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/MemoryBuiltins.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/VectorUtils.h"
23 #include "llvm/IR/ConstantRange.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/PatternMatch.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/KnownBits.h"
32 using namespace PatternMatch;
34 #define DEBUG_TYPE "instcombine"
36 // How many times is a select replaced by one of its operands?
37 STATISTIC(NumSel, "Number of select opts");
40 static ConstantInt *extractElement(Constant *V, Constant *Idx) {
41 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
44 static bool hasAddOverflow(ConstantInt *Result,
45 ConstantInt *In1, ConstantInt *In2,
48 return Result->getValue().ult(In1->getValue());
50 if (In2->isNegative())
51 return Result->getValue().sgt(In1->getValue());
52 return Result->getValue().slt(In1->getValue());
55 /// Compute Result = In1+In2, returning true if the result overflowed for this
57 static bool addWithOverflow(Constant *&Result, Constant *In1,
58 Constant *In2, bool IsSigned = false) {
59 Result = ConstantExpr::getAdd(In1, In2);
61 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
62 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
63 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
64 if (hasAddOverflow(extractElement(Result, Idx),
65 extractElement(In1, Idx),
66 extractElement(In2, Idx),
73 return hasAddOverflow(cast<ConstantInt>(Result),
74 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
78 static bool hasSubOverflow(ConstantInt *Result,
79 ConstantInt *In1, ConstantInt *In2,
82 return Result->getValue().ugt(In1->getValue());
84 if (In2->isNegative())
85 return Result->getValue().slt(In1->getValue());
87 return Result->getValue().sgt(In1->getValue());
90 /// Compute Result = In1-In2, returning true if the result overflowed for this
92 static bool subWithOverflow(Constant *&Result, Constant *In1,
93 Constant *In2, bool IsSigned = false) {
94 Result = ConstantExpr::getSub(In1, In2);
96 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
97 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
98 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
99 if (hasSubOverflow(extractElement(Result, Idx),
100 extractElement(In1, Idx),
101 extractElement(In2, Idx),
108 return hasSubOverflow(cast<ConstantInt>(Result),
109 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
113 /// Given an icmp instruction, return true if any use of this comparison is a
114 /// branch on sign bit comparison.
115 static bool hasBranchUse(ICmpInst &I) {
116 for (auto *U : I.users())
117 if (isa<BranchInst>(U))
122 /// Given an exploded icmp instruction, return true if the comparison only
123 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
124 /// result of the comparison is true when the input value is signed.
125 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
126 bool &TrueIfSigned) {
128 case ICmpInst::ICMP_SLT: // True if LHS s< 0
130 return RHS.isNullValue();
131 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
133 return RHS.isAllOnesValue();
134 case ICmpInst::ICMP_SGT: // True if LHS s> -1
135 TrueIfSigned = false;
136 return RHS.isAllOnesValue();
137 case ICmpInst::ICMP_UGT:
138 // True if LHS u> RHS and RHS == high-bit-mask - 1
140 return RHS.isMaxSignedValue();
141 case ICmpInst::ICMP_UGE:
142 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
144 return RHS.isSignMask();
150 /// Returns true if the exploded icmp can be expressed as a signed comparison
151 /// to zero and updates the predicate accordingly.
152 /// The signedness of the comparison is preserved.
153 /// TODO: Refactor with decomposeBitTestICmp()?
154 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
155 if (!ICmpInst::isSigned(Pred))
159 return ICmpInst::isRelational(Pred);
161 if (C.isOneValue()) {
162 if (Pred == ICmpInst::ICMP_SLT) {
163 Pred = ICmpInst::ICMP_SLE;
166 } else if (C.isAllOnesValue()) {
167 if (Pred == ICmpInst::ICMP_SGT) {
168 Pred = ICmpInst::ICMP_SGE;
176 /// Given a signed integer type and a set of known zero and one bits, compute
177 /// the maximum and minimum values that could have the specified known zero and
178 /// known one bits, returning them in Min/Max.
179 /// TODO: Move to method on KnownBits struct?
180 static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known,
181 APInt &Min, APInt &Max) {
182 assert(Known.getBitWidth() == Min.getBitWidth() &&
183 Known.getBitWidth() == Max.getBitWidth() &&
184 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
185 APInt UnknownBits = ~(Known.Zero|Known.One);
187 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
188 // bit if it is unknown.
190 Max = Known.One|UnknownBits;
192 if (UnknownBits.isNegative()) { // Sign bit is unknown
198 /// Given an unsigned integer type and a set of known zero and one bits, compute
199 /// the maximum and minimum values that could have the specified known zero and
200 /// known one bits, returning them in Min/Max.
201 /// TODO: Move to method on KnownBits struct?
202 static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known,
203 APInt &Min, APInt &Max) {
204 assert(Known.getBitWidth() == Min.getBitWidth() &&
205 Known.getBitWidth() == Max.getBitWidth() &&
206 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
207 APInt UnknownBits = ~(Known.Zero|Known.One);
209 // The minimum value is when the unknown bits are all zeros.
211 // The maximum value is when the unknown bits are all ones.
212 Max = Known.One|UnknownBits;
215 /// This is called when we see this pattern:
216 /// cmp pred (load (gep GV, ...)), cmpcst
217 /// where GV is a global variable with a constant initializer. Try to simplify
218 /// this into some simple computation that does not need the load. For example
219 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
221 /// If AndCst is non-null, then the loaded value is masked with that constant
222 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
223 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
226 ConstantInt *AndCst) {
227 Constant *Init = GV->getInitializer();
228 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
231 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
232 // Don't blow up on huge arrays.
233 if (ArrayElementCount > MaxArraySizeForCombine)
236 // There are many forms of this optimization we can handle, for now, just do
237 // the simple index into a single-dimensional array.
239 // Require: GEP GV, 0, i {{, constant indices}}
240 if (GEP->getNumOperands() < 3 ||
241 !isa<ConstantInt>(GEP->getOperand(1)) ||
242 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
243 isa<Constant>(GEP->getOperand(2)))
246 // Check that indices after the variable are constants and in-range for the
247 // type they index. Collect the indices. This is typically for arrays of
249 SmallVector<unsigned, 4> LaterIndices;
251 Type *EltTy = Init->getType()->getArrayElementType();
252 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
253 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
254 if (!Idx) return nullptr; // Variable index.
256 uint64_t IdxVal = Idx->getZExtValue();
257 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
259 if (StructType *STy = dyn_cast<StructType>(EltTy))
260 EltTy = STy->getElementType(IdxVal);
261 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
262 if (IdxVal >= ATy->getNumElements()) return nullptr;
263 EltTy = ATy->getElementType();
265 return nullptr; // Unknown type.
268 LaterIndices.push_back(IdxVal);
271 enum { Overdefined = -3, Undefined = -2 };
273 // Variables for our state machines.
275 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
276 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
277 // and 87 is the second (and last) index. FirstTrueElement is -2 when
278 // undefined, otherwise set to the first true element. SecondTrueElement is
279 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
280 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
282 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
283 // form "i != 47 & i != 87". Same state transitions as for true elements.
284 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
286 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
287 /// define a state machine that triggers for ranges of values that the index
288 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
289 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
290 /// index in the range (inclusive). We use -2 for undefined here because we
291 /// use relative comparisons and don't want 0-1 to match -1.
292 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
294 // MagicBitvector - This is a magic bitvector where we set a bit if the
295 // comparison is true for element 'i'. If there are 64 elements or less in
296 // the array, this will fully represent all the comparison results.
297 uint64_t MagicBitvector = 0;
299 // Scan the array and see if one of our patterns matches.
300 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
301 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
302 Constant *Elt = Init->getAggregateElement(i);
303 if (!Elt) return nullptr;
305 // If this is indexing an array of structures, get the structure element.
306 if (!LaterIndices.empty())
307 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
309 // If the element is masked, handle it.
310 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
312 // Find out if the comparison would be true or false for the i'th element.
313 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
314 CompareRHS, DL, &TLI);
315 // If the result is undef for this element, ignore it.
316 if (isa<UndefValue>(C)) {
317 // Extend range state machines to cover this element in case there is an
318 // undef in the middle of the range.
319 if (TrueRangeEnd == (int)i-1)
321 if (FalseRangeEnd == (int)i-1)
326 // If we can't compute the result for any of the elements, we have to give
327 // up evaluating the entire conditional.
328 if (!isa<ConstantInt>(C)) return nullptr;
330 // Otherwise, we know if the comparison is true or false for this element,
331 // update our state machines.
332 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
334 // State machine for single/double/range index comparison.
336 // Update the TrueElement state machine.
337 if (FirstTrueElement == Undefined)
338 FirstTrueElement = TrueRangeEnd = i; // First true element.
340 // Update double-compare state machine.
341 if (SecondTrueElement == Undefined)
342 SecondTrueElement = i;
344 SecondTrueElement = Overdefined;
346 // Update range state machine.
347 if (TrueRangeEnd == (int)i-1)
350 TrueRangeEnd = Overdefined;
353 // Update the FalseElement state machine.
354 if (FirstFalseElement == Undefined)
355 FirstFalseElement = FalseRangeEnd = i; // First false element.
357 // Update double-compare state machine.
358 if (SecondFalseElement == Undefined)
359 SecondFalseElement = i;
361 SecondFalseElement = Overdefined;
363 // Update range state machine.
364 if (FalseRangeEnd == (int)i-1)
367 FalseRangeEnd = Overdefined;
371 // If this element is in range, update our magic bitvector.
372 if (i < 64 && IsTrueForElt)
373 MagicBitvector |= 1ULL << i;
375 // If all of our states become overdefined, bail out early. Since the
376 // predicate is expensive, only check it every 8 elements. This is only
377 // really useful for really huge arrays.
378 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
379 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
380 FalseRangeEnd == Overdefined)
384 // Now that we've scanned the entire array, emit our new comparison(s). We
385 // order the state machines in complexity of the generated code.
386 Value *Idx = GEP->getOperand(2);
388 // If the index is larger than the pointer size of the target, truncate the
389 // index down like the GEP would do implicitly. We don't have to do this for
390 // an inbounds GEP because the index can't be out of range.
391 if (!GEP->isInBounds()) {
392 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
393 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
394 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
395 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
398 // If the comparison is only true for one or two elements, emit direct
400 if (SecondTrueElement != Overdefined) {
401 // None true -> false.
402 if (FirstTrueElement == Undefined)
403 return replaceInstUsesWith(ICI, Builder->getFalse());
405 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
407 // True for one element -> 'i == 47'.
408 if (SecondTrueElement == Undefined)
409 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
411 // True for two elements -> 'i == 47 | i == 72'.
412 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
413 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
414 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
415 return BinaryOperator::CreateOr(C1, C2);
418 // If the comparison is only false for one or two elements, emit direct
420 if (SecondFalseElement != Overdefined) {
421 // None false -> true.
422 if (FirstFalseElement == Undefined)
423 return replaceInstUsesWith(ICI, Builder->getTrue());
425 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
427 // False for one element -> 'i != 47'.
428 if (SecondFalseElement == Undefined)
429 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
431 // False for two elements -> 'i != 47 & i != 72'.
432 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
433 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
434 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
435 return BinaryOperator::CreateAnd(C1, C2);
438 // If the comparison can be replaced with a range comparison for the elements
439 // where it is true, emit the range check.
440 if (TrueRangeEnd != Overdefined) {
441 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
443 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
444 if (FirstTrueElement) {
445 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
446 Idx = Builder->CreateAdd(Idx, Offs);
449 Value *End = ConstantInt::get(Idx->getType(),
450 TrueRangeEnd-FirstTrueElement+1);
451 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
454 // False range check.
455 if (FalseRangeEnd != Overdefined) {
456 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
457 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
458 if (FirstFalseElement) {
459 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
460 Idx = Builder->CreateAdd(Idx, Offs);
463 Value *End = ConstantInt::get(Idx->getType(),
464 FalseRangeEnd-FirstFalseElement);
465 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
468 // If a magic bitvector captures the entire comparison state
469 // of this load, replace it with computation that does:
470 // ((magic_cst >> i) & 1) != 0
474 // Look for an appropriate type:
475 // - The type of Idx if the magic fits
476 // - The smallest fitting legal type if we have a DataLayout
478 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
481 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
484 Value *V = Builder->CreateIntCast(Idx, Ty, false);
485 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
486 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
487 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
494 /// Return a value that can be used to compare the *offset* implied by a GEP to
495 /// zero. For example, if we have &A[i], we want to return 'i' for
496 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
497 /// are involved. The above expression would also be legal to codegen as
498 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
499 /// This latter form is less amenable to optimization though, and we are allowed
500 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
502 /// If we can't emit an optimized form for this expression, this returns null.
504 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC,
505 const DataLayout &DL) {
506 gep_type_iterator GTI = gep_type_begin(GEP);
508 // Check to see if this gep only has a single variable index. If so, and if
509 // any constant indices are a multiple of its scale, then we can compute this
510 // in terms of the scale of the variable index. For example, if the GEP
511 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
512 // because the expression will cross zero at the same point.
513 unsigned i, e = GEP->getNumOperands();
515 for (i = 1; i != e; ++i, ++GTI) {
516 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
517 // Compute the aggregate offset of constant indices.
518 if (CI->isZero()) continue;
520 // Handle a struct index, which adds its field offset to the pointer.
521 if (StructType *STy = GTI.getStructTypeOrNull()) {
522 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
524 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
525 Offset += Size*CI->getSExtValue();
528 // Found our variable index.
533 // If there are no variable indices, we must have a constant offset, just
534 // evaluate it the general way.
535 if (i == e) return nullptr;
537 Value *VariableIdx = GEP->getOperand(i);
538 // Determine the scale factor of the variable element. For example, this is
539 // 4 if the variable index is into an array of i32.
540 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
542 // Verify that there are no other variable indices. If so, emit the hard way.
543 for (++i, ++GTI; i != e; ++i, ++GTI) {
544 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
545 if (!CI) return nullptr;
547 // Compute the aggregate offset of constant indices.
548 if (CI->isZero()) continue;
550 // Handle a struct index, which adds its field offset to the pointer.
551 if (StructType *STy = GTI.getStructTypeOrNull()) {
552 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
554 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
555 Offset += Size*CI->getSExtValue();
559 // Okay, we know we have a single variable index, which must be a
560 // pointer/array/vector index. If there is no offset, life is simple, return
562 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
563 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
565 // Cast to intptrty in case a truncation occurs. If an extension is needed,
566 // we don't need to bother extending: the extension won't affect where the
567 // computation crosses zero.
568 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
569 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
574 // Otherwise, there is an index. The computation we will do will be modulo
575 // the pointer size, so get it.
576 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
578 Offset &= PtrSizeMask;
579 VariableScale &= PtrSizeMask;
581 // To do this transformation, any constant index must be a multiple of the
582 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
583 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
584 // multiple of the variable scale.
585 int64_t NewOffs = Offset / (int64_t)VariableScale;
586 if (Offset != NewOffs*(int64_t)VariableScale)
589 // Okay, we can do this evaluation. Start by converting the index to intptr.
590 if (VariableIdx->getType() != IntPtrTy)
591 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
593 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
594 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
597 /// Returns true if we can rewrite Start as a GEP with pointer Base
598 /// and some integer offset. The nodes that need to be re-written
599 /// for this transformation will be added to Explored.
600 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
601 const DataLayout &DL,
602 SetVector<Value *> &Explored) {
603 SmallVector<Value *, 16> WorkList(1, Start);
604 Explored.insert(Base);
606 // The following traversal gives us an order which can be used
607 // when doing the final transformation. Since in the final
608 // transformation we create the PHI replacement instructions first,
609 // we don't have to get them in any particular order.
611 // However, for other instructions we will have to traverse the
612 // operands of an instruction first, which means that we have to
613 // do a post-order traversal.
614 while (!WorkList.empty()) {
615 SetVector<PHINode *> PHIs;
617 while (!WorkList.empty()) {
618 if (Explored.size() >= 100)
621 Value *V = WorkList.back();
623 if (Explored.count(V) != 0) {
628 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
629 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
630 // We've found some value that we can't explore which is different from
631 // the base. Therefore we can't do this transformation.
634 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
635 auto *CI = dyn_cast<CastInst>(V);
636 if (!CI->isNoopCast(DL))
639 if (Explored.count(CI->getOperand(0)) == 0)
640 WorkList.push_back(CI->getOperand(0));
643 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
644 // We're limiting the GEP to having one index. This will preserve
645 // the original pointer type. We could handle more cases in the
647 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
648 GEP->getType() != Start->getType())
651 if (Explored.count(GEP->getOperand(0)) == 0)
652 WorkList.push_back(GEP->getOperand(0));
655 if (WorkList.back() == V) {
657 // We've finished visiting this node, mark it as such.
661 if (auto *PN = dyn_cast<PHINode>(V)) {
662 // We cannot transform PHIs on unsplittable basic blocks.
663 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
670 // Explore the PHI nodes further.
671 for (auto *PN : PHIs)
672 for (Value *Op : PN->incoming_values())
673 if (Explored.count(Op) == 0)
674 WorkList.push_back(Op);
677 // Make sure that we can do this. Since we can't insert GEPs in a basic
678 // block before a PHI node, we can't easily do this transformation if
679 // we have PHI node users of transformed instructions.
680 for (Value *Val : Explored) {
681 for (Value *Use : Val->uses()) {
683 auto *PHI = dyn_cast<PHINode>(Use);
684 auto *Inst = dyn_cast<Instruction>(Val);
686 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
687 Explored.count(PHI) == 0)
690 if (PHI->getParent() == Inst->getParent())
697 // Sets the appropriate insert point on Builder where we can add
698 // a replacement Instruction for V (if that is possible).
