity(Op0);
4767 unsigned Op1Cplxity = getComplexity(Op1);
4769 /// Orders the operands of the compare so that they are listed from most
4770 /// complex to least complex. This puts constants before unary operators,
4771 /// before binary operators.
4772 if (Op0Cplxity < Op1Cplxity ||
4773 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
4775 std::swap(Op0, Op1);
4779 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1,
4780 SQ.getWithInstruction(&I)))
4781 return replaceInstUsesWith(I, V);
4783 // Comparing -val or val with non-zero is the same as just comparing val
4784 // ie, abs(val) != 0 -> val != 0
4785 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
4786 Value *Cond, *SelectTrue, *SelectFalse;
4787 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
4788 m_Value(SelectFalse)))) {
4789 if (Value *V = dyn_castNegVal(SelectTrue)) {
4790 if (V == SelectFalse)
4791 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4793 else if (Value *V = dyn_castNegVal(SelectFalse)) {
4794 if (V == SelectTrue)
4795 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4800 if (Op0->getType()->isIntOrIntVectorTy(1))
4801 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
4804 if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
4807 if (Instruction *Res = foldICmpWithConstant(I))
4810 if (Instruction *Res = foldICmpWithDominatingICmp(I))
4813 if (Instruction *Res = foldICmpUsingKnownBits(I))
4816 // Test if the ICmpInst instruction is used exclusively by a select as
4817 // part of a minimum or maximum operation. If so, refrain from doing
4818 // any other folding. This helps out other analyses which understand
4819 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4820 // and CodeGen. And in this case, at least one of the comparison
4821 // operands has at least one user besides the compare (the select),
4822 // which would often largely negate the benefit of folding anyway.
4824 // Do the same for the other patterns recognized by matchSelectPattern.
4826 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
4828 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
4829 if (SPR.Flavor != SPF_UNKNOWN)
4833 // Do this after checking for min/max to prevent infinite looping.
4834 if (Instruction *Res = foldICmpWithZero(I))
4837 // FIXME: We only do this after checking for min/max to prevent infinite
4838 // looping caused by a reverse canonicalization of these patterns for min/max.
4839 // FIXME: The organization of folds is a mess. These would naturally go into
4840 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
4841 // down here after the min/max restriction.
4842 ICmpInst::Predicate Pred = I.getPredicate();
4844 if (match(Op1, m_APInt(C))) {
4845 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
4846 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
4847 Constant *Zero = Constant::getNullValue(Op0->getType());
4848 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
4851 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
4852 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
4853 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
4854 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
4858 if (Instruction *Res = foldICmpInstWithConstant(I))
4861 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
4864 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4865 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
4866 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
4868 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
4869 if (Instruction *NI = foldGEPICmp(GEP, Op0,
4870 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
4873 // Try to optimize equality comparisons against alloca-based pointers.
4874 if (Op0->getType()->isPointerTy() && I.isEquality()) {
4875 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
4876 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
4877 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
4879 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
4880 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
4884 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
4886 if (match(Op0, m_BitCast(m_SIToFP(m_Value(X))))) {
4887 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
4888 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
4889 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
4890 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
4891 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
4892 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
4893 match(Op1, m_Zero()))
4894 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
4896 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
4897 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
4898 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
4900 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
4901 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
4902 return new ICmpInst(Pred, X, ConstantInt::getAllOnesValue(X->getType()));
4905 // Zero-equality checks are preserved through unsigned floating-point casts:
4906 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
4907 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
4908 if (match(Op0, m_BitCast(m_UIToFP(m_Value(X)))))
4909 if (I.isEquality() && match(Op1, m_Zero()))
4910 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
4912 // Test to see if the operands of the icmp are casted versions of other
4913 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4915 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4916 if (Op0->getType()->isPointerTy() &&
4917 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4918 // We keep moving the cast from the left operand over to the right
4919 // operand, where it can often be eliminated completely.
4920 Op0 = CI->getOperand(0);
4922 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
4923 // so eliminate it as well.
