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"
31 using namespace PatternMatch;
33 #define DEBUG_TYPE "instcombine"
35 // How many times is a select replaced by one of its operands?
36 STATISTIC(NumSel, "Number of select opts");
39 static ConstantInt *extractElement(Constant *V, Constant *Idx) {
40 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
43 static bool hasAddOverflow(ConstantInt *Result,
44 ConstantInt *In1, ConstantInt *In2,
47 return Result->getValue().ult(In1->getValue());
49 if (In2->isNegative())
50 return Result->getValue().sgt(In1->getValue());
51 return Result->getValue().slt(In1->getValue());
54 /// Compute Result = In1+In2, returning true if the result overflowed for this
56 static bool addWithOverflow(Constant *&Result, Constant *In1,
57 Constant *In2, bool IsSigned = false) {
58 Result = ConstantExpr::getAdd(In1, In2);
60 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
61 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
62 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
63 if (hasAddOverflow(extractElement(Result, Idx),
64 extractElement(In1, Idx),
65 extractElement(In2, Idx),
72 return hasAddOverflow(cast<ConstantInt>(Result),
73 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
77 static bool hasSubOverflow(ConstantInt *Result,
78 ConstantInt *In1, ConstantInt *In2,
81 return Result->getValue().ugt(In1->getValue());
83 if (In2->isNegative())
84 return Result->getValue().slt(In1->getValue());
86 return Result->getValue().sgt(In1->getValue());
89 /// Compute Result = In1-In2, returning true if the result overflowed for this
91 static bool subWithOverflow(Constant *&Result, Constant *In1,
92 Constant *In2, bool IsSigned = false) {
93 Result = ConstantExpr::getSub(In1, In2);
95 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
96 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
97 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
98 if (hasSubOverflow(extractElement(Result, Idx),
99 extractElement(In1, Idx),
100 extractElement(In2, Idx),
107 return hasSubOverflow(cast<ConstantInt>(Result),
108 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
112 /// Given an icmp instruction, return true if any use of this comparison is a
113 /// branch on sign bit comparison.
114 static bool isBranchOnSignBitCheck(ICmpInst &I, bool isSignBit) {
115 for (auto *U : I.users())
116 if (isa<BranchInst>(U))
121 /// Given an exploded icmp instruction, return true if the comparison only
122 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
123 /// result of the comparison is true when the input value is signed.
124 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
125 bool &TrueIfSigned) {
127 case ICmpInst::ICMP_SLT: // True if LHS s< 0
130 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
132 return RHS.isAllOnesValue();
133 case ICmpInst::ICMP_SGT: // True if LHS s> -1
134 TrueIfSigned = false;
135 return RHS.isAllOnesValue();
136 case ICmpInst::ICMP_UGT:
137 // True if LHS u> RHS and RHS == high-bit-mask - 1
139 return RHS.isMaxSignedValue();
140 case ICmpInst::ICMP_UGE:
141 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
143 return RHS.isSignBit();
149 /// Returns true if the exploded icmp can be expressed as a signed comparison
150 /// to zero and updates the predicate accordingly.
151 /// The signedness of the comparison is preserved.
152 /// TODO: Refactor with decomposeBitTestICmp()?
153 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
154 if (!ICmpInst::isSigned(Pred))
158 return ICmpInst::isRelational(Pred);
161 if (Pred == ICmpInst::ICMP_SLT) {
162 Pred = ICmpInst::ICMP_SLE;
165 } else if (C.isAllOnesValue()) {
166 if (Pred == ICmpInst::ICMP_SGT) {
167 Pred = ICmpInst::ICMP_SGE;
175 /// Given a signed integer type and a set of known zero and one bits, compute
176 /// the maximum and minimum values that could have the specified known zero and
177 /// known one bits, returning them in Min/Max.
178 static void computeSignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
179 const APInt &KnownOne,
180 APInt &Min, APInt &Max) {
181 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
182 KnownZero.getBitWidth() == Min.getBitWidth() &&
183 KnownZero.getBitWidth() == Max.getBitWidth() &&
184 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
185 APInt UnknownBits = ~(KnownZero|KnownOne);
187 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
188 // bit if it is unknown.
190 Max = KnownOne|UnknownBits;
192 if (UnknownBits.isNegative()) { // Sign bit is unknown
193 Min.setBit(Min.getBitWidth()-1);
194 Max.clearBit(Max.getBitWidth()-1);
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 static void computeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
202 const APInt &KnownOne,
203 APInt &Min, APInt &Max) {
204 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
205 KnownZero.getBitWidth() == Min.getBitWidth() &&
206 KnownZero.getBitWidth() == Max.getBitWidth() &&
207 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
208 APInt UnknownBits = ~(KnownZero|KnownOne);
210 // The minimum value is when the unknown bits are all zeros.
212 // The maximum value is when the unknown bits are all ones.