699 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
700 bool Before = true) {
701 if (auto *PHI = dyn_cast<PHINode>(V)) {
702 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
705 if (auto *I = dyn_cast<Instruction>(V)) {
707 I = &*std::next(I->getIterator());
708 Builder.SetInsertPoint(I);
711 if (auto *A = dyn_cast<Argument>(V)) {
712 // Set the insertion point in the entry block.
713 BasicBlock &Entry = A->getParent()->getEntryBlock();
714 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
717 // Otherwise, this is a constant and we don't need to set a new
719 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
722 /// Returns a re-written value of Start as an indexed GEP using Base as a
724 static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
725 const DataLayout &DL,
726 SetVector<Value *> &Explored) {
727 // Perform all the substitutions. This is a bit tricky because we can
728 // have cycles in our use-def chains.
729 // 1. Create the PHI nodes without any incoming values.
730 // 2. Create all the other values.
731 // 3. Add the edges for the PHI nodes.
732 // 4. Emit GEPs to get the original pointers.
733 // 5. Remove the original instructions.
734 Type *IndexType = IntegerType::get(
735 Base->getContext(), DL.getPointerTypeSizeInBits(Start->getType()));
737 DenseMap<Value *, Value *> NewInsts;
738 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
740 // Create the new PHI nodes, without adding any incoming values.
741 for (Value *Val : Explored) {
744 // Create empty phi nodes. This avoids cyclic dependencies when creating
745 // the remaining instructions.
746 if (auto *PHI = dyn_cast<PHINode>(Val))
747 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
748 PHI->getName() + ".idx", PHI);
750 IRBuilder<> Builder(Base->getContext());
752 // Create all the other instructions.
753 for (Value *Val : Explored) {
755 if (NewInsts.find(Val) != NewInsts.end())
758 if (auto *CI = dyn_cast<CastInst>(Val)) {
759 NewInsts[CI] = NewInsts[CI->getOperand(0)];
762 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
763 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
764 : GEP->getOperand(1);
765 setInsertionPoint(Builder, GEP);
766 // Indices might need to be sign extended. GEPs will magically do
767 // this, but we need to do it ourselves here.
768 if (Index->getType()->getScalarSizeInBits() !=
769 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
770 Index = Builder.CreateSExtOrTrunc(
771 Index, NewInsts[GEP->getOperand(0)]->getType(),
772 GEP->getOperand(0)->getName() + ".sext");
775 auto *Op = NewInsts[GEP->getOperand(0)];
776 if (isa<ConstantInt>(Op) && dyn_cast<ConstantInt>(Op)->isZero())
777 NewInsts[GEP] = Index;
779 NewInsts[GEP] = Builder.CreateNSWAdd(
780 Op, Index, GEP->getOperand(0)->getName() + ".add");
783 if (isa<PHINode>(Val))
786 llvm_unreachable("Unexpected instruction type");
789 // Add the incoming values to the PHI nodes.
790 for (Value *Val : Explored) {
793 // All the instructions have been created, we can now add edges to the
795 if (auto *PHI = dyn_cast<PHINode>(Val)) {
796 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
797 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
798 Value *NewIncoming = PHI->getIncomingValue(I);
800 if (NewInsts.find(NewIncoming) != NewInsts.end())
801 NewIncoming = NewInsts[NewIncoming];
803 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
808 for (Value *Val : Explored) {
812 // Depending on the type, for external users we have to emit
813 // a GEP or a GEP + ptrtoint.
814 setInsertionPoint(Builder, Val, false);
816 // If required, create an inttoptr instruction for Base.
817 Value *NewBase = Base;
818 if (!Base->getType()->isPointerTy())
819 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
820 Start->getName() + "to.ptr");
822 Value *GEP = Builder.CreateInBoundsGEP(
823 Start->getType()->getPointerElementType(), NewBase,
824 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
826 if (!Val->getType()->isPointerTy()) {
827 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
828 Val->getName() + ".conv");
831 Val->replaceAllUsesWith(GEP);
834 return NewInsts[Start];
837 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
838 /// the input Value as a constant indexed GEP. Returns a pair containing
839 /// the GEPs Pointer and Index.
840 static std::pair<Value *, Value *>
841 getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
842 Type *IndexType = IntegerType::get(V->getContext(),
843 DL.getPointerTypeSizeInBits(V->getType()));
845 Constant *Index = ConstantInt::getNullValue(IndexType);
847 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
848 // We accept only inbouds GEPs here to exclude the possibility of
850 if (!GEP->isInBounds())
852 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
853 GEP->getType() == V->getType()) {
854 V = GEP->getOperand(0);
855 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
856 Index = ConstantExpr::getAdd(
857 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
862 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
863 if (!CI->isNoopCast(DL))
865 V = CI->getOperand(0);
868 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
869 if (!CI->isNoopCast(DL))
871 V = CI->getOperand(0);
879 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
880 /// We can look through PHIs, GEPs and casts in order to determine a common base
881 /// between GEPLHS and RHS.
882 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
883 ICmpInst::Predicate Cond,
884 const DataLayout &DL) {
885 if (!GEPLHS->hasAllConstantIndices())
888 // Make sure the pointers have the same type.
889 if (GEPLHS->getType() != RHS->getType())
892 Value *PtrBase, *Index;
893 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
895 // The set of nodes that will take part in this transformation.
896 SetVector<Value *> Nodes;
898 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
901 // We know we can re-write this as
902 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
903 // Since we've only looked through inbouds GEPs we know that we
904 // can't have overflow on either side. We can therefore re-write
906 // OFFSET1 cmp OFFSET2
907 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
909 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
910 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
911 // offset. Since Index is the offset of LHS to the base pointer, we will now
912 // compare the offsets instead of comparing the pointers.
913 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
916 /// Fold comparisons between a GEP instruction and something else. At this point
917 /// we know that the GEP is on the LHS of the comparison.
918 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
919 ICmpInst::Predicate Cond,
921 // Don't transform signed compares of GEPs into index compares. Even if the
922 // GEP is inbounds, the final add of the base pointer can have signed overflow
923 // and would change the result of the icmp.
924 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
925 // the maximum signed value for the pointer type.
926 if (ICmpInst::isSigned(Cond))
929 // Look through bitcasts and addrspacecasts. We do not however want to remove
931 if (!isa<GetElementPtrInst>(RHS))
932 RHS = RHS->stripPointerCasts();
934 Value *PtrBase = GEPLHS->getOperand(0);
935 if (PtrBase == RHS && GEPLHS->isInBounds()) {
936 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
937 // This transformation (ignoring the base and scales) is valid because we
938 // know pointers can't overflow since the gep is inbounds. See if we can
939 // output an optimized form.
940 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
942 // If not, synthesize the offset the hard way.
944 Offset = EmitGEPOffset(GEPLHS);
945 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
946 Constant::getNullValue(Offset->getType()));
947 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
948 // If the base pointers are different, but the indices are the same, just
949 // compare the base pointer.
950 if (PtrBase != GEPRHS->getOperand(0)) {
951 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
952 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
953 GEPRHS->getOperand(0)->getType();
955 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
956 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
957 IndicesTheSame = false;
961 // If all indices are the same, just compare the base pointers.
963 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
965 // If we're comparing GEPs with two base pointers that only differ in type
966 // and both GEPs have only constant indices or just one use, then fold
967 // the compare with the adjusted indices.
968 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
969 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
970 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
971 PtrBase->stripPointerCasts() ==
972 GEPRHS->getOperand(0)->stripPointerCasts()) {
973 Value *LOffset = EmitGEPOffset(GEPLHS);
974 Value *ROffset = EmitGEPOffset(GEPRHS);
976 // If we looked through an addrspacecast between different sized address
977 // spaces, the LHS and RHS pointers are different sized
978 // integers. Truncate to the smaller one.
979 Type *LHSIndexTy = LOffset->getType();
980 Type *RHSIndexTy = ROffset->getType();
981 if (LHSIndexTy != RHSIndexTy) {
982 if (LHSIndexTy->getPrimitiveSizeInBits() <
983 RHSIndexTy->getPrimitiveSizeInBits()) {
984 ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy);
986 LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy);
989 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
991 return replaceInstUsesWith(I, Cmp);
994 // Otherwise, the base pointers are different and the indices are
995 // different. Try convert this to an indexed compare by looking through
997 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1000 // If one of the GEPs has all zero indices, recurse.
1001 if (GEPLHS->hasAllZeroIndices())
1002 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
1003 ICmpInst::getSwappedPredicate(Cond), I);
1005 // If the other GEP has all zero indices, recurse.
1006 if (GEPRHS->hasAllZeroIndices())
1007 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
1009 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
1010 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
1011 // If the GEPs only differ by one index, compare it.
1012 unsigned NumDifferences = 0; // Keep track of # differences.
1013 unsigned DiffOperand = 0; // The operand that differs.
1014 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
1015 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
1016 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
1017 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
1018 // Irreconcilable differences.
1022 if (NumDifferences++) break;
1027 if (NumDifferences == 0) // SAME GEP?
1028 return replaceInstUsesWith(I, // No comparison is needed here.
1029 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
1031 else if (NumDifferences == 1 && GEPsInBounds) {
1032 Value *LHSV = GEPLHS->getOperand(DiffOperand);
1033 Value *RHSV = GEPRHS->getOperand(DiffOperand);
1034 // Make sure we do a signed comparison here.
1035 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1039 // Only lower this if the icmp is the only user of the GEP or if we expect
1040 // the result to fold to a constant!
1041 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1042 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1043 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1044 Value *L = EmitGEPOffset(GEPLHS);
1045 Value *R = EmitGEPOffset(GEPRHS);
1046 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1050 // Try convert this to an indexed compare by looking through PHIs/casts as a
1052 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1055 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1056 const AllocaInst *Alloca,
1057 const Value *Other) {
1058 assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1060 // It would be tempting to fold away comparisons between allocas and any
1061 // pointer not based on that alloca (e.g. an argument). However, even
1062 // though such pointers cannot alias, they can still compare equal.
1064 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1065 // doesn't escape we can argue that it's impossible to guess its value, and we
1066 // can therefore act as if any such guesses are wrong.
1068 // The code below checks that the alloca doesn't escape, and that it's only
1069 // used in a comparison once (the current instruction). The
1070 // single-comparison-use condition ensures that we're trivially folding all
1071 // comparisons against the alloca consistently, and avoids the risk of
1072 // erroneously folding a comparison of the pointer with itself.
1074 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1076 SmallVector<const Use *, 32> Worklist;
1077 for (const Use &U : Alloca->uses()) {
1078 if (Worklist.size() >= MaxIter)
1080 Worklist.push_back(&U);
1083 unsigned NumCmps = 0;
1084 while (!Worklist.empty()) {
1085 assert(Worklist.size() <= MaxIter);
1086 const Use *U = Worklist.pop_back_val();
1087 const Value *V = U->getUser();
1090 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1091 isa<SelectInst>(V)) {
1093 } else if (isa<LoadInst>(V)) {
1094 // Loading from the pointer doesn't escape it.
1096 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1097 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1098 if (SI->getValueOperand() == U->get())
1101 } else if (isa<ICmpInst>(V)) {
1103 return nullptr; // Found more than one cmp.
1105 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1106 switch (Intrin->getIntrinsicID()) {
1107 // These intrinsics don't escape or compare the pointer. Memset is safe
1108 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1109 // we don't allow stores, so src cannot point to V.
1110 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1111 case Intrinsic::dbg_declare: case Intrinsic::dbg_value:
1112 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1120 for (const Use &U : V->uses()) {
1121 if (Worklist.size() >= MaxIter)
1123 Worklist.push_back(&U);
1127 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1128 return replaceInstUsesWith(
1130 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1133 /// Fold "icmp pred (X+CI), X".
1134 Instruction *InstCombiner::foldICmpAddOpConst(Instruction &ICI,
1135 Value *X, ConstantInt *CI,
1136 ICmpInst::Predicate Pred) {
1137 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1138 // so the values can never be equal. Similarly for all other "or equals"
1141 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1142 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1143 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1144 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1146 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
1147 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1150 // (X+1) >u X --> X <u (0-1) --> X != 255
1151 // (X+2) >u X --> X <u (0-2) --> X <u 254
1152 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1153 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1154 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
1156 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
1157 ConstantInt *SMax = ConstantInt::get(X->getContext(),
1158 APInt::getSignedMaxValue(BitWidth));
1160 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1161 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1162 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1163 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1164 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1165 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1166 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1167 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
1169 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1170 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1171 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1172 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1173 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1174 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1176 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1177 Constant *C = Builder->getInt(CI->getValue()-1);
1178 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
1181 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1182 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1183 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1184 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1187 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1189 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1190 if (I.getPredicate() == I.ICMP_NE)
1191 Pred = CmpInst::getInversePredicate(Pred);
1192 return new ICmpInst(Pred, LHS, RHS);
1195 // Don't bother doing any work for cases which InstSimplify handles.
1196 if (AP2.isNullValue())
1199 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1201 if (AP2.isAllOnesValue())
1203 if (AP2.isNegative() != AP1.isNegative())
1210 // 'A' must be large enough to shift out the highest set bit.
1211 return getICmp(I.ICMP_UGT, A,
1212 ConstantInt::get(A->getType(), AP2.logBase2()));
1215 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1218 if (IsAShr && AP1.isNegative())
1219 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1221 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1224 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1225 // There are multiple solutions if we are comparing against -1 and the LHS
1226 // of the ashr is not a power of two.
1227 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1228 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1229 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1230 } else if (AP1 == AP2.lshr(Shift)) {
1231 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1235 // Shifting const2 will never be equal to const1.
1236 // FIXME: This should always be handled by InstSimplify?
1237 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1238 return replaceInstUsesWith(I, TorF);
1241 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1242 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1243 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1246 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1248 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1249 if (I.getPredicate() == I.ICMP_NE)
1250 Pred = CmpInst::getInversePredicate(Pred);
1251 return new ICmpInst(Pred, LHS, RHS);
1254 // Don't bother doing any work for cases which InstSimplify handles.
1255 if (AP2.isNullValue())
1258 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1260 if (!AP1 && AP2TrailingZeros != 0)
1263 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1266 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1268 // Get the distance between the lowest bits that are set.
1269 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1271 if (Shift > 0 && AP2.shl(Shift) == AP1)
1272 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1274 // Shifting const2 will never be equal to const1.
1275 // FIXME: This should always be handled by InstSimplify?
1276 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1277 return replaceInstUsesWith(I, TorF);
1280 /// The caller has matched a pattern of the form:
1281 /// I = icmp ugt (add (add A, B), CI2), CI1
1282 /// If this is of the form:
1284 /// if (sum+128 >u 255)
1285 /// Then replace it with llvm.sadd.with.overflow.i8.
1287 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1288 ConstantInt *CI2, ConstantInt *CI1,
1290 // The transformation we're trying to do here is to transform this into an
1291 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1292 // with a narrower add, and discard the add-with-constant that is part of the
1293 // range check (if we can't eliminate it, this isn't profitable).
1295 // In order to eliminate the add-with-constant, the compare can be its only
1297 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1298 if (!AddWithCst->hasOneUse())
1301 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1302 if (!CI2->getValue().isPowerOf2())
1304 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1305 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1308 // The width of the new add formed is 1 more than the bias.
1311 // Check to see that CI1 is an all-ones value with NewWidth bits.
1312 if (CI1->getBitWidth() == NewWidth ||
1313 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1316 // This is only really a signed overflow check if the inputs have been
1317 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1318 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1319 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1320 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1321 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1324 // In order to replace the original add with a narrower
1325 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1326 // and truncates that discard the high bits of the add. Verify that this is
1328 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1329 for (User *U : OrigAdd->users()) {
1330 if (U == AddWithCst)
1333 // Only accept truncates for now. We would really like a nice recursive
1334 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1335 // chain to see which bits of a value are actually demanded. If the
1336 // original add had another add which was then immediately truncated, we
1337 // could still do the transformation.
1338 TruncInst *TI = dyn_cast<TruncInst>(U);
1339 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1343 // If the pattern matches, truncate the inputs to the narrower type and
1344 // use the sadd_with_overflow intrinsic to efficiently compute both the
1345 // result and the overflow bit.
1346 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1347 Value *F = Intrinsic::getDeclaration(I.getModule(),
1348 Intrinsic::sadd_with_overflow, NewType);
1350 InstCombiner::BuilderTy *Builder = IC.Builder;
1352 // Put the new code above the original add, in case there are any uses of the
1353 // add between the add and the compare.
1354 Builder->SetInsertPoint(OrigAdd);
1356 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName() + ".trunc");
1357 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName() + ".trunc");
1358 CallInst *Call = Builder->CreateCall(F, {TruncA, TruncB}, "sadd");
1359 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1360 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1362 // The inner add was the result of the narrow add, zero extended to the
1363 // wider type. Replace it with the result computed by the intrinsic.
1364 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1366 // The original icmp gets replaced with the overflow value.
1367 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1370 // Fold icmp Pred X, C.
1371 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1372 CmpInst::Predicate Pred = Cmp.getPredicate();
1373 Value *X = Cmp.getOperand(0);
1376 if (!match(Cmp.getOperand(1), m_APInt(C)))
1379 Value *A = nullptr, *B = nullptr;
1381 // Match the following pattern, which is a common idiom when writing
1382 // overflow-safe integer arithmetic functions. The source performs an addition
1383 // in wider type and explicitly checks for overflow using comparisons against
1384 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1386 // TODO: This could probably be generalized to handle other overflow-safe
1387 // operations if we worked out the formulas to compute the appropriate magic
1391 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1393 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1394 if (Pred == ICmpInst::ICMP_UGT &&
1395 match(X, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1396 if (Instruction *Res = processUGT_ADDCST_ADD(
1397 Cmp, A, B, CI2, cast<ConstantInt>(Cmp.getOperand(1)), *this))
1401 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1402 if (C->isNullValue() && Pred == ICmpInst::ICMP_SGT) {
1403 SelectPatternResult SPR = matchSelectPattern(X, A, B);
1404 if (SPR.Flavor == SPF_SMIN) {
1405 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1406 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1407 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1408 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1412 // FIXME: Use m_APInt to allow folds for splat constants.