4924 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
4925 Op1 = CI2->getOperand(0);
4927 // If Op1 is a constant, we can fold the cast into the constant.
4928 if (Op0->getType() != Op1->getType()) {
4929 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4930 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4932 // Otherwise, cast the RHS right before the icmp
4933 Op1 = Builder.CreateBitCast(Op1, Op0->getType());
4936 return new ICmpInst(I.getPredicate(), Op0, Op1);
4940 if (isa<CastInst>(Op0)) {
4941 // Handle the special case of: icmp (cast bool to X), <cst>
4942 // This comes up when you have code like
4945 // For generality, we handle any zero-extension of any operand comparison
4946 // with a constant or another cast from the same type.
4947 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
4948 if (Instruction *R = foldICmpWithCastAndCast(I))
4952 if (Instruction *Res = foldICmpBinOp(I))
4955 if (Instruction *Res = foldICmpWithMinMax(I))
4960 // Transform (A & ~B) == 0 --> (A & B) != 0
4961 // and (A & ~B) != 0 --> (A & B) == 0
4962 // if A is a power of 2.
4963 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
4964 match(Op1, m_Zero()) &&
4965 isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
4966 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
4969 // ~X < ~Y --> Y < X
4970 // ~X < C --> X > ~C
4971 if (match(Op0, m_Not(m_Value(A)))) {
4972 if (match(Op1, m_Not(m_Value(B))))
4973 return new ICmpInst(I.getPredicate(), B, A);
4976 if (match(Op1, m_APInt(C)))
4977 return new ICmpInst(I.getSwappedPredicate(), A,
4978 ConstantInt::get(Op1->getType(), ~(*C)));
4981 Instruction *AddI = nullptr;
4982 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
4983 m_Instruction(AddI))) &&
4984 isa<IntegerType>(A->getType())) {
4987 if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result,
4989 replaceInstUsesWith(*AddI, Result);
4990 return replaceInstUsesWith(I, Overflow);
4994 // (zext a) * (zext b) --> llvm.umul.with.overflow.
4995 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4996 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
4999 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5000 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
5005 if (Instruction *Res = foldICmpEquality(I))
5008 // The 'cmpxchg' instruction returns an aggregate containing the old value and
5009 // an i1 which indicates whether or not we successfully did the swap.
5011 // Replace comparisons between the old value and the expected value with the
5012 // indicator that 'cmpxchg' returns.
5014 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
5015 // spuriously fail. In those cases, the old value may equal the expected
5016 // value but it is possible for the swap to not occur.
5017 if (I.getPredicate() == ICmpInst::ICMP_EQ)
5018 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
5019 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
5020 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
5022 return ExtractValueInst::Create(ACXI, 1);
5028 if (match(Op0, m_Add(m_Value(X), m_APInt(C))) && Op1 == X)
5029 return foldICmpAddOpConst(X, *C, I.getPredicate());
5032 if (match(Op1, m_Add(m_Value(X), m_APInt(C))) && Op0 == X)
5033 return foldICmpAddOpConst(X, *C, I.getSwappedPredicate());
5036 if (Instruction *Res = foldICmpWithHighBitMask(I, Builder))
5039 if (I.getType()->isVectorTy())
5040 if (Instruction *Res = foldVectorCmp(I, Builder))
5043 return Changed ? &I : nullptr;
5046 /// Fold fcmp ([us]itofp x, cst) if possible.
5047 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
5049 if (!isa<ConstantFP>(RHSC)) return nullptr;
5050 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5052 // Get the width of the mantissa. We don't want to hack on conversions that
5053 // might lose information from the integer, e.g. "i64 -> float"
5054 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5055 if (MantissaWidth == -1) return nullptr; // Unknown.
5057 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5059 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5061 if (I.isEquality()) {
5062 FCmpInst::Predicate P = I.getPredicate();
5063 bool IsExact = false;
5064 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
5065 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
5067 // If the floating point constant isn't an integer value, we know if we will
5068 // ever compare equal / not equal to it.