213 Max = KnownOne|UnknownBits;
216 /// This is called when we see this pattern:
217 /// cmp pred (load (gep GV, ...)), cmpcst
218 /// where GV is a global variable with a constant initializer. Try to simplify
219 /// this into some simple computation that does not need the load. For example
220 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
222 /// If AndCst is non-null, then the loaded value is masked with that constant
223 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
224 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
227 ConstantInt *AndCst) {
228 Constant *Init = GV->getInitializer();
229 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
232 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
233 if (ArrayElementCount > 1024) return nullptr; // Don't blow up on huge arrays.
235 // There are many forms of this optimization we can handle, for now, just do
236 // the simple index into a single-dimensional array.
238 // Require: GEP GV, 0, i {{, constant indices}}
239 if (GEP->getNumOperands() < 3 ||
240 !isa<ConstantInt>(GEP->getOperand(1)) ||
241 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
242 isa<Constant>(GEP->getOperand(2)))
245 // Check that indices after the variable are constants and in-range for the
246 // type they index. Collect the indices. This is typically for arrays of
248 SmallVector<unsigned, 4> LaterIndices;
250 Type *EltTy = Init->getType()->getArrayElementType();
251 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
252 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
253 if (!Idx) return nullptr; // Variable index.
255 uint64_t IdxVal = Idx->getZExtValue();
256 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
258 if (StructType *STy = dyn_cast<StructType>(EltTy))
259 EltTy = STy->getElementType(IdxVal);
260 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
261 if (IdxVal >= ATy->getNumElements()) return nullptr;
262 EltTy = ATy->getElementType();
264 return nullptr; // Unknown type.
267 LaterIndices.push_back(IdxVal);
270 enum { Overdefined = -3, Undefined = -2 };
272 // Variables for our state machines.
274 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
275 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
276 // and 87 is the second (and last) index. FirstTrueElement is -2 when
277 // undefined, otherwise set to the first true element. SecondTrueElement is
278 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
279 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
281 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
282 // form "i != 47 & i != 87". Same state transitions as for true elements.
283 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
285 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
286 /// define a state machine that triggers for ranges of values that the index
287 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
288 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
289 /// index in the range (inclusive). We use -2 for undefined here because we
290 /// use relative comparisons and don't want 0-1 to match -1.
291 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
293 // MagicBitvector - This is a magic bitvector where we set a bit if the
294 // comparison is true for element 'i'. If there are 64 elements or less in
295 // the array, this will fully represent all the comparison results.
296 uint64_t MagicBitvector = 0;
298 // Scan the array and see if one of our patterns matches.
299 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
300 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
301 Constant *Elt = Init->getAggregateElement(i);
302 if (!Elt) return nullptr;
304 // If this is indexing an array of structures, get the structure element.
305 if (!LaterIndices.empty())
306 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
308 // If the element is masked, handle it.
309 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
311 // Find out if the comparison would be true or false for the i'th element.
312 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
313 CompareRHS, DL, &TLI);
314 // If the result is undef for this element, ignore it.
315 if (isa<UndefValue>(C)) {
316 // Extend range state machines to cover this element in case there is an
317 // undef in the middle of the range.
318 if (TrueRangeEnd == (int)i-1)
320 if (FalseRangeEnd == (int)i-1)
325 // If we can't compute the result for any of the elements, we have to give
326 // up evaluating the entire conditional.
327 if (!isa<ConstantInt>(C)) return nullptr;
329 // Otherwise, we know if the comparison is true or false for this element,
330 // update our state machines.
331 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
333 // State machine for single/double/range index comparison.
335 // Update the TrueElement state machine.
336 if (FirstTrueElement == Undefined)
337 FirstTrueElement = TrueRangeEnd = i; // First true element.
339 // Update double-compare state machine.
340 if (SecondTrueElement == Undefined)
341 SecondTrueElement = i;
343 SecondTrueElement = Overdefined;
345 // Update range state machine.
346 if (TrueRangeEnd == (int)i-1)
349 TrueRangeEnd = Overdefined;
352 // Update the FalseElement state machine.
353 if (FirstFalseElement == Undefined)
354 FirstFalseElement = FalseRangeEnd = i; // First false element.
356 // Update double-compare state machine.
357 if (SecondFalseElement == Undefined)
358 SecondFalseElement = i;
360 SecondFalseElement = Overdefined;
362 // Update range state machine.
363 if (FalseRangeEnd == (int)i-1)
366 FalseRangeEnd = Overdefined;
370 // If this element is in range, update our magic bitvector.
371 if (i < 64 && IsTrueForElt)
372 MagicBitvector |= 1ULL << i;
374 // If all of our states become overdefined, bail out early. Since the
375 // predicate is expensive, only check it every 8 elements. This is only
376 // really useful for really huge arrays.