1413 ConstantInt *CI = dyn_cast<ConstantInt>(Cmp.getOperand(1));
1417 // Canonicalize icmp instructions based on dominating conditions.
1418 BasicBlock *Parent = Cmp.getParent();
1419 BasicBlock *Dom = Parent->getSinglePredecessor();
1420 auto *BI = Dom ? dyn_cast<BranchInst>(Dom->getTerminator()) : nullptr;
1421 ICmpInst::Predicate Pred2;
1422 BasicBlock *TrueBB, *FalseBB;
1424 if (BI && match(BI, m_Br(m_ICmp(Pred2, m_Specific(X), m_ConstantInt(CI2)),
1425 TrueBB, FalseBB)) &&
1426 TrueBB != FalseBB) {
1428 ConstantRange::makeAllowedICmpRegion(Pred, CI->getValue());
1429 ConstantRange DominatingCR =
1431 ? ConstantRange::makeExactICmpRegion(Pred2, CI2->getValue())
1432 : ConstantRange::makeExactICmpRegion(
1433 CmpInst::getInversePredicate(Pred2), CI2->getValue());
1434 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1435 ConstantRange Difference = DominatingCR.difference(CR);
1436 if (Intersection.isEmptySet())
1437 return replaceInstUsesWith(Cmp, Builder->getFalse());
1438 if (Difference.isEmptySet())
1439 return replaceInstUsesWith(Cmp, Builder->getTrue());
1441 // If this is a normal comparison, it demands all bits. If it is a sign
1442 // bit comparison, it only demands the sign bit.
1444 bool IsSignBit = isSignBitCheck(Pred, CI->getValue(), UnusedBit);
1446 // Canonicalizing a sign bit comparison that gets used in a branch,
1447 // pessimizes codegen by generating branch on zero instruction instead
1448 // of a test and branch. So we avoid canonicalizing in such situations
1449 // because test and branch instruction has better branch displacement
1450 // than compare and branch instruction.
1451 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1454 if (auto *AI = Intersection.getSingleElement())
1455 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder->getInt(*AI));
1456 if (auto *AD = Difference.getSingleElement())
1457 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder->getInt(*AD));
1463 /// Fold icmp (trunc X, Y), C.
1464 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1467 ICmpInst::Predicate Pred = Cmp.getPredicate();
1468 Value *X = Trunc->getOperand(0);
1469 if (C->isOneValue() && C->getBitWidth() > 1) {
1470 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1472 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1473 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1474 ConstantInt::get(V->getType(), 1));
1477 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1478 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1479 // of the high bits truncated out of x are known.
1480 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1481 SrcBits = X->getType()->getScalarSizeInBits();
1482 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1484 // If all the high bits are known, we can do this xform.
1485 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1486 // Pull in the high bits from known-ones set.
1487 APInt NewRHS = C->zext(SrcBits);
1488 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1489 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1496 /// Fold icmp (xor X, Y), C.
1497 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1498 BinaryOperator *Xor,
1500 Value *X = Xor->getOperand(0);
1501 Value *Y = Xor->getOperand(1);
1503 if (!match(Y, m_APInt(XorC)))
1506 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1508 ICmpInst::Predicate Pred = Cmp.getPredicate();
1509 if ((Pred == ICmpInst::ICMP_SLT && C->isNullValue()) ||
1510 (Pred == ICmpInst::ICMP_SGT && C->isAllOnesValue())) {
1512 // If the sign bit of the XorCst is not set, there is no change to
1513 // the operation, just stop using the Xor.
1514 if (!XorC->isNegative()) {
1515 Cmp.setOperand(0, X);
1520 // Was the old condition true if the operand is positive?
1521 bool isTrueIfPositive = Pred == ICmpInst::ICMP_SGT;
1523 // If so, the new one isn't.
1524 isTrueIfPositive ^= true;
1526 Constant *CmpConstant = cast<Constant>(Cmp.getOperand(1));
1527 if (isTrueIfPositive)
1528 return new ICmpInst(ICmpInst::ICMP_SGT, X, SubOne(CmpConstant));
1530 return new ICmpInst(ICmpInst::ICMP_SLT, X, AddOne(CmpConstant));
1533 if (Xor->hasOneUse()) {
1534 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1535 if (!Cmp.isEquality() && XorC->isSignMask()) {
1536 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1537 : Cmp.getSignedPredicate();
1538 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), *C ^ *XorC));
1541 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1542 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1543 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1544 : Cmp.getSignedPredicate();
1545 Pred = Cmp.getSwappedPredicate(Pred);
1546 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), *C ^ *XorC));
1550 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1551 // iff -C is a power of 2
1552 if (Pred == ICmpInst::ICMP_UGT && *XorC == ~(*C) && (*C + 1).isPowerOf2())
1553 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1555 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1556 // iff -C is a power of 2
1557 if (Pred == ICmpInst::ICMP_ULT && *XorC == -(*C) && C->isPowerOf2())
1558 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
1563 /// Fold icmp (and (sh X, Y), C2), C1.
1564 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1565 const APInt *C1, const APInt *C2) {
1566 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1567 if (!Shift || !Shift->isShift())
1570 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1571 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1572 // code produced by the clang front-end, for bitfield access.
1573 // This seemingly simple opportunity to fold away a shift turns out to be
1574 // rather complicated. See PR17827 for details.
1575 unsigned ShiftOpcode = Shift->getOpcode();
1576 bool IsShl = ShiftOpcode == Instruction::Shl;
1578 if (match(Shift->getOperand(1), m_APInt(C3))) {
1579 bool CanFold = false;
1580 if (ShiftOpcode == Instruction::AShr) {
1581 // There may be some constraints that make this possible, but nothing
1582 // simple has been discovered yet.
1584 } else if (ShiftOpcode == Instruction::Shl) {
1585 // For a left shift, we can fold if the comparison is not signed. We can
1586 // also fold a signed comparison if the mask value and comparison value
1587 // are not negative. These constraints may not be obvious, but we can
1588 // prove that they are correct using an SMT solver.
1589 if (!Cmp.isSigned() || (!C2->isNegative() && !C1->isNegative()))
1591 } else if (ShiftOpcode == Instruction::LShr) {
1592 // For a logical right shift, we can fold if the comparison is not signed.
1593 // We can also fold a signed comparison if the shifted mask value and the
1594 // shifted comparison value are not negative. These constraints may not be
1595 // obvious, but we can prove that they are correct using an SMT solver.
1596 if (!Cmp.isSigned() ||
1597 (!C2->shl(*C3).isNegative() && !C1->shl(*C3).isNegative()))
1602 APInt NewCst = IsShl ? C1->lshr(*C3) : C1->shl(*C3);
1603 APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1604 // Check to see if we are shifting out any of the bits being compared.
1605 if (SameAsC1 != *C1) {
1606 // If we shifted bits out, the fold is not going to work out. As a
1607 // special case, check to see if this means that the result is always
1608 // true or false now.
1609 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1610 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1611 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1612 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1614 Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1615 APInt NewAndCst = IsShl ? C2->lshr(*C3) : C2->shl(*C3);
1616 And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1617 And->setOperand(0, Shift->getOperand(0));
1618 Worklist.Add(Shift); // Shift is dead.
1624 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1625 // preferable because it allows the C2 << Y expression to be hoisted out of a
1626 // loop if Y is invariant and X is not.
1627 if (Shift->hasOneUse() && C1->isNullValue() && Cmp.isEquality() &&
1628 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1631 IsShl ? Builder->CreateLShr(And->getOperand(1), Shift->getOperand(1))
1632 : Builder->CreateShl(And->getOperand(1), Shift->getOperand(1));
1634 // Compute X & (C2 << Y).
1635 Value *NewAnd = Builder->CreateAnd(Shift->getOperand(0), NewShift);
1636 Cmp.setOperand(0, NewAnd);
1643 /// Fold icmp (and X, C2), C1.
1644 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1645 BinaryOperator *And,
1648 if (!match(And->getOperand(1), m_APInt(C2)))
1651 if (!And->hasOneUse() || !And->getOperand(0)->hasOneUse())
1654 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1655 // the input width without changing the value produced, eliminate the cast:
1657 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1659 // We can do this transformation if the constants do not have their sign bits
1660 // set or if it is an equality comparison. Extending a relational comparison
1661 // when we're checking the sign bit would not work.
1663 if (match(And->getOperand(0), m_Trunc(m_Value(W))) &&
1664 (Cmp.isEquality() || (!C1->isNegative() && !C2->isNegative()))) {
1665 // TODO: Is this a good transform for vectors? Wider types may reduce
1666 // throughput. Should this transform be limited (even for scalars) by using
1667 // shouldChangeType()?
1668 if (!Cmp.getType()->isVectorTy()) {
1669 Type *WideType = W->getType();
1670 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1671 Constant *ZextC1 = ConstantInt::get(WideType, C1->zext(WideScalarBits));
1672 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1673 Value *NewAnd = Builder->CreateAnd(W, ZextC2, And->getName());
1674 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1678 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, C2))
1681 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1682 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1684 // iff pred isn't signed
1685 if (!Cmp.isSigned() && C1->isNullValue() &&
1686 match(And->getOperand(1), m_One())) {
1687 Constant *One = cast<Constant>(And->getOperand(1));
1688 Value *Or = And->getOperand(0);
1689 Value *A, *B, *LShr;
1690 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1691 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1692 unsigned UsesRemoved = 0;
1693 if (And->hasOneUse())
1695 if (Or->hasOneUse())
1697 if (LShr->hasOneUse())
1700 // Compute A & ((1 << B) | 1)
1701 Value *NewOr = nullptr;
1702 if (auto *C = dyn_cast<Constant>(B)) {
1703 if (UsesRemoved >= 1)
1704 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1706 if (UsesRemoved >= 3)
1707 NewOr = Builder->CreateOr(Builder->CreateShl(One, B, LShr->getName(),
1709 One, Or->getName());
1712 Value *NewAnd = Builder->CreateAnd(A, NewOr, And->getName());
1713 Cmp.setOperand(0, NewAnd);
1719 // (X & C2) > C1 --> (X & C2) != 0, if any bit set in (X & C2) will produce a
1720 // result greater than C1.
1721 unsigned NumTZ = C2->countTrailingZeros();
1722 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && NumTZ < C2->getBitWidth() &&
1723 APInt::getOneBitSet(C2->getBitWidth(), NumTZ).ugt(*C1)) {
1724 Constant *Zero = Constant::getNullValue(And->getType());
1725 return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
1731 /// Fold icmp (and X, Y), C.
1732 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1733 BinaryOperator *And,
1735 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1738 // TODO: These all require that Y is constant too, so refactor with the above.
1740 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1741 Value *X = And->getOperand(0);
1742 Value *Y = And->getOperand(1);
1743 if (auto *LI = dyn_cast<LoadInst>(X))
1744 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1745 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1746 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1747 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1748 ConstantInt *C2 = cast<ConstantInt>(Y);
1749 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1753 if (!Cmp.isEquality())
1756 // X & -C == -C -> X > u ~C
1757 // X & -C != -C -> X <= u ~C
1758 // iff C is a power of 2
1759 if (Cmp.getOperand(1) == Y && (-(*C)).isPowerOf2()) {
1760 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1761 : CmpInst::ICMP_ULE;
1762 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1765 // (X & C2) == 0 -> (trunc X) >= 0
1766 // (X & C2) != 0 -> (trunc X) < 0
1767 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1769 if (And->hasOneUse() && C->isNullValue() && match(Y, m_APInt(C2))) {
1770 int32_t ExactLogBase2 = C2->exactLogBase2();
1771 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1772 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1773 if (And->getType()->isVectorTy())
1774 NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1775 Value *Trunc = Builder->CreateTrunc(X, NTy);
1776 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1777 : CmpInst::ICMP_SLT;
1778 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1785 /// Fold icmp (or X, Y), C.
1786 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1788 ICmpInst::Predicate Pred = Cmp.getPredicate();
1789 if (C->isOneValue()) {
1790 // icmp slt signum(V) 1 --> icmp slt V, 1
1792 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1793 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1794 ConstantInt::get(V->getType(), 1));
1797 // X | C == C --> X <=u C
1798 // X | C != C --> X >u C
1799 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1800 if (Cmp.isEquality() && Cmp.getOperand(1) == Or->getOperand(1) &&
1801 (*C + 1).isPowerOf2()) {
1802 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1803 return new ICmpInst(Pred, Or->getOperand(0), Or->getOperand(1));
1806 if (!Cmp.isEquality() || !C->isNullValue() || !Or->hasOneUse())
1810 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1811 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1812 // -> and (icmp eq P, null), (icmp eq Q, null).
1814 Builder->CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1816 Builder->CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1817 auto LogicOpc = Pred == ICmpInst::Predicate::ICMP_EQ ? Instruction::And
1819 return BinaryOperator::Create(LogicOpc, CmpP, CmpQ);
1825 /// Fold icmp (mul X, Y), C.
1826 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1827 BinaryOperator *Mul,
1830 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1833 // If this is a test of the sign bit and the multiply is sign-preserving with
1834 // a constant operand, use the multiply LHS operand instead.
1835 ICmpInst::Predicate Pred = Cmp.getPredicate();
1836 if (isSignTest(Pred, *C) && Mul->hasNoSignedWrap()) {
1837 if (MulC->isNegative())
1838 Pred = ICmpInst::getSwappedPredicate(Pred);
1839 return new ICmpInst(Pred, Mul->getOperand(0),
1840 Constant::getNullValue(Mul->getType()));
1846 /// Fold icmp (shl 1, Y), C.
1847 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
1850 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1853 Type *ShiftType = Shl->getType();
1854 uint32_t TypeBits = C->getBitWidth();
1855 bool CIsPowerOf2 = C->isPowerOf2();
1856 ICmpInst::Predicate Pred = Cmp.getPredicate();
1857 if (Cmp.isUnsigned()) {
1858 // (1 << Y) pred C -> Y pred Log2(C)
1860 // (1 << Y) < 30 -> Y <= 4
1861 // (1 << Y) <= 30 -> Y <= 4
1862 // (1 << Y) >= 30 -> Y > 4
1863 // (1 << Y) > 30 -> Y > 4
1864 if (Pred == ICmpInst::ICMP_ULT)
1865 Pred = ICmpInst::ICMP_ULE;
1866 else if (Pred == ICmpInst::ICMP_UGE)
1867 Pred = ICmpInst::ICMP_UGT;
1870 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1871 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1872 unsigned CLog2 = C->logBase2();
1873 if (CLog2 == TypeBits - 1) {
1874 if (Pred == ICmpInst::ICMP_UGE)
1875 Pred = ICmpInst::ICMP_EQ;
1876 else if (Pred == ICmpInst::ICMP_ULT)
1877 Pred = ICmpInst::ICMP_NE;
1879 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1880 } else if (Cmp.isSigned()) {
1881 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1882 if (C->isAllOnesValue()) {
1883 // (1 << Y) <= -1 -> Y == 31
1884 if (Pred == ICmpInst::ICMP_SLE)
1885 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1887 // (1 << Y) > -1 -> Y != 31
1888 if (Pred == ICmpInst::ICMP_SGT)
1889 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1891 // (1 << Y) < 0 -> Y == 31
1892 // (1 << Y) <= 0 -> Y == 31
1893 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1894 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1896 // (1 << Y) >= 0 -> Y != 31
1897 // (1 << Y) > 0 -> Y != 31
1898 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1899 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1901 } else if (Cmp.isEquality() && CIsPowerOf2) {
1902 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C->logBase2()));
1908 /// Fold icmp (shl X, Y), C.
1909 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1910 BinaryOperator *Shl,
1912 const APInt *ShiftVal;
1913 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1914 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), *C, *ShiftVal);
1916 const APInt *ShiftAmt;
1917 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1918 return foldICmpShlOne(Cmp, Shl, C);
1920 // Check that the shift amount is in range. If not, don't perform undefined
1921 // shifts. When the shift is visited, it will be simplified.
1922 unsigned TypeBits = C->getBitWidth();
1923 if (ShiftAmt->uge(TypeBits))
1926 ICmpInst::Predicate Pred = Cmp.getPredicate();
1927 Value *X = Shl->getOperand(0);
1928 Type *ShType = Shl->getType();
1930 // NSW guarantees that we are only shifting out sign bits from the high bits,
1931 // so we can ASHR the compare constant without needing a mask and eliminate
1933 if (Shl->hasNoSignedWrap()) {
1934 if (Pred == ICmpInst::ICMP_SGT) {
1935 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
1936 APInt ShiftedC = C->ashr(*ShiftAmt);
1937 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1939 if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) {
1940 // This is the same code as the SGT case, but assert the pre-condition
1941 // that is needed for this to work with equality predicates.
1942 assert(C->ashr(*ShiftAmt).shl(*ShiftAmt) == *C &&
1943 "Compare known true or false was not folded");
1944 APInt ShiftedC = C->ashr(*ShiftAmt);
1945 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1947 if (Pred == ICmpInst::ICMP_SLT) {
1948 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
1949 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
1950 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
1951 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
1952 assert(!C->isMinSignedValue() && "Unexpected icmp slt");
1953 APInt ShiftedC = (*C - 1).ashr(*ShiftAmt) + 1;
1954 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1956 // If this is a signed comparison to 0 and the shift is sign preserving,
1957 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
1958 // do that if we're sure to not continue on in this function.