5070 // TODO: Can never be -0.0 and other non-representable values
5071 APFloat RHSRoundInt(RHS);
5072 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
5073 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
5074 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
5075 return replaceInstUsesWith(I, Builder.getFalse());
5077 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
5078 return replaceInstUsesWith(I, Builder.getTrue());
5082 // TODO: If the constant is exactly representable, is it always OK to do
5083 // equality compares as integer?
5086 // Check to see that the input is converted from an integer type that is small
5087 // enough that preserves all bits. TODO: check here for "known" sign bits.
5088 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5089 unsigned InputSize = IntTy->getScalarSizeInBits();
5091 // Following test does NOT adjust InputSize downwards for signed inputs,
5092 // because the most negative value still requires all the mantissa bits
5093 // to distinguish it from one less than that value.
5094 if ((int)InputSize > MantissaWidth) {
5095 // Conversion would lose accuracy. Check if loss can impact comparison.
5096 int Exp = ilogb(RHS);
5097 if (Exp == APFloat::IEK_Inf) {
5098 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
5099 if (MaxExponent < (int)InputSize - !LHSUnsigned)
5100 // Conversion could create infinity.
5103 // Note that if RHS is zero or NaN, then Exp is negative
5104 // and first condition is trivially false.
5105 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
5106 // Conversion could affect comparison.
5111 // Otherwise, we can potentially simplify the comparison. We know that it
5112 // will always come through as an integer value and we know the constant is
5113 // not a NAN (it would have been previously simplified).
5114 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5116 ICmpInst::Predicate Pred;
5117 switch (I.getPredicate()) {
5118 default: llvm_unreachable("Unexpected predicate!");
5119 case FCmpInst::FCMP_UEQ:
5120 case FCmpInst::FCMP_OEQ:
5121 Pred = ICmpInst::ICMP_EQ;
5123 case FCmpInst::FCMP_UGT:
5124 case FCmpInst::FCMP_OGT:
5125 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5127 case FCmpInst::FCMP_UGE:
5128 case FCmpInst::FCMP_OGE:
5129 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5131 case FCmpInst::FCMP_ULT:
5132 case FCmpInst::FCMP_OLT:
5133 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5135 case FCmpInst::FCMP_ULE:
5136 case FCmpInst::FCMP_OLE:
5137 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5139 case FCmpInst::FCMP_UNE:
5140 case FCmpInst::FCMP_ONE:
5141 Pred = ICmpInst::ICMP_NE;
5143 case FCmpInst::FCMP_ORD:
5144 return replaceInstUsesWith(I, Builder.getTrue());
5145 case FCmpInst::FCMP_UNO:
5146 return replaceInstUsesWith(I, Builder.getFalse());
5149 // Now we know that the APFloat is a normal number, zero or inf.
5151 // See if the FP constant is too large for the integer. For example,
5152 // comparing an i8 to 300.0.
5153 unsigned IntWidth = IntTy->getScalarSizeInBits();
5156 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5157 // and large values.
5158 APFloat SMax(RHS.getSemantics());
5159 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5160 APFloat::rmNearestTiesToEven);
5161 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5162 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5163 Pred == ICmpInst::ICMP_SLE)
5164 return replaceInstUsesWith(I, Builder.getTrue());
5165 return replaceInstUsesWith(I, Builder.getFalse());
5168 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5169 // +INF and large values.
5170 APFloat UMax(RHS.getSemantics());
5171 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5172 APFloat::rmNearestTiesToEven);
5173 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5174 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5175 Pred == ICmpInst::ICMP_ULE)
5176 return replaceInstUsesWith(I, Builder.getTrue());
5177 return replaceInstUsesWith(I, Builder.getFalse());
5182 // See if the RHS value is < SignedMin.
5183 APFloat SMin(RHS.getSemantics());
5184 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5185 APFloat::rmNearestTiesToEven);
5186 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5187 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5188 Pred == ICmpInst::ICMP_SGE)
5189 return replaceInstUsesWith(I, Builder.getTrue());
5190 return replaceInstUsesWith(I, Builder.getFalse());
5193 // See if the RHS value is < UnsignedMin.