1959 if (isSignTest(Pred, *C))
1960 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
1963 // NUW guarantees that we are only shifting out zero bits from the high bits,
1964 // so we can LSHR the compare constant without needing a mask and eliminate
1966 if (Shl->hasNoUnsignedWrap()) {
1967 if (Pred == ICmpInst::ICMP_UGT) {
1968 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
1969 APInt ShiftedC = C->lshr(*ShiftAmt);
1970 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1972 if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) {
1973 // This is the same code as the UGT case, but assert the pre-condition
1974 // that is needed for this to work with equality predicates.
1975 assert(C->lshr(*ShiftAmt).shl(*ShiftAmt) == *C &&
1976 "Compare known true or false was not folded");
1977 APInt ShiftedC = C->lshr(*ShiftAmt);
1978 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1980 if (Pred == ICmpInst::ICMP_ULT) {
1981 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
1982 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
1983 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
1984 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
1985 assert(C->ugt(0) && "ult 0 should have been eliminated");
1986 APInt ShiftedC = (*C - 1).lshr(*ShiftAmt) + 1;
1987 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1991 if (Cmp.isEquality() && Shl->hasOneUse()) {
1992 // Strength-reduce the shift into an 'and'.
1993 Constant *Mask = ConstantInt::get(
1995 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
1996 Value *And = Builder->CreateAnd(X, Mask, Shl->getName() + ".mask");
1997 Constant *LShrC = ConstantInt::get(ShType, C->lshr(*ShiftAmt));
1998 return new ICmpInst(Pred, And, LShrC);
2001 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2002 bool TrueIfSigned = false;
2003 if (Shl->hasOneUse() && isSignBitCheck(Pred, *C, TrueIfSigned)) {
2004 // (X << 31) <s 0 --> (X & 1) != 0
2005 Constant *Mask = ConstantInt::get(
2007 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2008 Value *And = Builder->CreateAnd(X, Mask, Shl->getName() + ".mask");
2009 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2010 And, Constant::getNullValue(ShType));
2013 // Transform (icmp pred iM (shl iM %v, N), C)
2014 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2015 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2016 // This enables us to get rid of the shift in favor of a trunc that may be
2017 // free on the target. It has the additional benefit of comparing to a
2018 // smaller constant that may be more target-friendly.
2019 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2020 if (Shl->hasOneUse() && Amt != 0 && C->countTrailingZeros() >= Amt &&
2021 DL.isLegalInteger(TypeBits - Amt)) {
2022 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2023 if (ShType->isVectorTy())
2024 TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2026 ConstantInt::get(TruncTy, C->ashr(*ShiftAmt).trunc(TypeBits - Amt));
2027 return new ICmpInst(Pred, Builder->CreateTrunc(X, TruncTy), NewC);
2033 /// Fold icmp ({al}shr X, Y), C.
2034 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2035 BinaryOperator *Shr,
2037 // An exact shr only shifts out zero bits, so:
2038 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2039 Value *X = Shr->getOperand(0);
2040 CmpInst::Predicate Pred = Cmp.getPredicate();
2041 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2043 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2045 const APInt *ShiftVal;
2046 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2047 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), *C, *ShiftVal);
2049 const APInt *ShiftAmt;
2050 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2053 // Check that the shift amount is in range. If not, don't perform undefined
2054 // shifts. When the shift is visited it will be simplified.
2055 unsigned TypeBits = C->getBitWidth();
2056 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2057 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2060 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2061 if (!Cmp.isEquality()) {
2062 // If we have an unsigned comparison and an ashr, we can't simplify this.
2063 // Similarly for signed comparisons with lshr.
2064 if (Cmp.isSigned() != IsAShr)
2067 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
2068 // by a power of 2. Since we already have logic to simplify these,
2069 // transform to div and then simplify the resultant comparison.
2070 if (IsAShr && (!Shr->isExact() || ShAmtVal == TypeBits - 1))
2073 // Revisit the shift (to delete it).
2076 Constant *DivCst = ConstantInt::get(
2077 Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
2079 Value *Tmp = IsAShr ? Builder->CreateSDiv(X, DivCst, "", Shr->isExact())
2080 : Builder->CreateUDiv(X, DivCst, "", Shr->isExact());
2082 Cmp.setOperand(0, Tmp);
2084 // If the builder folded the binop, just return it.
2085 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
2089 // Otherwise, fold this div/compare.
2090 assert(TheDiv->getOpcode() == Instruction::SDiv ||
2091 TheDiv->getOpcode() == Instruction::UDiv);
2093 Instruction *Res = foldICmpDivConstant(Cmp, TheDiv, C);
2094 assert(Res && "This div/cst should have folded!");
2098 // Handle equality comparisons of shift-by-constant.
2100 // If the comparison constant changes with the shift, the comparison cannot
2101 // succeed (bits of the comparison constant cannot match the shifted value).
2102 // This should be known by InstSimplify and already be folded to true/false.
2103 assert(((IsAShr && C->shl(ShAmtVal).ashr(ShAmtVal) == *C) ||
2104 (!IsAShr && C->shl(ShAmtVal).lshr(ShAmtVal) == *C)) &&
2105 "Expected icmp+shr simplify did not occur.");
2107 // Check if the bits shifted out are known to be zero. If so, we can compare
2108 // against the unshifted value:
2109 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2110 Constant *ShiftedCmpRHS = ConstantInt::get(Shr->getType(), *C << ShAmtVal);
2111 if (Shr->hasOneUse()) {
2113 return new ICmpInst(Pred, X, ShiftedCmpRHS);
2115 // Otherwise strength reduce the shift into an 'and'.
2116 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2117 Constant *Mask = ConstantInt::get(Shr->getType(), Val);
2118 Value *And = Builder->CreateAnd(X, Mask, Shr->getName() + ".mask");
2119 return new ICmpInst(Pred, And, ShiftedCmpRHS);
2125 /// Fold icmp (udiv X, Y), C.
2126 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2127 BinaryOperator *UDiv,
2130 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2133 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2135 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2136 Value *Y = UDiv->getOperand(1);
2137 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2138 assert(!C->isMaxValue() &&
2139 "icmp ugt X, UINT_MAX should have been simplified already.");
2140 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2141 ConstantInt::get(Y->getType(), C2->udiv(*C + 1)));
2144 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2145 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2146 assert(*C != 0 && "icmp ult X, 0 should have been simplified already.");
2147 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2148 ConstantInt::get(Y->getType(), C2->udiv(*C)));
2154 /// Fold icmp ({su}div X, Y), C.
2155 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2156 BinaryOperator *Div,
2158 // Fold: icmp pred ([us]div X, C2), C -> range test
2159 // Fold this div into the comparison, producing a range check.
2160 // Determine, based on the divide type, what the range is being
2161 // checked. If there is an overflow on the low or high side, remember
2162 // it, otherwise compute the range [low, hi) bounding the new value.
2163 // See: InsertRangeTest above for the kinds of replacements possible.
2165 if (!match(Div->getOperand(1), m_APInt(C2)))
2168 // FIXME: If the operand types don't match the type of the divide
2169 // then don't attempt this transform. The code below doesn't have the
2170 // logic to deal with a signed divide and an unsigned compare (and
2171 // vice versa). This is because (x /s C2) <s C produces different
2172 // results than (x /s C2) <u C or (x /u C2) <s C or even
2173 // (x /u C2) <u C. Simply casting the operands and result won't
2174 // work. :( The if statement below tests that condition and bails
2176 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2177 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2180 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2181 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2182 // division-by-constant cases should be present, we can not assert that they
2183 // have happened before we reach this icmp instruction.
2184 if (C2->isNullValue() || C2->isOneValue() ||
2185 (DivIsSigned && C2->isAllOnesValue()))
2188 // TODO: We could do all of the computations below using APInt.
2189 Constant *CmpRHS = cast<Constant>(Cmp.getOperand(1));
2190 Constant *DivRHS = cast<Constant>(Div->getOperand(1));
2192 // Compute Prod = CmpRHS * DivRHS. We are essentially solving an equation of
2193 // form X / C2 = C. We solve for X by multiplying C2 (DivRHS) and C (CmpRHS).
2194 // By solving for X, we can turn this into a range check instead of computing
2196 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
2198 // Determine if the product overflows by seeing if the product is not equal to
2199 // the divide. Make sure we do the same kind of divide as in the LHS
2200 // instruction that we're folding.
2201 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS)
2202 : ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
2204 ICmpInst::Predicate Pred = Cmp.getPredicate();
2206 // If the division is known to be exact, then there is no remainder from the
2207 // divide, so the covered range size is unit, otherwise it is the divisor.
2208 Constant *RangeSize =
2209 Div->isExact() ? ConstantInt::get(Div->getType(), 1) : DivRHS;
2211 // Figure out the interval that is being checked. For example, a comparison
2212 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2213 // Compute this interval based on the constants involved and the signedness of
2214 // the compare/divide. This computes a half-open interval, keeping track of
2215 // whether either value in the interval overflows. After analysis each
2216 // overflow variable is set to 0 if it's corresponding bound variable is valid
2217 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2218 int LoOverflow = 0, HiOverflow = 0;
2219 Constant *LoBound = nullptr, *HiBound = nullptr;
2221 if (!DivIsSigned) { // udiv
2222 // e.g. X/5 op 3 --> [15, 20)
2224 HiOverflow = LoOverflow = ProdOV;
2226 // If this is not an exact divide, then many values in the range collapse
2227 // to the same result value.
2228 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2230 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2231 if (C->isNullValue()) { // (X / pos) op 0
2232 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2233 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
2234 HiBound = RangeSize;
2235 } else if (C->isStrictlyPositive()) { // (X / pos) op pos
2236 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2237 HiOverflow = LoOverflow = ProdOV;
2239 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2240 } else { // (X / pos) op neg
2241 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2242 HiBound = AddOne(Prod);
2243 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2245 Constant *DivNeg = ConstantExpr::getNeg(RangeSize);
2246 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2249 } else if (C2->isNegative()) { // Divisor is < 0.
2251 RangeSize = ConstantExpr::getNeg(RangeSize);
2252 if (C->isNullValue()) { // (X / neg) op 0
2253 // e.g. X/-5 op 0 --> [-4, 5)
2254 LoBound = AddOne(RangeSize);
2255 HiBound = ConstantExpr::getNeg(RangeSize);
2256 if (HiBound == DivRHS) { // -INTMIN = INTMIN
2257 HiOverflow = 1; // [INTMIN+1, overflow)
2258 HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
2260 } else if (C->isStrictlyPositive()) { // (X / neg) op pos
2261 // e.g. X/-5 op 3 --> [-19, -14)
2262 HiBound = AddOne(Prod);
2263 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2265 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2266 } else { // (X / neg) op neg
2267 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2268 LoOverflow = HiOverflow = ProdOV;
2270 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2273 // Dividing by a negative swaps the condition. LT <-> GT
2274 Pred = ICmpInst::getSwappedPredicate(Pred);
2277 Value *X = Div->getOperand(0);
2279 default: llvm_unreachable("Unhandled icmp opcode!");
2280 case ICmpInst::ICMP_EQ:
2281 if (LoOverflow && HiOverflow)
2282 return replaceInstUsesWith(Cmp, Builder->getFalse());
2284 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2285 ICmpInst::ICMP_UGE, X, LoBound);
2287 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2288 ICmpInst::ICMP_ULT, X, HiBound);
2289 return replaceInstUsesWith(
2290 Cmp, insertRangeTest(X, LoBound->getUniqueInteger(),
2291 HiBound->getUniqueInteger(), DivIsSigned, true));
2292 case ICmpInst::ICMP_NE:
2293 if (LoOverflow && HiOverflow)
2294 return replaceInstUsesWith(Cmp, Builder->getTrue());
2296 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2297 ICmpInst::ICMP_ULT, X, LoBound);
2299 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2300 ICmpInst::ICMP_UGE, X, HiBound);
2301 return replaceInstUsesWith(Cmp,
2302 insertRangeTest(X, LoBound->getUniqueInteger(),
2303 HiBound->getUniqueInteger(),
2304 DivIsSigned, false));
2305 case ICmpInst::ICMP_ULT:
2306 case ICmpInst::ICMP_SLT:
2307 if (LoOverflow == +1) // Low bound is greater than input range.
2308 return replaceInstUsesWith(Cmp, Builder->getTrue());
2309 if (LoOverflow == -1) // Low bound is less than input range.
2310 return replaceInstUsesWith(Cmp, Builder->getFalse());
2311 return new ICmpInst(Pred, X, LoBound);
2312 case ICmpInst::ICMP_UGT:
2313 case ICmpInst::ICMP_SGT:
2314 if (HiOverflow == +1) // High bound greater than input range.
2315 return replaceInstUsesWith(Cmp, Builder->getFalse());
2316 if (HiOverflow == -1) // High bound less than input range.
2317 return replaceInstUsesWith(Cmp, Builder->getTrue());
2318 if (Pred == ICmpInst::ICMP_UGT)
2319 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
2320 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
2326 /// Fold icmp (sub X, Y), C.
2327 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2328 BinaryOperator *Sub,
2330 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2331 ICmpInst::Predicate Pred = Cmp.getPredicate();
2333 // The following transforms are only worth it if the only user of the subtract
2335 if (!Sub->hasOneUse())
2338 if (Sub->hasNoSignedWrap()) {
2339 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2340 if (Pred == ICmpInst::ICMP_SGT && C->isAllOnesValue())
2341 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2343 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2344 if (Pred == ICmpInst::ICMP_SGT && C->isNullValue())
2345 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2347 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2348 if (Pred == ICmpInst::ICMP_SLT && C->isNullValue())
2349 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2351 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2352 if (Pred == ICmpInst::ICMP_SLT && C->isOneValue())
2353 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2357 if (!match(X, m_APInt(C2)))
2360 // C2 - Y <u C -> (Y | (C - 1)) == C2
2361 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2362 if (Pred == ICmpInst::ICMP_ULT && C->isPowerOf2() &&
2363 (*C2 & (*C - 1)) == (*C - 1))
2364 return new ICmpInst(ICmpInst::ICMP_EQ, Builder->CreateOr(Y, *C - 1), X);
2366 // C2 - Y >u C -> (Y | C) != C2
2367 // iff C2 & C == C and C + 1 is a power of 2
2368 if (Pred == ICmpInst::ICMP_UGT && (*C + 1).isPowerOf2() && (*C2 & *C) == *C)
2369 return new ICmpInst(ICmpInst::ICMP_NE, Builder->CreateOr(Y, *C), X);
2374 /// Fold icmp (add X, Y), C.
2375 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2376 BinaryOperator *Add,
2378 Value *Y = Add->getOperand(1);
2380 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2383 // Fold icmp pred (add X, C2), C.
2384 Value *X = Add->getOperand(0);
2385 Type *Ty = Add->getType();
2386 CmpInst::Predicate Pred = Cmp.getPredicate();
2388 // If the add does not wrap, we can always adjust the compare by subtracting
2389 // the constants. Equality comparisons are handled elsewhere. SGE/SLE are
2390 // canonicalized to SGT/SLT.
2391 if (Add->hasNoSignedWrap() &&
2392 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) {
2394 APInt NewC = C->ssub_ov(*C2, Overflow);
2395 // If there is overflow, the result must be true or false.
2396 // TODO: Can we assert there is no overflow because InstSimplify always
2397 // handles those cases?
2399 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2400 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2403 auto CR = ConstantRange::makeExactICmpRegion(Pred, *C).subtract(*C2);
2404 const APInt &Upper = CR.getUpper();
2405 const APInt &Lower = CR.getLower();
2406 if (Cmp.isSigned()) {
2407 if (Lower.isSignMask())
2408 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2409 if (Upper.isSignMask())
2410 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2412 if (Lower.isMinValue())
2413 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2414 if (Upper.isMinValue())
2415 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2418 if (!Add->hasOneUse())
2421 // X+C <u C2 -> (X & -C2) == C
2422 // iff C & (C2-1) == 0
2423 // C2 is a power of 2
2424 if (Pred == ICmpInst::ICMP_ULT && C->isPowerOf2() && (*C2 & (*C - 1)) == 0)
2425 return new ICmpInst(ICmpInst::ICMP_EQ, Builder->CreateAnd(X, -(*C)),
2426 ConstantExpr::getNeg(cast<Constant>(Y)));
2428 // X+C >u C2 -> (X & ~C2) != C
2430 // C2+1 is a power of 2
2431 if (Pred == ICmpInst::ICMP_UGT && (*C + 1).isPowerOf2() && (*C2 & *C) == 0)
2432 return new ICmpInst(ICmpInst::ICMP_NE, Builder->CreateAnd(X, ~(*C)),
2433 ConstantExpr::getNeg(cast<Constant>(Y)));
2438 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2439 Value *&RHS, ConstantInt *&Less,
2440 ConstantInt *&Equal,
2441 ConstantInt *&Greater) {
2442 // TODO: Generalize this to work with other comparison idioms or ensure
2443 // they get canonicalized into this form.
2445 // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32
2446 // Greater), where Equal, Less and Greater are placeholders for any three
2448 ICmpInst::Predicate PredA, PredB;
2449 if (match(SI->getTrueValue(), m_ConstantInt(Equal)) &&
2450 match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) &&
2451 PredA == ICmpInst::ICMP_EQ &&
2452 match(SI->getFalseValue(),
2453 m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)),
2454 m_ConstantInt(Less), m_ConstantInt(Greater))) &&
2455 PredB == ICmpInst::ICMP_SLT) {
2461 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2462 Instruction *Select,
2465 assert(C && "Cmp RHS should be a constant int!");
2466 // If we're testing a constant value against the result of a three way
2467 // comparison, the result can be expressed directly in terms of the
2468 // original values being compared. Note: We could possibly be more
2469 // aggressive here and remove the hasOneUse test. The original select is
2470 // really likely to simplify or sink when we remove a test of the result.
2471 Value *OrigLHS, *OrigRHS;
2472 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2473 if (Cmp.hasOneUse() &&
2474 matchThreeWayIntCompare(cast<SelectInst>(Select), OrigLHS, OrigRHS,
2475 C1LessThan, C2Equal, C3GreaterThan)) {
2476 assert(C1LessThan && C2Equal && C3GreaterThan);
2478 bool TrueWhenLessThan =
2479 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2481 bool TrueWhenEqual =
2482 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2484 bool TrueWhenGreaterThan =
2485 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2488 // This generates the new instruction that will replace the original Cmp
2489 // Instruction. Instead of enumerating the various combinations when
2490 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2491 // false, we rely on chaining of ORs and future passes of InstCombine to
2492 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2494 // When none of the three constants satisfy the predicate for the RHS (C),
2495 // the entire original Cmp can be simplified to a false.
2496 Value *Cond = Builder->getFalse();
2497 if (TrueWhenLessThan)
2498 Cond = Builder->CreateOr(Cond, Builder->CreateICmp(ICmpInst::ICMP_SLT, OrigLHS, OrigRHS));
2500 Cond = Builder->CreateOr(Cond, Builder->CreateICmp(ICmpInst::ICMP_EQ, OrigLHS, OrigRHS));
2501 if (TrueWhenGreaterThan)
2502 Cond = Builder->CreateOr(Cond, Builder->CreateICmp(ICmpInst::ICMP_SGT, OrigLHS, OrigRHS));
2504 return replaceInstUsesWith(Cmp, Cond);
2509 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2510 /// where X is some kind of instruction.
2511 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2513 if (!match(Cmp.getOperand(1), m_APInt(C)))
2517 if (match(Cmp.getOperand(0), m_BinOp(BO))) {
2518 switch (BO->getOpcode()) {
2519 case Instruction::Xor:
2520 if (Instruction *I = foldICmpXorConstant(Cmp, BO, C))
2523 case Instruction::And:
2524 if (Instruction *I = foldICmpAndConstant(Cmp, BO, C))
2527 case Instruction::Or:
2528 if (Instruction *I = foldICmpOrConstant(Cmp, BO, C))
2531 case Instruction::Mul:
2532 if (Instruction *I = foldICmpMulConstant(Cmp, BO, C))
2535 case Instruction::Shl:
2536 if (Instruction *I = foldICmpShlConstant(Cmp, BO, C))
2539 case Instruction::LShr:
2540 case Instruction::AShr:
2541 if (Instruction *I = foldICmpShrConstant(Cmp, BO, C))
2544 case Instruction::UDiv:
2545 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, C))
2548 case Instruction::SDiv:
2549 if (Instruction *I = foldICmpDivConstant(Cmp, BO, C))
2552 case Instruction::Sub:
2553 if (Instruction *I = foldICmpSubConstant(Cmp, BO, C))
2556 case Instruction::Add:
2557 if (Instruction *I = foldICmpAddConstant(Cmp, BO, C))
2563 // TODO: These folds could be refactored to be part of the above calls.
2564 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, C))
2568 // Match against CmpInst LHS being instructions other than binary operators.
2570 if (match(Cmp.getOperand(0), m_Instruction(LHSI))) {
2571 switch (LHSI->getOpcode()) {
2572 case Instruction::Select:
2574 // For now, we only support constant integers while folding the
2575 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2576 // similar to the cases handled by binary ops above.
2577 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2578 if (Instruction *I = foldICmpSelectConstant(Cmp, LHSI, ConstRHS))
2582 case Instruction::Trunc:
2583 if (Instruction *I = foldICmpTruncConstant(Cmp, LHSI, C))
2591 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, C))
2597 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2598 /// icmp eq/ne BO, C.
2599 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2602 // TODO: Some of these folds could work with arbitrary constants, but this
2603 // function is limited to scalar and vector splat constants.
2604 if (!Cmp.isEquality())
2607 ICmpInst::Predicate Pred = Cmp.getPredicate();
2608 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2609 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2610 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2612 switch (BO->getOpcode()) {
2613 case Instruction::SRem:
2614 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2615 if (C->isNullValue() && BO->hasOneUse()) {
2617 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2618 Value *NewRem = Builder->CreateURem(BOp0, BOp1, BO->getName());
2619 return new ICmpInst(Pred, NewRem,
2620 Constant::getNullValue(BO->getType()));
2624 case Instruction::Add: {
2625 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2627 if (match(BOp1, m_APInt(BOC))) {
2628 if (BO->hasOneUse()) {
2629 Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2630 return new ICmpInst(Pred, BOp0, SubC);
2632 } else if (C->isNullValue()) {
2633 // Replace ((add A, B) != 0) with (A != -B) if A or B is
2634 // efficiently invertible, or if the add has just this one use.
2635 if (Value *NegVal = dyn_castNegVal(BOp1))
2636 return new ICmpInst(Pred, BOp0, NegVal);
2637 if (Value *NegVal = dyn_castNegVal(BOp0))
2638 return new ICmpInst(Pred, NegVal, BOp1);
2639 if (BO->hasOneUse()) {
2640 Value *Neg = Builder->CreateNeg(BOp1);
2642 return new ICmpInst(Pred, BOp0, Neg);
2647 case Instruction::Xor:
2648 if (BO->hasOneUse()) {
2649 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2650 // For the xor case, we can xor two constants together, eliminating
2651 // the explicit xor.
2652 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2653 } else if (C->isNullValue()) {
2654 // Replace ((xor A, B) != 0) with (A != B)
2655 return new ICmpInst(Pred, BOp0, BOp1);
2659 case Instruction::Sub:
2660 if (BO->hasOneUse()) {
2662 if (match(BOp0, m_APInt(BOC))) {
2663 // Replace ((sub BOC, B) != C) with (B != BOC-C).
2664 Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2665 return new ICmpInst(Pred, BOp1, SubC);
2666 } else if (C->isNullValue()) {
2667 // Replace ((sub A, B) != 0) with (A != B).
2668 return new ICmpInst(Pred, BOp0, BOp1);
2672 case Instruction::Or: {
2674 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2675 // Comparing if all bits outside of a constant mask are set?
2676 // Replace (X | C) == -1 with (X & ~C) == ~C.
2677 // This removes the -1 constant.
2678 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2679 Value *And = Builder->CreateAnd(BOp0, NotBOC);
2680 return new ICmpInst(Pred, And, NotBOC);
2684 case Instruction::And: {
2686 if (match(BOp1, m_APInt(BOC))) {
2687 // If we have ((X & C) == C), turn it into ((X & C) != 0).
2688 if (C == BOC && C->isPowerOf2())
2689 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2690 BO, Constant::getNullValue(RHS->getType()));
2692 // Don't perform the following transforms if the AND has multiple uses
2693 if (!BO->hasOneUse())
2696 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
2697 if (BOC->isSignMask()) {
2698 Constant *Zero = Constant::getNullValue(BOp0->getType());
2699 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
2700 return new ICmpInst(NewPred, BOp0, Zero);
2703 // ((X & ~7) == 0) --> X < 8
2704 if (C->isNullValue() && (~(*BOC) + 1).isPowerOf2()) {
2705 Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1));
2706 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
2707 return new ICmpInst(NewPred, BOp0, NegBOC);
2712 case Instruction::Mul:
2713 if (C->isNullValue() && BO->hasNoSignedWrap()) {
2715 if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2716 // The trivial case (mul X, 0) is handled by InstSimplify.
2717 // General case : (mul X, C) != 0 iff X != 0
2718 // (mul X, C) == 0 iff X == 0
2719 return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2723 case Instruction::UDiv:
2724 if (C->isNullValue()) {
2725 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2726 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2727 return new ICmpInst(NewPred, BOp1, BOp0);
2736 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
2737 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2739 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0));
2740 if (!II || !Cmp.isEquality())
2743 // Handle icmp {eq|ne} <intrinsic>, intcst.
2744 switch (II->getIntrinsicID()) {
2745 case Intrinsic::bswap:
2747 Cmp.setOperand(0, II->getArgOperand(0));
2748 Cmp.setOperand(1, Builder->getInt(C->byteSwap()));
2750 case Intrinsic::ctlz:
2751 case Intrinsic::cttz:
2752 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
2753 if (*C == C->getBitWidth()) {
2755 Cmp.setOperand(0, II->getArgOperand(0));
2756 Cmp.setOperand(1, ConstantInt::getNullValue(II->getType()));
2760 case Intrinsic::ctpop: {
2761 // popcount(A) == 0 -> A == 0 and likewise for !=
2762 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
2763 bool IsZero = C->isNullValue();
2764 if (IsZero || *C == C->getBitWidth()) {
2766 Cmp.setOperand(0, II->getArgOperand(0));
2767 auto *NewOp = IsZero ? Constant::getNullValue(II->getType())
2768 : Constant::getAllOnesValue(II->getType());
2769 Cmp.setOperand(1, NewOp);
2780 /// Handle icmp with constant (but not simple integer constant) RHS.
2781 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
2782 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2783 Constant *RHSC = dyn_cast<Constant>(Op1);
2784 Instruction *LHSI = dyn_cast<Instruction>(Op0);
2788 switch (LHSI->getOpcode()) {
2789 case Instruction::GetElementPtr:
2790 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2791 if (RHSC->isNullValue() &&
2792 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2793 return new ICmpInst(
2794 I.getPredicate(), LHSI->getOperand(0),
2795 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2797 case Instruction::PHI:
2798 // Only fold icmp into the PHI if the phi and icmp are in the same
2799 // block. If in the same block, we're encouraging jump threading. If
2800 // not, we are just pessimizing the code by making an i1 phi.
2801 if (LHSI->getParent() == I.getParent())
2802 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
2805 case Instruction::Select: {
2806 // If either operand of the select is a constant, we can fold the
2807 // comparison into the select arms, which will cause one to be
2808 // constant folded and the select turned into a bitwise or.
2809 Value *Op1 = nullptr, *Op2 = nullptr;
2810 ConstantInt *CI = nullptr;
2811 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2812 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2813 CI = dyn_cast<ConstantInt>(Op1);
2815 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2816 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2817 CI = dyn_cast<ConstantInt>(Op2);
2820 // We only want to perform this transformation if it will not lead to
2821 // additional code. This is true if either both sides of the select
2822 // fold to a constant (in which case the icmp is replaced with a select
2823 // which will usually simplify) or this is the only user of the
2824 // select (in which case we are trading a select+icmp for a simpler
2825 // select+icmp) or all uses of the select can be replaced based on
2826 // dominance information ("Global cases").
2827 bool Transform = false;
2830 else if (Op1 || Op2) {
2832 if (LHSI->hasOneUse())
2835 else if (CI && !CI->isZero())
2836 // When Op1 is constant try replacing select with second operand.
2837 // Otherwise Op2 is constant and try replacing select with first
2840 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
2844 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
2847 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
2849 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2853 case Instruction::IntToPtr:
2854 // icmp pred inttoptr(X), null -> icmp pred X, 0
2855 if (RHSC->isNullValue() &&
2856 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
2857 return new ICmpInst(
2858 I.getPredicate(), LHSI->getOperand(0),
2859 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2862 case Instruction::Load:
2863 // Try to optimize things like "A[i] > 4" to index computations.
2864 if (GetElementPtrInst *GEP =
2865 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2866 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2867 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2868 !cast<LoadInst>(LHSI)->isVolatile())
2869 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
2878 /// Try to fold icmp (binop), X or icmp X, (binop).
2879 /// TODO: A large part of this logic is duplicated in InstSimplify's
2880 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
2882 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
2883 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2885 // Special logic for binary operators.
2886 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2887 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2891 const CmpInst::Predicate Pred = I.getPredicate();
2892 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2893 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2895 ICmpInst::isEquality(Pred) ||
2896 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2897 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2898 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2900 ICmpInst::isEquality(Pred) ||
2901 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2902 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2904 // Analyze the case when either Op0 or Op1 is an add instruction.
2905 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2906 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2907 if (BO0 && BO0->getOpcode() == Instruction::Add) {
2908 A = BO0->getOperand(0);
2909 B = BO0->getOperand(1);
2911 if (BO1 && BO1->getOpcode() == Instruction::Add) {
2912 C = BO1->getOperand(0);
2913 D = BO1->getOperand(1);
2916 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2917 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2918 return new ICmpInst(Pred, A == Op1 ? B : A,
2919 Constant::getNullValue(Op1->getType()));
2921 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2922 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2923 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2926 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2927 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
2929 // Try not to increase register pressure.
2930 BO0->hasOneUse() && BO1->hasOneUse()) {
2931 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2934 // C + B == C + D -> B == D
2937 } else if (A == D) {
2938 // D + B == C + D -> B == C
2941 } else if (B == C) {
2942 // A + C == C + D -> A == D
2947 // A + D == C + D -> A == C
2951 return new ICmpInst(Pred, Y, Z);
2954 // icmp slt (X + -1), Y -> icmp sle X, Y
2955 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2956 match(B, m_AllOnes()))
2957 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2959 // icmp sge (X + -1), Y -> icmp sgt X, Y
2960 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2961 match(B, m_AllOnes()))
2962 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2964 // icmp sle (X + 1), Y -> icmp slt X, Y
2965 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
2966 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2968 // icmp sgt (X + 1), Y -> icmp sge X, Y
2969 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
2970 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2972 // icmp sgt X, (Y + -1) -> icmp sge X, Y
2973 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
2974 match(D, m_AllOnes()))
2975 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
2977 // icmp sle X, (Y + -1) -> icmp slt X, Y
2978 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
2979 match(D, m_AllOnes()))
2980 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
2982 // icmp sge X, (Y + 1) -> icmp sgt X, Y
2983 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
2984 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
2986 // icmp slt X, (Y + 1) -> icmp sle X, Y
2987 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
2988 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
2990 // TODO: The subtraction-related identities shown below also hold, but
2991 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
2992 // wouldn't happen even if they were implemented.
2994 // icmp ult (X - 1), Y -> icmp ule X, Y
2995 // icmp uge (X - 1), Y -> icmp ugt X, Y
2996 // icmp ugt X, (Y - 1) -> icmp uge X, Y
2997 // icmp ule X, (Y - 1) -> icmp ult X, Y
2999 // icmp ule (X + 1), Y -> icmp ult X, Y
3000 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3001 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3003 // icmp ugt (X + 1), Y -> icmp uge X, Y
3004 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3005 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3007 // icmp uge X, (Y + 1) -> icmp ugt X, Y
3008 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3009 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3011 // icmp ult X, (Y + 1) -> icmp ule X, Y
3012 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3013 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3015 // if C1 has greater magnitude than C2:
3016 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3017 // s.t. C3 = C1 - C2
3019 // if C2 has greater magnitude than C1:
3020 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3021 // s.t. C3 = C2 - C1
3022 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3023 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3024 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3025 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3026 const APInt &AP1 = C1->getValue();
3027 const APInt &AP2 = C2->getValue();
3028 if (AP1.isNegative() == AP2.isNegative()) {
3029 APInt AP1Abs = C1->getValue().abs();
3030 APInt AP2Abs = C2->getValue().abs();
3031 if (AP1Abs.uge(AP2Abs)) {
3032 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
3033 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
3034 return new ICmpInst(Pred, NewAdd, C);
3036 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
3037 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
3038 return new ICmpInst(Pred, A, NewAdd);
3043 // Analyze the case when either Op0 or Op1 is a sub instruction.
3044 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3049 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3050 A = BO0->getOperand(0);
3051 B = BO0->getOperand(1);
3053 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3054 C = BO1->getOperand(0);
3055 D = BO1->getOperand(1);
3058 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3059 if (A == Op1 && NoOp0WrapProblem)
3060 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3062 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3063 if (C == Op0 && NoOp1WrapProblem)
3064 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3066 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3067 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3068 // Try not to increase register pressure.
3069 BO0->hasOneUse() && BO1->hasOneUse())
3070 return new ICmpInst(Pred, A, C);
3072 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3073 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3074 // Try not to increase register pressure.
3075 BO0->hasOneUse() && BO1->hasOneUse())
3076 return new ICmpInst(Pred, D, B);
3078 // icmp (0-X) < cst --> x > -cst
3079 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3081 if (match(BO0, m_Neg(m_Value(X))))
3082 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3083 if (!RHSC->isMinValue(/*isSigned=*/true))
3084 return new ICmpInst(I.getSwappedPredicate(), X,
3085 ConstantExpr::getNeg(RHSC));
3088 BinaryOperator *SRem = nullptr;
3089 // icmp (srem X, Y), Y
3090 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3092 // icmp Y, (srem X, Y)
3093 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3094 Op0 == BO1->getOperand(1))
3097 // We don't check hasOneUse to avoid increasing register pressure because
3098 // the value we use is the same value this instruction was already using.
3099 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3102 case ICmpInst::ICMP_EQ:
3103 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3104 case ICmpInst::ICMP_NE:
3105 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3106 case ICmpInst::ICMP_SGT:
3107 case ICmpInst::ICMP_SGE:
3108 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3109 Constant::getAllOnesValue(SRem->getType()));
3110 case ICmpInst::ICMP_SLT:
3111 case ICmpInst::ICMP_SLE:
3112 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3113 Constant::getNullValue(SRem->getType()));
3117 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3118 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3119 switch (BO0->getOpcode()) {
3122 case Instruction::Add:
3123 case Instruction::Sub:
3124 case Instruction::Xor: {
3125 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3126 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3129 if (match(BO0->getOperand(1), m_APInt(C))) {
3130 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3131 if (C->isSignMask()) {
3132 ICmpInst::Predicate NewPred =
3133 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3134 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3137 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3138 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3139 ICmpInst::Predicate NewPred =
3140 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3141 NewPred = I.getSwappedPredicate(NewPred);
3142 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3147 case Instruction::Mul: {
3148 if (!I.isEquality())
3152 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3154 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3155 // Mask = -1 >> count-trailing-zeros(C).