5194 APFloat SMin(RHS.getSemantics());
5195 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
5196 APFloat::rmNearestTiesToEven);
5197 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
5198 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
5199 Pred == ICmpInst::ICMP_UGE)
5200 return replaceInstUsesWith(I, Builder.getTrue());
5201 return replaceInstUsesWith(I, Builder.getFalse());
5205 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5206 // [0, UMAX], but it may still be fractional. See if it is fractional by
5207 // casting the FP value to the integer value and back, checking for equality.
5208 // Don't do this for zero, because -0.0 is not fractional.
5209 Constant *RHSInt = LHSUnsigned
5210 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5211 : ConstantExpr::getFPToSI(RHSC, IntTy);
5212 if (!RHS.isZero()) {
5213 bool Equal = LHSUnsigned
5214 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5215 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5217 // If we had a comparison against a fractional value, we have to adjust
5218 // the compare predicate and sometimes the value. RHSC is rounded towards
5219 // zero at this point.
5221 default: llvm_unreachable("Unexpected integer comparison!");
5222 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5223 return replaceInstUsesWith(I, Builder.getTrue());
5224 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5225 return replaceInstUsesWith(I, Builder.getFalse());
5226 case ICmpInst::ICMP_ULE:
5227 // (float)int <= 4.4 --> int <= 4
5228 // (float)int <= -4.4 --> false
5229 if (RHS.isNegative())
5230 return replaceInstUsesWith(I, Builder.getFalse());
5232 case ICmpInst::ICMP_SLE:
5233 // (float)int <= 4.4 --> int <= 4
5234 // (float)int <= -4.4 --> int < -4
5235 if (RHS.isNegative())
5236 Pred = ICmpInst::ICMP_SLT;
5238 case ICmpInst::ICMP_ULT:
5239 // (float)int < -4.4 --> false
5240 // (float)int < 4.4 --> int <= 4
5241 if (RHS.isNegative())
5242 return replaceInstUsesWith(I, Builder.getFalse());
5243 Pred = ICmpInst::ICMP_ULE;
5245 case ICmpInst::ICMP_SLT:
5246 // (float)int < -4.4 --> int < -4
5247 // (float)int < 4.4 --> int <= 4
5248 if (!RHS.isNegative())
5249 Pred = ICmpInst::ICMP_SLE;
5251 case ICmpInst::ICMP_UGT:
5252 // (float)int > 4.4 --> int > 4
5253 // (float)int > -4.4 --> true
5254 if (RHS.isNegative())
5255 return replaceInstUsesWith(I, Builder.getTrue());
5257 case ICmpInst::ICMP_SGT:
5258 // (float)int > 4.4 --> int > 4
5259 // (float)int > -4.4 --> int >= -4
5260 if (RHS.isNegative())
5261 Pred = ICmpInst::ICMP_SGE;
5263 case ICmpInst::ICMP_UGE:
5264 // (float)int >= -4.4 --> true
5265 // (float)int >= 4.4 --> int > 4
5266 if (RHS.isNegative())
5267 return replaceInstUsesWith(I, Builder.getTrue());
5268 Pred = ICmpInst::ICMP_UGT;
5270 case ICmpInst::ICMP_SGE:
5271 // (float)int >= -4.4 --> int >= -4
5272 // (float)int >= 4.4 --> int > 4
5273 if (!RHS.isNegative())
5274 Pred = ICmpInst::ICMP_SGT;
5280 // Lower this FP comparison into an appropriate integer version of the
5282 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5285 /// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary.
5286 static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI,
5288 // When C is not 0.0 and infinities are not allowed:
5289 // (C / X) < 0.0 is a sign-bit test of X
5290 // (C / X) < 0.0 --> X < 0.0 (if C is positive)
5291 // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate)
5294 // Multiply (C / X) < 0.0 by X * X / C.
5295 // - X is non zero, if it is the flag 'ninf' is violated.