3156 if (unsigned TZs = C->countTrailingZeros()) {
3157 Constant *Mask = ConstantInt::get(
3159 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3160 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
3161 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
3162 return new ICmpInst(Pred, And1, And2);
3164 // If there are no trailing zeros in the multiplier, just eliminate
3165 // the multiplies (no masking is needed):
3166 // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3167 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3171 case Instruction::UDiv:
3172 case Instruction::LShr:
3173 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3175 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3177 case Instruction::SDiv:
3178 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3180 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3182 case Instruction::AShr:
3183 if (!BO0->isExact() || !BO1->isExact())
3185 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3187 case Instruction::Shl: {
3188 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3189 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3192 if (!NSW && I.isSigned())
3194 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3200 // Transform A & (L - 1) `ult` L --> L != 0
3201 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3202 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3204 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3205 auto *Zero = Constant::getNullValue(BO0->getType());
3206 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3213 /// Fold icmp Pred min|max(X, Y), X.
3214 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
3215 ICmpInst::Predicate Pred = Cmp.getPredicate();
3216 Value *Op0 = Cmp.getOperand(0);
3217 Value *X = Cmp.getOperand(1);
3219 // Canonicalize minimum or maximum operand to LHS of the icmp.
3220 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3221 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3222 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3223 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3225 Pred = Cmp.getSwappedPredicate();
3229 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3230 // smin(X, Y) == X --> X s<= Y
3231 // smin(X, Y) s>= X --> X s<= Y
3232 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3233 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3235 // smin(X, Y) != X --> X s> Y
3236 // smin(X, Y) s< X --> X s> Y
3237 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3238 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3240 // These cases should be handled in InstSimplify:
3241 // smin(X, Y) s<= X --> true
3242 // smin(X, Y) s> X --> false
3246 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3247 // smax(X, Y) == X --> X s>= Y
3248 // smax(X, Y) s<= X --> X s>= Y
3249 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3250 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3252 // smax(X, Y) != X --> X s< Y
3253 // smax(X, Y) s> X --> X s< Y
3254 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3255 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3257 // These cases should be handled in InstSimplify:
3258 // smax(X, Y) s>= X --> true
3259 // smax(X, Y) s< X --> false
3263 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3264 // umin(X, Y) == X --> X u<= Y
3265 // umin(X, Y) u>= X --> X u<= Y
3266 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3267 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3269 // umin(X, Y) != X --> X u> Y
3270 // umin(X, Y) u< X --> X u> Y
3271 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3272 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3274 // These cases should be handled in InstSimplify:
3275 // umin(X, Y) u<= X --> true
3276 // umin(X, Y) u> X --> false
3280 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3281 // umax(X, Y) == X --> X u>= Y
3282 // umax(X, Y) u<= X --> X u>= Y
3283 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3284 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3286 // umax(X, Y) != X --> X u< Y
3287 // umax(X, Y) u> X --> X u< Y
3288 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3289 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3291 // These cases should be handled in InstSimplify:
3292 // umax(X, Y) u>= X --> true
3293 // umax(X, Y) u< X --> false
3300 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3301 if (!I.isEquality())
3304 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3305 const CmpInst::Predicate Pred = I.getPredicate();
3306 Value *A, *B, *C, *D;
3307 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3308 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3309 Value *OtherVal = A == Op1 ? B : A;
3310 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3313 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3314 // A^c1 == C^c2 --> A == C^(c1^c2)
3315 ConstantInt *C1, *C2;
3316 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3318 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3319 Value *Xor = Builder->CreateXor(C, NC);
3320 return new ICmpInst(Pred, A, Xor);
3323 // A^B == A^D -> B == D
3325 return new ICmpInst(Pred, B, D);
3327 return new ICmpInst(Pred, B, C);
3329 return new ICmpInst(Pred, A, D);
3331 return new ICmpInst(Pred, A, C);
3335 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3336 // A == (A^B) -> B == 0
3337 Value *OtherVal = A == Op0 ? B : A;
3338 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3341 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3342 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3343 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3344 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3350 } else if (A == D) {
3354 } else if (B == C) {
3358 } else if (B == D) {
3364 if (X) { // Build (X^Y) & Z
3365 Op1 = Builder->CreateXor(X, Y);
3366 Op1 = Builder->CreateAnd(Op1, Z);
3367 I.setOperand(0, Op1);
3368 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3373 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3374 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3376 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3377 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3378 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3379 match(Op1, m_ZExt(m_Value(A))))) {
3380 APInt Pow2 = Cst1->getValue() + 1;
3381 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3382 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3383 return new ICmpInst(Pred, A, Builder->CreateTrunc(B, A->getType()));
3386 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3387 // For lshr and ashr pairs.
3388 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3389 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3390 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3391 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3392 unsigned TypeBits = Cst1->getBitWidth();
3393 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3394 if (ShAmt < TypeBits && ShAmt != 0) {
3395 ICmpInst::Predicate NewPred =
3396 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
3397 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3398 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3399 return new ICmpInst(NewPred, Xor, Builder->getInt(CmpVal));
3403 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3404 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3405 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3406 unsigned TypeBits = Cst1->getBitWidth();
3407 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3408 if (ShAmt < TypeBits && ShAmt != 0) {
3409 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3410 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3411 Value *And = Builder->CreateAnd(Xor, Builder->getInt(AndVal),
3412 I.getName() + ".mask");
3413 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
3417 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3418 // "icmp (and X, mask), cst"
3420 if (Op0->hasOneUse() &&
3421 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3422 match(Op1, m_ConstantInt(Cst1)) &&
3423 // Only do this when A has multiple uses. This is most important to do
3424 // when it exposes other optimizations.
3426 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3428 if (ShAmt < ASize) {
3430 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3433 APInt CmpV = Cst1->getValue().zext(ASize);
3436 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3437 return new ICmpInst(Pred, Mask, Builder->getInt(CmpV));
3444 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3446 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3447 const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3448 Value *LHSCIOp = LHSCI->getOperand(0);
3449 Type *SrcTy = LHSCIOp->getType();
3450 Type *DestTy = LHSCI->getType();
3453 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3454 // integer type is the same size as the pointer type.
3455 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3456 DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
3457 Value *RHSOp = nullptr;
3458 if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3459 Value *RHSCIOp = RHSC->getOperand(0);
3460 if (RHSCIOp->getType()->getPointerAddressSpace() ==
3461 LHSCIOp->getType()->getPointerAddressSpace()) {
3462 RHSOp = RHSC->getOperand(0);
3463 // If the pointer types don't match, insert a bitcast.
3464 if (LHSCIOp->getType() != RHSOp->getType())
3465 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
3467 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
3468 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
3472 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
3475 // The code below only handles extension cast instructions, so far.
3477 if (LHSCI->getOpcode() != Instruction::ZExt &&
3478 LHSCI->getOpcode() != Instruction::SExt)
3481 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
3482 bool isSignedCmp = ICmp.isSigned();
3484 if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
3485 // Not an extension from the same type?
3486 RHSCIOp = CI->getOperand(0);
3487 if (RHSCIOp->getType() != LHSCIOp->getType())
3490 // If the signedness of the two casts doesn't agree (i.e. one is a sext
3491 // and the other is a zext), then we can't handle this.
3492 if (CI->getOpcode() != LHSCI->getOpcode())
3495 // Deal with equality cases early.
3496 if (ICmp.isEquality())
3497 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3499 // A signed comparison of sign extended values simplifies into a
3500 // signed comparison.
3501 if (isSignedCmp && isSignedExt)
3502 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3504 // The other three cases all fold into an unsigned comparison.
3505 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
3508 // If we aren't dealing with a constant on the RHS, exit early.
3509 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
3513 // Compute the constant that would happen if we truncated to SrcTy then
3514 // re-extended to DestTy.
3515 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
3516 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
3518 // If the re-extended constant didn't change...
3520 // Deal with equality cases early.
3521 if (ICmp.isEquality())
3522 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3524 // A signed comparison of sign extended values simplifies into a
3525 // signed comparison.
3526 if (isSignedExt && isSignedCmp)
3527 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3529 // The other three cases all fold into an unsigned comparison.
3530 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
3533 // The re-extended constant changed, partly changed (in the case of a vector),
3534 // or could not be determined to be equal (in the case of a constant
3535 // expression), so the constant cannot be represented in the shorter type.
3536 // Consequently, we cannot emit a simple comparison.
3537 // All the cases that fold to true or false will have already been handled
3538 // by SimplifyICmpInst, so only deal with the tricky case.
3540 if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
3543 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
3544 // should have been folded away previously and not enter in here.
3546 // We're performing an unsigned comp with a sign extended value.
3547 // This is true if the input is >= 0. [aka >s -1]
3548 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
3549 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
3551 // Finally, return the value computed.
3552 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
3553 return replaceInstUsesWith(ICmp, Result);
3555 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
3556 return BinaryOperator::CreateNot(Result);
3559 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
3560 Value *RHS, Instruction &OrigI,
3561 Value *&Result, Constant *&Overflow) {
3562 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
3563 std::swap(LHS, RHS);
3565 auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
3567 Overflow = OverflowVal;
3569 Result->takeName(&OrigI);
3573 // If the overflow check was an add followed by a compare, the insertion point
3574 // may be pointing to the compare. We want to insert the new instructions
3575 // before the add in case there are uses of the add between the add and the
3577 Builder->SetInsertPoint(&OrigI);
3581 llvm_unreachable("bad overflow check kind!");
3583 case OCF_UNSIGNED_ADD: {
3584 OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
3585 if (OR == OverflowResult::NeverOverflows)
3586 return SetResult(Builder->CreateNUWAdd(LHS, RHS), Builder->getFalse(),
3589 if (OR == OverflowResult::AlwaysOverflows)
3590 return SetResult(Builder->CreateAdd(LHS, RHS), Builder->getTrue(), true);
3592 // Fall through uadd into sadd
3595 case OCF_SIGNED_ADD: {
3596 // X + 0 -> {X, false}
3597 if (match(RHS, m_Zero()))
3598 return SetResult(LHS, Builder->getFalse(), false);
3600 // We can strength reduce this signed add into a regular add if we can prove
3601 // that it will never overflow.
3602 if (OCF == OCF_SIGNED_ADD)
3603 if (willNotOverflowSignedAdd(LHS, RHS, OrigI))
3604 return SetResult(Builder->CreateNSWAdd(LHS, RHS), Builder->getFalse(),
3609 case OCF_UNSIGNED_SUB:
3610 case OCF_SIGNED_SUB: {
3611 // X - 0 -> {X, false}
3612 if (match(RHS, m_Zero()))
3613 return SetResult(LHS, Builder->getFalse(), false);
3615 if (OCF == OCF_SIGNED_SUB) {
3616 if (willNotOverflowSignedSub(LHS, RHS, OrigI))
3617 return SetResult(Builder->CreateNSWSub(LHS, RHS), Builder->getFalse(),
3620 if (willNotOverflowUnsignedSub(LHS, RHS, OrigI))
3621 return SetResult(Builder->CreateNUWSub(LHS, RHS), Builder->getFalse(),
3627 case OCF_UNSIGNED_MUL: {
3628 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
3629 if (OR == OverflowResult::NeverOverflows)
3630 return SetResult(Builder->CreateNUWMul(LHS, RHS), Builder->getFalse(),
3632 if (OR == OverflowResult::AlwaysOverflows)
3633 return SetResult(Builder->CreateMul(LHS, RHS), Builder->getTrue(), true);
3636 case OCF_SIGNED_MUL:
3637 // X * undef -> undef
3638 if (isa<UndefValue>(RHS))
3639 return SetResult(RHS, UndefValue::get(Builder->getInt1Ty()), false);
3641 // X * 0 -> {0, false}
3642 if (match(RHS, m_Zero()))
3643 return SetResult(RHS, Builder->getFalse(), false);
3645 // X * 1 -> {X, false}
3646 if (match(RHS, m_One()))
3647 return SetResult(LHS, Builder->getFalse(), false);
3649 if (OCF == OCF_SIGNED_MUL)
3650 if (willNotOverflowSignedMul(LHS, RHS, OrigI))
3651 return SetResult(Builder->CreateNSWMul(LHS, RHS), Builder->getFalse(),
3659 /// \brief Recognize and process idiom involving test for multiplication
3662 /// The caller has matched a pattern of the form:
3663 /// I = cmp u (mul(zext A, zext B), V
3664 /// The function checks if this is a test for overflow and if so replaces
3665 /// multiplication with call to 'mul.with.overflow' intrinsic.
3667 /// \param I Compare instruction.
3668 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
3669 /// the compare instruction. Must be of integer type.
3670 /// \param OtherVal The other argument of compare instruction.
3671 /// \returns Instruction which must replace the compare instruction, NULL if no
3672 /// replacement required.
3673 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
3674 Value *OtherVal, InstCombiner &IC) {
3675 // Don't bother doing this transformation for pointers, don't do it for
3677 if (!isa<IntegerType>(MulVal->getType()))
3680 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
3681 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
3682 auto *MulInstr = dyn_cast<Instruction>(MulVal);
3685 assert(MulInstr->getOpcode() == Instruction::Mul);
3687 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
3688 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
3689 assert(LHS->getOpcode() == Instruction::ZExt);
3690 assert(RHS->getOpcode() == Instruction::ZExt);
3691 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
3693 // Calculate type and width of the result produced by mul.with.overflow.
3694 Type *TyA = A->getType(), *TyB = B->getType();
3695 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
3696 WidthB = TyB->getPrimitiveSizeInBits();
3699 if (WidthB > WidthA) {
3707 // In order to replace the original mul with a narrower mul.with.overflow,
3708 // all uses must ignore upper bits of the product. The number of used low
3709 // bits must be not greater than the width of mul.with.overflow.
3710 if (MulVal->hasNUsesOrMore(2))
3711 for (User *U : MulVal->users()) {
3714 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3715 // Check if truncation ignores bits above MulWidth.
3716 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
3717 if (TruncWidth > MulWidth)
3719 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3720 // Check if AND ignores bits above MulWidth.
3721 if (BO->getOpcode() != Instruction::And)
3723 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3724 const APInt &CVal = CI->getValue();
3725 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
3729 // Other uses prohibit this transformation.
3734 // Recognize patterns
3735 switch (I.getPredicate()) {
3736 case ICmpInst::ICMP_EQ:
3737 case ICmpInst::ICMP_NE:
3738 // Recognize pattern:
3739 // mulval = mul(zext A, zext B)
3740 // cmp eq/neq mulval, zext trunc mulval
3741 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
3742 if (Zext->hasOneUse()) {
3743 Value *ZextArg = Zext->getOperand(0);
3744 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
3745 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
3749 // Recognize pattern:
3750 // mulval = mul(zext A, zext B)
3751 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
3754 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
3755 if (ValToMask != MulVal)
3757 const APInt &CVal = CI->getValue() + 1;
3758 if (CVal.isPowerOf2()) {
3759 unsigned MaskWidth = CVal.logBase2();
3760 if (MaskWidth == MulWidth)
3761 break; // Recognized
3766 case ICmpInst::ICMP_UGT:
3767 // Recognize pattern:
3768 // mulval = mul(zext A, zext B)
3769 // cmp ugt mulval, max
3770 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3771 APInt MaxVal = APInt::getMaxValue(MulWidth);
3772 MaxVal = MaxVal.zext(CI->getBitWidth());
3773 if (MaxVal.eq(CI->getValue()))
3774 break; // Recognized
3778 case ICmpInst::ICMP_UGE:
3779 // Recognize pattern:
3780 // mulval = mul(zext A, zext B)
3781 // cmp uge mulval, max+1
3782 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3783 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
3784 if (MaxVal.eq(CI->getValue()))
3785 break; // Recognized
3789 case ICmpInst::ICMP_ULE:
3790 // Recognize pattern:
3791 // mulval = mul(zext A, zext B)
3792 // cmp ule mulval, max
3793 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3794 APInt MaxVal = APInt::getMaxValue(MulWidth);
3795 MaxVal = MaxVal.zext(CI->getBitWidth());
3796 if (MaxVal.eq(CI->getValue()))
3797 break; // Recognized
3801 case ICmpInst::ICMP_ULT:
3802 // Recognize pattern:
3803 // mulval = mul(zext A, zext B)
3804 // cmp ule mulval, max + 1
3805 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3806 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
3807 if (MaxVal.eq(CI->getValue()))
3808 break; // Recognized
3816 InstCombiner::BuilderTy *Builder = IC.Builder;
3817 Builder->SetInsertPoint(MulInstr);
3819 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
3820 Value *MulA = A, *MulB = B;
3821 if (WidthA < MulWidth)
3822 MulA = Builder->CreateZExt(A, MulType);
3823 if (WidthB < MulWidth)
3824 MulB = Builder->CreateZExt(B, MulType);
3825 Value *F = Intrinsic::getDeclaration(I.getModule(),
3826 Intrinsic::umul_with_overflow, MulType);
3827 CallInst *Call = Builder->CreateCall(F, {MulA, MulB}, "umul");
3828 IC.Worklist.Add(MulInstr);
3830 // If there are uses of mul result other than the comparison, we know that
3831 // they are truncation or binary AND. Change them to use result of
3832 // mul.with.overflow and adjust properly mask/size.
3833 if (MulVal->hasNUsesOrMore(2)) {
3834 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
3835 for (User *U : MulVal->users()) {
3836 if (U == &I || U == OtherVal)
3838 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3839 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
3840 IC.replaceInstUsesWith(*TI, Mul);
3842 TI->setOperand(0, Mul);
3843 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3844 assert(BO->getOpcode() == Instruction::And);
3845 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
3846 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
3847 APInt ShortMask = CI->getValue().trunc(MulWidth);
3848 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
3850 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
3851 IC.Worklist.Add(Zext);
3852 IC.replaceInstUsesWith(*BO, Zext);
3854 llvm_unreachable("Unexpected Binary operation");
3856 IC.Worklist.Add(cast<Instruction>(U));
3859 if (isa<Instruction>(OtherVal))
3860 IC.Worklist.Add(cast<Instruction>(OtherVal));
3862 // The original icmp gets replaced with the overflow value, maybe inverted
3863 // depending on predicate.