5296 // - C defines the sign of X * X * C. Thus it also defines whether to swap
5297 // the predicate. C is also non zero by definition.
5299 // Thus X * X / C is non zero and the transformation is valid. [qed]
5301 FCmpInst::Predicate Pred = I.getPredicate();
5303 // Check that predicates are valid.
5304 if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) &&
5305 (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE))
5308 // Check that RHS operand is zero.
5309 if (!match(RHSC, m_AnyZeroFP()))
5312 // Check fastmath flags ('ninf').
5313 if (!LHSI->hasNoInfs() || !I.hasNoInfs())
5316 // Check the properties of the dividend. It must not be zero to avoid a
5317 // division by zero (see Proof).
5319 if (!match(LHSI->getOperand(0), m_APFloat(C)))
5325 // Get swapped predicate if necessary.
5326 if (C->isNegative())
5327 Pred = I.getSwappedPredicate();
5329 return new FCmpInst(Pred, LHSI->getOperand(1), RHSC, "", &I);
5332 /// Optimize fabs(X) compared with zero.
5333 static Instruction *foldFabsWithFcmpZero(FCmpInst &I) {
5335 if (!match(I.getOperand(0), m_Intrinsic<Intrinsic::fabs>(m_Value(X))) ||
5336 !match(I.getOperand(1), m_PosZeroFP()))
5339 auto replacePredAndOp0 = [](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
5341 I->setOperand(0, X);
5345 switch (I.getPredicate()) {
5346 case FCmpInst::FCMP_UGE:
5347 case FCmpInst::FCMP_OLT:
5348 // fabs(X) >= 0.0 --> true
5349 // fabs(X) < 0.0 --> false
5350 llvm_unreachable("fcmp should have simplified");
5352 case FCmpInst::FCMP_OGT:
5353 // fabs(X) > 0.0 --> X != 0.0
5354 return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
5356 case FCmpInst::FCMP_UGT:
5357 // fabs(X) u> 0.0 --> X u!= 0.0
5358 return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
5360 case FCmpInst::FCMP_OLE:
5361 // fabs(X) <= 0.0 --> X == 0.0
5362 return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
5364 case FCmpInst::FCMP_ULE:
5365 // fabs(X) u<= 0.0 --> X u== 0.0
5366 return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
5368 case FCmpInst::FCMP_OGE:
5369 // fabs(X) >= 0.0 --> !isnan(X)
5370 assert(!I.hasNoNaNs() && "fcmp should have simplified");
5371 return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
5373 case FCmpInst::FCMP_ULT:
5374 // fabs(X) u< 0.0 --> isnan(X)
5375 assert(!I.hasNoNaNs() && "fcmp should have simplified");
5376 return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
5378 case FCmpInst::FCMP_OEQ:
5379 case FCmpInst::FCMP_UEQ:
5380 case FCmpInst::FCMP_ONE:
5381 case FCmpInst::FCMP_UNE:
5382 case FCmpInst::FCMP_ORD:
5383 case FCmpInst::FCMP_UNO:
5384 // Look through the fabs() because it doesn't change anything but the sign.
5385 // fabs(X) == 0.0 --> X == 0.0,
5386 // fabs(X) != 0.0 --> X != 0.0
5387 // isnan(fabs(X)) --> isnan(X)
5388 // !isnan(fabs(X) --> !isnan(X)
5389 return replacePredAndOp0(&I, I.getPredicate(), X);
5396 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5397 bool Changed = false;
5399 /// Orders the operands of the compare so that they are listed from most
5400 /// complex to least complex. This puts constants before unary operators,
5401 /// before binary operators.
5402 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
5407 const CmpInst::Predicate Pred = I.getPredicate();
5408 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5409 if (Value *V = SimplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(),
5410 SQ.getWithInstruction(&I)))
5411 return replaceInstUsesWith(I, V);
5413 // Simplify 'fcmp pred X, X'
5417 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5418 case FCmpInst::FCMP_ULT: // True if unordered or less than
5419 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5420 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5421 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5422 I.setPredicate(FCmpInst::FCMP_UNO);
5423 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5426 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5427 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5428 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5429 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5430 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5431 I.setPredicate(FCmpInst::FCMP_ORD);
5432 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5437 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
5438 // then canonicalize the operand to 0.0.
5439 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
5440 if (!match(Op0, m_PosZeroFP()) && isKnownNeverNaN(Op0, &TLI)) {
5441 I.setOperand(0, ConstantFP::getNullValue(Op0->getType()));
5444 if (!match(Op1, m_PosZeroFP()) && isKnownNeverNaN(Op1, &TLI)) {
5445 I.setOperand(1, ConstantFP::getNullValue(Op0->getType()));
5450 // Test if the FCmpInst instruction is used exclusively by a select as
5451 // part of a minimum or maximum operation. If so, refrain from doing
5452 // any other folding. This helps out other analyses which understand
5453 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5454 // and CodeGen. And in this case, at least one of the comparison
5455 // operands has at least one user besides the compare (the select),
5456 // which would often largely negate the benefit of folding anyway.
5458 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
5460 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
5461 if (SPR.Flavor != SPF_UNKNOWN)
5465 // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0:
5466 // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0
5467 if (match(Op1, m_AnyZeroFP()) && !match(Op1, m_PosZeroFP())) {
5468 I.setOperand(1, ConstantFP::getNullValue(Op1->getType()));
5472 // Handle fcmp with instruction LHS and constant RHS.
5475 if (match(Op0, m_Instruction(LHSI)) && match(Op1, m_Constant(RHSC))) {
5476 switch (LHSI->getOpcode()) {
5477 case Instruction::PHI:
5478 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5479 // block. If in the same block, we're encouraging jump threading. If
5480 // not, we are just pessimizing the code by making an i1 phi.
5481 if (LHSI->getParent() == I.getParent())
5482 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
5485 case Instruction::SIToFP:
5486 case Instruction::UIToFP:
5487 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
5490 case Instruction::FDiv:
5491 if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC))
5494 case Instruction::Load:
5495 if (auto *GEP = dyn_cast<GetElementPtrInst>(LHSI->getOperand(0)))
5496 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
5497 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
5498 !cast<LoadInst>(LHSI)->isVolatile())
5499 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
5505 if (Instruction *R = foldFabsWithFcmpZero(I))
5509 if (match(Op0, m_FNeg(m_Value(X)))) {
5510 // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y
5511 if (match(Op1, m_FNeg(m_Value(Y))))
5512 return new FCmpInst(I.getSwappedPredicate(), X, Y, "", &I);
5514 // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
5516 if (match(Op1, m_Constant(C))) {
5517 Constant *NegC = ConstantExpr::getFNeg(C);
5518 return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
5522 if (match(Op0, m_FPExt(m_Value(X)))) {
5523 // fcmp (fpext X), (fpext Y) -> fcmp X, Y
5524 if (match(Op1, m_FPExt(m_Value(Y))) && X->getType() == Y->getType())
5525 return new FCmpInst(Pred, X, Y, "", &I);
5527 // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless
5529 if (match(Op1, m_APFloat(C))) {
5530 const fltSemantics &FPSem =
5531 X->getType()->getScalarType()->getFltSemantics();
5533 APFloat TruncC = *C;
5534 TruncC.convert(FPSem, APFloat::rmNearestTiesToEven, &Lossy);
5536 // Avoid lossy conversions and denormals.
5537 // Zero is a special case that's OK to convert.
5538 APFloat Fabs = TruncC;
5541 ((Fabs.compare(APFloat::getSmallestNormalized(FPSem)) !=
5542 APFloat::cmpLessThan) || Fabs.isZero())) {
5543 Constant *NewC = ConstantFP::get(X->getType(), TruncC);
5544 return new FCmpInst(Pred, X, NewC, "", &I);
5549 if (I.getType()->isVectorTy())
5550 if (Instruction *Res = foldVectorCmp(I, Builder))
5553 return Changed ? &I : nullptr;