3864 bool Inverse = false;
3865 switch (I.getPredicate()) {
3866 case ICmpInst::ICMP_NE:
3868 case ICmpInst::ICMP_EQ:
3871 case ICmpInst::ICMP_UGT:
3872 case ICmpInst::ICMP_UGE:
3873 if (I.getOperand(0) == MulVal)
3877 case ICmpInst::ICMP_ULT:
3878 case ICmpInst::ICMP_ULE:
3879 if (I.getOperand(1) == MulVal)
3884 llvm_unreachable("Unexpected predicate");
3887 Value *Res = Builder->CreateExtractValue(Call, 1);
3888 return BinaryOperator::CreateNot(Res);
3891 return ExtractValueInst::Create(Call, 1);
3894 /// When performing a comparison against a constant, it is possible that not all
3895 /// the bits in the LHS are demanded. This helper method computes the mask that
3897 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth,
3900 return APInt::getSignMask(BitWidth);
3902 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
3903 if (!CI) return APInt::getAllOnesValue(BitWidth);
3904 const APInt &RHS = CI->getValue();
3906 switch (I.getPredicate()) {
3907 // For a UGT comparison, we don't care about any bits that
3908 // correspond to the trailing ones of the comparand. The value of these
3909 // bits doesn't impact the outcome of the comparison, because any value
3910 // greater than the RHS must differ in a bit higher than these due to carry.
3911 case ICmpInst::ICMP_UGT: {
3912 unsigned trailingOnes = RHS.countTrailingOnes();
3913 return APInt::getBitsSetFrom(BitWidth, trailingOnes);
3916 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
3917 // Any value less than the RHS must differ in a higher bit because of carries.
3918 case ICmpInst::ICMP_ULT: {
3919 unsigned trailingZeros = RHS.countTrailingZeros();
3920 return APInt::getBitsSetFrom(BitWidth, trailingZeros);
3924 return APInt::getAllOnesValue(BitWidth);
3928 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
3929 /// should be swapped.
3930 /// The decision is based on how many times these two operands are reused
3931 /// as subtract operands and their positions in those instructions.
3932 /// The rational is that several architectures use the same instruction for
3933 /// both subtract and cmp, thus it is better if the order of those operands
3935 /// \return true if Op0 and Op1 should be swapped.
3936 static bool swapMayExposeCSEOpportunities(const Value * Op0,
3937 const Value * Op1) {
3938 // Filter out pointer value as those cannot appears directly in subtract.
3939 // FIXME: we may want to go through inttoptrs or bitcasts.
3940 if (Op0->getType()->isPointerTy())
3942 // Count every uses of both Op0 and Op1 in a subtract.
3943 // Each time Op0 is the first operand, count -1: swapping is bad, the
3944 // subtract has already the same layout as the compare.
3945 // Each time Op0 is the second operand, count +1: swapping is good, the
3946 // subtract has a different layout as the compare.
3947 // At the end, if the benefit is greater than 0, Op0 should come second to
3948 // expose more CSE opportunities.
3949 int GlobalSwapBenefits = 0;
3950 for (const User *U : Op0->users()) {
3951 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
3952 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
3954 // If Op0 is the first argument, this is not beneficial to swap the
3956 int LocalSwapBenefits = -1;
3957 unsigned Op1Idx = 1;
3958 if (BinOp->getOperand(Op1Idx) == Op0) {
3960 LocalSwapBenefits = 1;
3962 if (BinOp->getOperand(Op1Idx) != Op1)
3964 GlobalSwapBenefits += LocalSwapBenefits;
3966 return GlobalSwapBenefits > 0;
3969 /// \brief Check that one use is in the same block as the definition and all
3970 /// other uses are in blocks dominated by a given block.
3972 /// \param DI Definition
3974 /// \param DB Block that must dominate all uses of \p DI outside
3975 /// the parent block
3976 /// \return true when \p UI is the only use of \p DI in the parent block
3977 /// and all other uses of \p DI are in blocks dominated by \p DB.
3979 bool InstCombiner::dominatesAllUses(const Instruction *DI,
3980 const Instruction *UI,
3981 const BasicBlock *DB) const {
3982 assert(DI && UI && "Instruction not defined\n");
3983 // Ignore incomplete definitions.
3984 if (!DI->getParent())
3986 // DI and UI must be in the same block.
3987 if (DI->getParent() != UI->getParent())
3989 // Protect from self-referencing blocks.
3990 if (DI->getParent() == DB)
3992 for (const User *U : DI->users()) {
3993 auto *Usr = cast<Instruction>(U);
3994 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4000 /// Return true when the instruction sequence within a block is select-cmp-br.
4001 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4002 const BasicBlock *BB = SI->getParent();
4005 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4006 if (!BI || BI->getNumSuccessors() != 2)
4008 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4009 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4014 /// \brief True when a select result is replaced by one of its operands
4015 /// in select-icmp sequence. This will eventually result in the elimination
4018 /// \param SI Select instruction
4019 /// \param Icmp Compare instruction
4020 /// \param SIOpd Operand that replaces the select
4023 /// - The replacement is global and requires dominator information
4024 /// - The caller is responsible for the actual replacement
4029 /// %4 = select i1 %3, %C* %0, %C* null
4030 /// %5 = icmp eq %C* %4, null
4031 /// br i1 %5, label %9, label %7
4033 /// ; <label>:7 ; preds = %entry
4034 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4037 /// can be transformed to
4039 /// %5 = icmp eq %C* %0, null
4040 /// %6 = select i1 %3, i1 %5, i1 true
4041 /// br i1 %6, label %9, label %7
4043 /// ; <label>:7 ; preds = %entry
4044 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4046 /// Similar when the first operand of the select is a constant or/and
4047 /// the compare is for not equal rather than equal.
4049 /// NOTE: The function is only called when the select and compare constants
4050 /// are equal, the optimization can work only for EQ predicates. This is not a
4051 /// major restriction since a NE compare should be 'normalized' to an equal
4052 /// compare, which usually happens in the combiner and test case
4053 /// select-cmp-br.ll checks for it.
4054 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
4055 const ICmpInst *Icmp,
4056 const unsigned SIOpd) {
4057 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4058 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4059 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4060 // The check for the single predecessor is not the best that can be
4061 // done. But it protects efficiently against cases like when SI's
4062 // home block has two successors, Succ and Succ1, and Succ1 predecessor
4063 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4064 // replaced can be reached on either path. So the uniqueness check
4065 // guarantees that the path all uses of SI (outside SI's parent) are on
4066 // is disjoint from all other paths out of SI. But that information
4067 // is more expensive to compute, and the trade-off here is in favor
4068 // of compile-time. It should also be noticed that we check for a single
4069 // predecessor and not only uniqueness. This to handle the situation when
4070 // Succ and Succ1 points to the same basic block.
4071 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4073 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4080 /// Try to fold the comparison based on range information we can get by checking
4081 /// whether bits are known to be zero or one in the inputs.
4082 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4083 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4084 Type *Ty = Op0->getType();
4085 ICmpInst::Predicate Pred = I.getPredicate();
4087 // Get scalar or pointer size.
4088 unsigned BitWidth = Ty->isIntOrIntVectorTy()
4089 ? Ty->getScalarSizeInBits()
4090 : DL.getTypeSizeInBits(Ty->getScalarType());
4095 // If this is a normal comparison, it demands all bits. If it is a sign bit
4096 // comparison, it only demands the sign bit.
4097 bool IsSignBit = false;
4099 if (match(Op1, m_APInt(CmpC))) {
4101 IsSignBit = isSignBitCheck(Pred, *CmpC, UnusedBit);
4104 KnownBits Op0Known(BitWidth);
4105 KnownBits Op1Known(BitWidth);
4107 if (SimplifyDemandedBits(&I, 0,
4108 getDemandedBitsLHSMask(I, BitWidth, IsSignBit),
4112 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4116 // Given the known and unknown bits, compute a range that the LHS could be
4117 // in. Compute the Min, Max and RHS values based on the known bits. For the
4118 // EQ and NE we use unsigned values.
4119 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4120 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4122 computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4123 computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4125 computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4126 computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4129 // If Min and Max are known to be the same, then SimplifyDemandedBits
4130 // figured out that the LHS is a constant. Constant fold this now, so that
4131 // code below can assume that Min != Max.
4132 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4133 return new ICmpInst(Pred, ConstantInt::get(Op0->getType(), Op0Min), Op1);
4134 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4135 return new ICmpInst(Pred, Op0, ConstantInt::get(Op1->getType(), Op1Min));
4137 // Based on the range information we know about the LHS, see if we can
4138 // simplify this comparison. For example, (x&4) < 8 is always true.
4141 llvm_unreachable("Unknown icmp opcode!");
4142 case ICmpInst::ICMP_EQ:
4143 case ICmpInst::ICMP_NE: {
4144 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4145 return Pred == CmpInst::ICMP_EQ
4146 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4147 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4150 // If all bits are known zero except for one, then we know at most one bit
4151 // is set. If the comparison is against zero, then this is a check to see if
4152 // *that* bit is set.
4153 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4154 if (Op1Known.isZero()) {
4155 // If the LHS is an AND with the same constant, look through it.
4156 Value *LHS = nullptr;
4158 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4159 *LHSC != Op0KnownZeroInverted)
4163 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4164 APInt ValToCheck = Op0KnownZeroInverted;
4165 Type *XTy = X->getType();
4166 if (ValToCheck.isPowerOf2()) {
4167 // ((1 << X) & 8) == 0 -> X != 3
4168 // ((1 << X) & 8) != 0 -> X == 3
4169 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4170 auto NewPred = ICmpInst::getInversePredicate(Pred);
4171 return new ICmpInst(NewPred, X, CmpC);
4172 } else if ((++ValToCheck).isPowerOf2()) {
4173 // ((1 << X) & 7) == 0 -> X >= 3
4174 // ((1 << X) & 7) != 0 -> X < 3
4175 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4177 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
4178 return new ICmpInst(NewPred, X, CmpC);
4182 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4184 if (Op0KnownZeroInverted.isOneValue() &&
4185 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4186 // ((8 >>u X) & 1) == 0 -> X != 3
4187 // ((8 >>u X) & 1) != 0 -> X == 3
4188 unsigned CmpVal = CI->countTrailingZeros();
4189 auto NewPred = ICmpInst::getInversePredicate(Pred);
4190 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4195 case ICmpInst::ICMP_ULT: {
4196 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4197 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4198 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4199 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4200 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4201 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4204 if (match(Op1, m_APInt(CmpC))) {
4205 // A <u C -> A == C-1 if min(A)+1 == C
4206 if (Op1Max == Op0Min + 1) {
4207 Constant *CMinus1 = ConstantInt::get(Op0->getType(), *CmpC - 1);
4208 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, CMinus1);
4213 case ICmpInst::ICMP_UGT: {
4214 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4215 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4217 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4218 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4220 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4221 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4224 if (match(Op1, m_APInt(CmpC))) {
4225 // A >u C -> A == C+1 if max(a)-1 == C
4226 if (*CmpC == Op0Max - 1)
4227 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4228 ConstantInt::get(Op1->getType(), *CmpC + 1));
4232 case ICmpInst::ICMP_SLT:
4233 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4234 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4235 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4236 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4237 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4238 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4239 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4240 if (Op1Max == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4241 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4242 Builder->getInt(CI->getValue() - 1));
4245 case ICmpInst::ICMP_SGT:
4246 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4247 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4248 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4249 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4251 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4252 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4253 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4254 if (Op1Min == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4255 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4256 Builder->getInt(CI->getValue() + 1));
4259 case ICmpInst::ICMP_SGE:
4260 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4261 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4262 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4263 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4264 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4266 case ICmpInst::ICMP_SLE:
4267 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4268 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4269 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4270 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4271 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4273 case ICmpInst::ICMP_UGE:
4274 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4275 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4276 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4277 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4278 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4280 case ICmpInst::ICMP_ULE:
4281 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4282 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4283 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4284 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4285 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4289 // Turn a signed comparison into an unsigned one if both operands are known to
4290 // have the same sign.
4292 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
4293 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
4294 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4299 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4300 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4301 /// allows them to be folded in visitICmpInst.
4302 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
4303 ICmpInst::Predicate Pred = I.getPredicate();
4304 if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4305 Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4308 Value *Op0 = I.getOperand(0);
4309 Value *Op1 = I.getOperand(1);
4310 auto *Op1C = dyn_cast<Constant>(Op1);
4314 // Check if the constant operand can be safely incremented/decremented without
4315 // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4316 // the edge cases for us, so we just assert on them. For vectors, we must
4317 // handle the edge cases.
4318 Type *Op1Type = Op1->getType();
4319 bool IsSigned = I.isSigned();
4320 bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4321 auto *CI = dyn_cast<ConstantInt>(Op1C);
4323 // A <= MAX -> TRUE ; A >= MIN -> TRUE
4324 assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4325 } else if (Op1Type->isVectorTy()) {
4326 // TODO? If the edge cases for vectors were guaranteed to be handled as they
4327 // are for scalar, we could remove the min/max checks. However, to do that,
4328 // we would have to use insertelement/shufflevector to replace edge values.
4329 unsigned NumElts = Op1Type->getVectorNumElements();
4330 for (unsigned i = 0; i != NumElts; ++i) {
4331 Constant *Elt = Op1C->getAggregateElement(i);
4335 if (isa<UndefValue>(Elt))
4338 // Bail out if we can't determine if this constant is min/max or if we
4339 // know that this constant is min/max.
4340 auto *CI = dyn_cast<ConstantInt>(Elt);
4341 if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4349 // Increment or decrement the constant and set the new comparison predicate:
4350 // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4351 Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4352 CmpInst::Predicate NewPred = IsLE ? ICmpInst::ICMP_ULT: ICmpInst::ICMP_UGT;
4353 NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4354 return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4357 /// Integer compare with boolean values can always be turned into bitwise ops.
4358 static Instruction *canonicalizeICmpBool(ICmpInst &I,
4359 InstCombiner::BuilderTy &Builder) {
4360 Value *A = I.getOperand(0), *B = I.getOperand(1);
4361 assert(A->getType()->getScalarType()->isIntegerTy(1) && "Bools only");
4363 // A boolean compared to true/false can be simplified to Op0/true/false in
4364 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
4365 // Cases not handled by InstSimplify are always 'not' of Op0.
4366 if (match(B, m_Zero())) {
4367 switch (I.getPredicate()) {
4368 case CmpInst::ICMP_EQ: // A == 0 -> !A
4369 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
4370 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
4371 return BinaryOperator::CreateNot(A);
4373 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4375 } else if (match(B, m_One())) {
4376 switch (I.getPredicate()) {
4377 case CmpInst::ICMP_NE: // A != 1 -> !A
4378 case CmpInst::ICMP_ULT: // A <u 1 -> !A
4379 case CmpInst::ICMP_SGT: // A >s -1 -> !A
4380 return BinaryOperator::CreateNot(A);
4382 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4386 switch (I.getPredicate()) {
4388 llvm_unreachable("Invalid icmp instruction!");
4389 case ICmpInst::ICMP_EQ:
4390 // icmp eq i1 A, B -> ~(A ^ B)
4391 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
4393 case ICmpInst::ICMP_NE:
4394 // icmp ne i1 A, B -> A ^ B
4395 return BinaryOperator::CreateXor(A, B);
4397 case ICmpInst::ICMP_UGT:
4398 // icmp ugt -> icmp ult
4401 case ICmpInst::ICMP_ULT:
4402 // icmp ult i1 A, B -> ~A & B
4403 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
4405 case ICmpInst::ICMP_SGT:
4406 // icmp sgt -> icmp slt
4409 case ICmpInst::ICMP_SLT:
4410 // icmp slt i1 A, B -> A & ~B
4411 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
4413 case ICmpInst::ICMP_UGE:
4414 // icmp uge -> icmp ule
4417 case ICmpInst::ICMP_ULE:
4418 // icmp ule i1 A, B -> ~A | B
4419 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
4421 case ICmpInst::ICMP_SGE:
4422 // icmp sge -> icmp sle
4425 case ICmpInst::ICMP_SLE:
4426 // icmp sle i1 A, B -> A | ~B
4427 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
4431 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4432 bool Changed = false;
4433 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4434 unsigned Op0Cplxity = getComplexity(Op0);
4435 unsigned Op1Cplxity = getComplexity(Op1);
4437 /// Orders the operands of the compare so that they are listed from most
4438 /// complex to least complex. This puts constants before unary operators,
4439 /// before binary operators.
4440 if (Op0Cplxity < Op1Cplxity ||
4441 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
4443 std::swap(Op0, Op1);
4447 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1,
4448 SQ.getWithInstruction(&I)))
4449 return replaceInstUsesWith(I, V);
4451 // comparing -val or val with non-zero is the same as just comparing val
4452 // ie, abs(val) != 0 -> val != 0
4453 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
4454 Value *Cond, *SelectTrue, *SelectFalse;
4455 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
4456 m_Value(SelectFalse)))) {
4457 if (Value *V = dyn_castNegVal(SelectTrue)) {
4458 if (V == SelectFalse)
4459 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4461 else if (Value *V = dyn_castNegVal(SelectFalse)) {
4462 if (V == SelectTrue)
4463 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4468 if (Op0->getType()->getScalarType()->isIntegerTy(1))
4469 if (Instruction *Res = canonicalizeICmpBool(I, *Builder))
4472 if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
4475 if (Instruction *Res = foldICmpWithConstant(I))
4478 if (Instruction *Res = foldICmpUsingKnownBits(I))
4481 // Test if the ICmpInst instruction is used exclusively by a select as
4482 // part of a minimum or maximum operation. If so, refrain from doing
4483 // any other folding. This helps out other analyses which understand
4484 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4485 // and CodeGen. And in this case, at least one of the comparison
4486 // operands has at least one user besides the compare (the select),
4487 // which would often largely negate the benefit of folding anyway.
4489 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
4490 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
4491 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
4494 // FIXME: We only do this after checking for min/max to prevent infinite
4495 // looping caused by a reverse canonicalization of these patterns for min/max.
4496 // FIXME: The organization of folds is a mess. These would naturally go into
4497 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
4498 // down here after the min/max restriction.
4499 ICmpInst::Predicate Pred = I.getPredicate();
4501 if (match(Op1, m_APInt(C))) {
4502 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
4503 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
4504 Constant *Zero = Constant::getNullValue(Op0->getType());
4505 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
4508 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
4509 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
4510 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
4511 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
4515 if (Instruction *Res = foldICmpInstWithConstant(I))
4518 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
4521 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4522 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
4523 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
4525 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
4526 if (Instruction *NI = foldGEPICmp(GEP, Op0,
4527 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
4530 // Try to optimize equality comparisons against alloca-based pointers.
4531 if (Op0->getType()->isPointerTy() && I.isEquality()) {
4532 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
4533 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
4534 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
4536 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
4537 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
4541 // Test to see if the operands of the icmp are casted versions of other
4542 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4544 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4545 if (Op0->getType()->isPointerTy() &&
4546 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4547 // We keep moving the cast from the left operand over to the right
4548 // operand, where it can often be eliminated completely.
4549 Op0 = CI->getOperand(0);
4551 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
4552 // so eliminate it as well.
4553 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
4554 Op1 = CI2->getOperand(0);
4556 // If Op1 is a constant, we can fold the cast into the constant.
4557 if (Op0->getType() != Op1->getType()) {
4558 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4559 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4561 // Otherwise, cast the RHS right before the icmp
4562 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
4565 return new ICmpInst(I.getPredicate(), Op0, Op1);
4569 if (isa<CastInst>(Op0)) {
4570 // Handle the special case of: icmp (cast bool to X), <cst>
4571 // This comes up when you have code like
4574 // For generality, we handle any zero-extension of any operand comparison
4575 // with a constant or another cast from the same type.
4576 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
4577 if (Instruction *R = foldICmpWithCastAndCast(I))
4581 if (Instruction *Res = foldICmpBinOp(I))
4584 if (Instruction *Res = foldICmpWithMinMax(I))
4589 // Transform (A & ~B) == 0 --> (A & B) != 0
4590 // and (A & ~B) != 0 --> (A & B) == 0
4591 // if A is a power of 2.
4592 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
4593 match(Op1, m_Zero()) &&
4594 isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
4595 return new ICmpInst(I.getInversePredicate(),
4596 Builder->CreateAnd(A, B),
4599 // ~X < ~Y --> Y < X
4600 // ~X < C --> X > ~C
4601 if (match(Op0, m_Not(m_Value(A)))) {
4602 if (match(Op1, m_Not(m_Value(B))))
4603 return new ICmpInst(I.getPredicate(), B, A);
4606 if (match(Op1, m_APInt(C)))
4607 return new ICmpInst(I.getSwappedPredicate(), A,
4608 ConstantInt::get(Op1->getType(), ~(*C)));
4611 Instruction *AddI = nullptr;
4612 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
4613 m_Instruction(AddI))) &&
4614 isa<IntegerType>(A->getType())) {
4617 if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result,
4619 replaceInstUsesWith(*AddI, Result);
4620 return replaceInstUsesWith(I, Overflow);
4624 // (zext a) * (zext b) --> llvm.umul.with.overflow.
4625 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4626 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
4629 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4630 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
4635 if (Instruction *Res = foldICmpEquality(I))
4638 // The 'cmpxchg' instruction returns an aggregate containing the old value and
4639 // an i1 which indicates whether or not we successfully did the swap.
4641 // Replace comparisons between the old value and the expected value with the
4642 // indicator that 'cmpxchg' returns.
4644 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
4645 // spuriously fail. In those cases, the old value may equal the expected
4646 // value but it is possible for the swap to not occur.
4647 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4648 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
4649 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
4650 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
4652 return ExtractValueInst::Create(ACXI, 1);
4655 Value *X; ConstantInt *Cst;
4657 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
4658 return foldICmpAddOpConst(I, X, Cst, I.getPredicate());
4661 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
4662 return foldICmpAddOpConst(I, X, Cst, I.getSwappedPredicate());
4664 return Changed ? &I : nullptr;
4667 /// Fold fcmp ([us]itofp x, cst) if possible.
4668 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
4670 if (!isa<ConstantFP>(RHSC)) return nullptr;
4671 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4673 // Get the width of the mantissa. We don't want to hack on conversions that
4674 // might lose information from the integer, e.g. "i64 -> float"
4675 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4676 if (MantissaWidth == -1) return nullptr; // Unknown.
4678 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4680 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
4682 if (I.isEquality()) {
4683 FCmpInst::Predicate P = I.getPredicate();
4684 bool IsExact = false;
4685 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
4686 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
4688 // If the floating point constant isn't an integer value, we know if we will
4689 // ever compare equal / not equal to it.
4691 // TODO: Can never be -0.0 and other non-representable values
4692 APFloat RHSRoundInt(RHS);
4693 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
4694 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
4695 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
4696 return replaceInstUsesWith(I, Builder->getFalse());
4698 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
4699 return replaceInstUsesWith(I, Builder->getTrue());
4703 // TODO: If the constant is exactly representable, is it always OK to do
4704 // equality compares as integer?
4707 // Check to see that the input is converted from an integer type that is small
4708 // enough that preserves all bits. TODO: check here for "known" sign bits.
4709 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4710 unsigned InputSize = IntTy->getScalarSizeInBits();
4712 // Following test does NOT adjust InputSize downwards for signed inputs,
4713 // because the most negative value still requires all the mantissa bits
4714 // to distinguish it from one less than that value.
4715 if ((int)InputSize > MantissaWidth) {
4716 // Conversion would lose accuracy. Check if loss can impact comparison.
4717 int Exp = ilogb(RHS);
4718 if (Exp == APFloat::IEK_Inf) {
4719 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
4720 if (MaxExponent < (int)InputSize - !LHSUnsigned)
4721 // Conversion could create infinity.
4724 // Note that if RHS is zero or NaN, then Exp is negative
4725 // and first condition is trivially false.
4726 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
4727 // Conversion could affect comparison.
4732 // Otherwise, we can potentially simplify the comparison. We know that it
4733 // will always come through as an integer value and we know the constant is
4734 // not a NAN (it would have been previously simplified).
4735 assert(!RHS.isNaN() && "NaN comparison not already folded!");
4737 ICmpInst::Predicate Pred;
4738 switch (I.getPredicate()) {
4739 default: llvm_unreachable("Unexpected predicate!");
4740 case FCmpInst::FCMP_UEQ:
4741 case FCmpInst::FCMP_OEQ:
4742 Pred = ICmpInst::ICMP_EQ;
4744 case FCmpInst::FCMP_UGT:
4745 case FCmpInst::FCMP_OGT:
4746 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
4748 case FCmpInst::FCMP_UGE:
4749 case FCmpInst::FCMP_OGE:
4750 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
4752 case FCmpInst::FCMP_ULT:
4753 case FCmpInst::FCMP_OLT:
4754 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
4756 case FCmpInst::FCMP_ULE:
4757 case FCmpInst::FCMP_OLE:
4758 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
4760 case FCmpInst::FCMP_UNE:
4761 case FCmpInst::FCMP_ONE:
4762 Pred = ICmpInst::ICMP_NE;
4764 case FCmpInst::FCMP_ORD:
4765 return replaceInstUsesWith(I, Builder->getTrue());
4766 case FCmpInst::FCMP_UNO:
4767 return replaceInstUsesWith(I, Builder->getFalse());
4770 // Now we know that the APFloat is a normal number, zero or inf.
4772 // See if the FP constant is too large for the integer. For example,
4773 // comparing an i8 to 300.0.
4774 unsigned IntWidth = IntTy->getScalarSizeInBits();
4777 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4778 // and large values.
4779 APFloat SMax(RHS.getSemantics());
4780 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4781 APFloat::rmNearestTiesToEven);
4782 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4783 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4784 Pred == ICmpInst::ICMP_SLE)
4785 return replaceInstUsesWith(I, Builder->getTrue());
4786 return replaceInstUsesWith(I, Builder->getFalse());
4789 // If the RHS value is > UnsignedMax, fold the comparison. This handles
4790 // +INF and large values.
4791 APFloat UMax(RHS.getSemantics());
4792 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
4793 APFloat::rmNearestTiesToEven);
4794 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
4795 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
4796 Pred == ICmpInst::ICMP_ULE)
4797 return replaceInstUsesWith(I, Builder->getTrue());
4798 return replaceInstUsesWith(I, Builder->getFalse());
4803 // See if the RHS value is < SignedMin.
4804 APFloat SMin(RHS.getSemantics());
4805 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
4806 APFloat::rmNearestTiesToEven);
4807 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
4808 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
4809 Pred == ICmpInst::ICMP_SGE)
4810 return replaceInstUsesWith(I, Builder->getTrue());
4811 return replaceInstUsesWith(I, Builder->getFalse());
4814 // See if the RHS value is < UnsignedMin.
4815 APFloat SMin(RHS.getSemantics());
4816 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
4817 APFloat::rmNearestTiesToEven);
4818 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
4819 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
4820 Pred == ICmpInst::ICMP_UGE)
4821 return replaceInstUsesWith(I, Builder->getTrue());
4822 return replaceInstUsesWith(I, Builder->getFalse());
4826 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
4827 // [0, UMAX], but it may still be fractional. See if it is fractional by
4828 // casting the FP value to the integer value and back, checking for equality.
4829 // Don't do this for zero, because -0.0 is not fractional.
4830 Constant *RHSInt = LHSUnsigned
4831 ? ConstantExpr::getFPToUI(RHSC, IntTy)
4832 : ConstantExpr::getFPToSI(RHSC, IntTy);
4833 if (!RHS.isZero()) {
4834 bool Equal = LHSUnsigned
4835 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
4836 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
4838 // If we had a comparison against a fractional value, we have to adjust
4839 // the compare predicate and sometimes the value. RHSC is rounded towards
4840 // zero at this point.
4842 default: llvm_unreachable("Unexpected integer comparison!");
4843 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
4844 return replaceInstUsesWith(I, Builder->getTrue());
4845 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
4846 return replaceInstUsesWith(I, Builder->getFalse());
4847 case ICmpInst::ICMP_ULE:
4848 // (float)int <= 4.4 --> int <= 4
4849 // (float)int <= -4.4 --> false
4850 if (RHS.isNegative())
4851 return replaceInstUsesWith(I, Builder->getFalse());
4853 case ICmpInst::ICMP_SLE:
4854 // (float)int <= 4.4 --> int <= 4
4855 // (float)int <= -4.4 --> int < -4
4856 if (RHS.isNegative())
4857 Pred = ICmpInst::ICMP_SLT;
4859 case ICmpInst::ICMP_ULT:
4860 // (float)int < -4.4 --> false
4861 // (float)int < 4.4 --> int <= 4
4862 if (RHS.isNegative())
4863 return replaceInstUsesWith(I, Builder->getFalse());
4864 Pred = ICmpInst::ICMP_ULE;
4866 case ICmpInst::ICMP_SLT:
4867 // (float)int < -4.4 --> int < -4
4868 // (float)int < 4.4 --> int <= 4
4869 if (!RHS.isNegative())
4870 Pred = ICmpInst::ICMP_SLE;
4872 case ICmpInst::ICMP_UGT:
4873 // (float)int > 4.4 --> int > 4
4874 // (float)int > -4.4 --> true
4875 if (RHS.isNegative())
4876 return replaceInstUsesWith(I, Builder->getTrue());
4878 case ICmpInst::ICMP_SGT:
4879 // (float)int > 4.4 --> int > 4
4880 // (float)int > -4.4 --> int >= -4
4881 if (RHS.isNegative())
4882 Pred = ICmpInst::ICMP_SGE;
4884 case ICmpInst::ICMP_UGE:
4885 // (float)int >= -4.4 --> true
4886 // (float)int >= 4.4 --> int > 4
4887 if (RHS.isNegative())
4888 return replaceInstUsesWith(I, Builder->getTrue());
4889 Pred = ICmpInst::ICMP_UGT;
4891 case ICmpInst::ICMP_SGE:
4892 // (float)int >= -4.4 --> int >= -4
4893 // (float)int >= 4.4 --> int > 4
4894 if (!RHS.isNegative())
4895 Pred = ICmpInst::ICMP_SGT;
4901 // Lower this FP comparison into an appropriate integer version of the
4903 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
4906 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4907 bool Changed = false;
4909 /// Orders the operands of the compare so that they are listed from most
4910 /// complex to least complex. This puts constants before unary operators,
4911 /// before binary operators.
4912 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
4917 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4920 SimplifyFCmpInst(I.getPredicate(), Op0, Op1, I.getFastMathFlags(),
4921 SQ.getWithInstruction(&I)))
4922 return replaceInstUsesWith(I, V);
4924 // Simplify 'fcmp pred X, X'
4926 switch (I.getPredicate()) {
4927 default: llvm_unreachable("Unknown predicate!");
4928 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4929 case FCmpInst::FCMP_ULT: // True if unordered or less than
4930 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4931 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4932 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4933 I.setPredicate(FCmpInst::FCMP_UNO);
4934 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4937 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4938 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4939 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4940 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4941 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4942 I.setPredicate(FCmpInst::FCMP_ORD);
4943 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4948 // Test if the FCmpInst instruction is used exclusively by a select as
4949 // part of a minimum or maximum operation. If so, refrain from doing
4950 // any other folding. This helps out other analyses which understand
4951 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4952 // and CodeGen. And in this case, at least one of the comparison
4953 // operands has at least one user besides the compare (the select),
4954 // which would often largely negate the benefit of folding anyway.
4956 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
4957 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
4958 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
4961 // Handle fcmp with constant RHS
4962 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4963 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4964 switch (LHSI->getOpcode()) {
4965 case Instruction::FPExt: {
4966 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
4967 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
4968 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
4972 const fltSemantics *Sem;
4973 // FIXME: This shouldn't be here.
4974 if (LHSExt->getSrcTy()->isHalfTy())
4975 Sem = &APFloat::IEEEhalf();
4976 else if (LHSExt->getSrcTy()->isFloatTy())
4977 Sem = &APFloat::IEEEsingle();
4978 else if (LHSExt->getSrcTy()->isDoubleTy())
4979 Sem = &APFloat::IEEEdouble();
4980 else if (LHSExt->getSrcTy()->isFP128Ty())
4981 Sem = &APFloat::IEEEquad();
4982 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
4983 Sem = &APFloat::x87DoubleExtended();
4984 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
4985 Sem = &APFloat::PPCDoubleDouble();
4990 APFloat F = RHSF->getValueAPF();
4991 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
4993 // Avoid lossy conversions and denormals. Zero is a special case
4994 // that's OK to convert.
4998 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
4999 APFloat::cmpLessThan) || Fabs.isZero()))
5001 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
5002 ConstantFP::get(RHSC->getContext(), F));
5005 case Instruction::PHI:
5006 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5007 // block. If in the same block, we're encouraging jump threading. If
5008 // not, we are just pessimizing the code by making an i1 phi.
5009 if (LHSI->getParent() == I.getParent())
5010 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
5013 case Instruction::SIToFP:
5014 case Instruction::UIToFP:
5015 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
5018 case Instruction::FSub: {
5019 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
5021 if (match(LHSI, m_FNeg(m_Value(Op))))
5022 return new FCmpInst(I.getSwappedPredicate(), Op,
5023 ConstantExpr::getFNeg(RHSC));
5026 case Instruction::Load:
5027 if (GetElementPtrInst *GEP =
5028 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
5029 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
5030 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
5031 !cast<LoadInst>(LHSI)->isVolatile())
5032 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
5036 case Instruction::Call: {
5037 if (!RHSC->isNullValue())
5040 CallInst *CI = cast<CallInst>(LHSI);
5041 Intrinsic::ID IID = getIntrinsicForCallSite(CI, &TLI);
5042 if (IID != Intrinsic::fabs)
5045 // Various optimization for fabs compared with zero.
5046 switch (I.getPredicate()) {
5049 // fabs(x) < 0 --> false
5050 case FCmpInst::FCMP_OLT:
5051 llvm_unreachable("handled by SimplifyFCmpInst");
5052 // fabs(x) > 0 --> x != 0
5053 case FCmpInst::FCMP_OGT:
5054 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), RHSC);
5055 // fabs(x) <= 0 --> x == 0
5056 case FCmpInst::FCMP_OLE:
5057 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), RHSC);
5058 // fabs(x) >= 0 --> !isnan(x)
5059 case FCmpInst::FCMP_OGE:
5060 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), RHSC);
5061 // fabs(x) == 0 --> x == 0
5062 // fabs(x) != 0 --> x != 0
5063 case FCmpInst::FCMP_OEQ:
5064 case FCmpInst::FCMP_UEQ:
5065 case FCmpInst::FCMP_ONE:
5066 case FCmpInst::FCMP_UNE:
5067 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), RHSC);
5073 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
5075 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
5076 return new FCmpInst(I.getSwappedPredicate(), X, Y);
5078 // fcmp (fpext x), (fpext y) -> fcmp x, y
5079 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
5080 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
5081 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
5082 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
5083 RHSExt->getOperand(0));
5085 return Changed ? &I : nullptr;