377 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
378 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
379 FalseRangeEnd == Overdefined)
383 // Now that we've scanned the entire array, emit our new comparison(s). We
384 // order the state machines in complexity of the generated code.
385 Value *Idx = GEP->getOperand(2);
387 // If the index is larger than the pointer size of the target, truncate the
388 // index down like the GEP would do implicitly. We don't have to do this for
389 // an inbounds GEP because the index can't be out of range.
390 if (!GEP->isInBounds()) {
391 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
392 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
393 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
394 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
397 // If the comparison is only true for one or two elements, emit direct
399 if (SecondTrueElement != Overdefined) {
400 // None true -> false.
401 if (FirstTrueElement == Undefined)
402 return replaceInstUsesWith(ICI, Builder->getFalse());
404 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
406 // True for one element -> 'i == 47'.
407 if (SecondTrueElement == Undefined)
408 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
410 // True for two elements -> 'i == 47 | i == 72'.
411 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
412 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
413 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
414 return BinaryOperator::CreateOr(C1, C2);
417 // If the comparison is only false for one or two elements, emit direct
419 if (SecondFalseElement != Overdefined) {
420 // None false -> true.
421 if (FirstFalseElement == Undefined)
422 return replaceInstUsesWith(ICI, Builder->getTrue());
424 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
426 // False for one element -> 'i != 47'.
427 if (SecondFalseElement == Undefined)
428 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
430 // False for two elements -> 'i != 47 & i != 72'.
431 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
432 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
433 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
434 return BinaryOperator::CreateAnd(C1, C2);
437 // If the comparison can be replaced with a range comparison for the elements
438 // where it is true, emit the range check.
439 if (TrueRangeEnd != Overdefined) {
440 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
442 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
443 if (FirstTrueElement) {
444 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
445 Idx = Builder->CreateAdd(Idx, Offs);
448 Value *End = ConstantInt::get(Idx->getType(),
449 TrueRangeEnd-FirstTrueElement+1);
450 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
453 // False range check.
454 if (FalseRangeEnd != Overdefined) {
455 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
456 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
457 if (FirstFalseElement) {
458 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
459 Idx = Builder->CreateAdd(Idx, Offs);
462 Value *End = ConstantInt::get(Idx->getType(),
463 FalseRangeEnd-FirstFalseElement);
464 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
467 // If a magic bitvector captures the entire comparison state
468 // of this load, replace it with computation that does:
469 // ((magic_cst >> i) & 1) != 0
473 // Look for an appropriate type:
474 // - The type of Idx if the magic fits
475 // - The smallest fitting legal type if we have a DataLayout
477 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
480 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
483 Value *V = Builder->CreateIntCast(Idx, Ty, false);
484 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
485 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
486 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
493 /// Return a value that can be used to compare the *offset* implied by a GEP to
494 /// zero. For example, if we have &A[i], we want to return 'i' for
495 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
496 /// are involved. The above expression would also be legal to codegen as
497 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
498 /// This latter form is less amenable to optimization though, and we are allowed
499 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
501 /// If we can't emit an optimized form for this expression, this returns null.
503 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC,
504 const DataLayout &DL) {
505 gep_type_iterator GTI = gep_type_begin(GEP);
507 // Check to see if this gep only has a single variable index. If so, and if
508 // any constant indices are a multiple of its scale, then we can compute this
509 // in terms of the scale of the variable index. For example, if the GEP
510 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
511 // because the expression will cross zero at the same point.
512 unsigned i, e = GEP->getNumOperands();
514 for (i = 1; i != e; ++i, ++GTI) {
515 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
516 // Compute the aggregate offset of constant indices.
517 if (CI->isZero()) continue;
519 // Handle a struct index, which adds its field offset to the pointer.
520 if (StructType *STy = GTI.getStructTypeOrNull()) {
521 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
523 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
524 Offset += Size*CI->getSExtValue();
527 // Found our variable index.
532 // If there are no variable indices, we must have a constant offset, just
533 // evaluate it the general way.
534 if (i == e) return nullptr;
536 Value *VariableIdx = GEP->getOperand(i);
537 // Determine the scale factor of the variable element. For example, this is
538 // 4 if the variable index is into an array of i32.
539 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
541 // Verify that there are no other variable indices. If so, emit the hard way.
542 for (++i, ++GTI; i != e; ++i, ++GTI) {
543 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
544 if (!CI) return nullptr;
546 // Compute the aggregate offset of constant indices.
547 if (CI->isZero()) continue;
549 // Handle a struct index, which adds its field offset to the pointer.
550 if (StructType *STy = GTI.getStructTypeOrNull()) {
551 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
553 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
554 Offset += Size*CI->getSExtValue();
558 // Okay, we know we have a single variable index, which must be a
559 // pointer/array/vector index. If there is no offset, life is simple, return
561 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
562 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
564 // Cast to intptrty in case a truncation occurs. If an extension is needed,
565 // we don't need to bother extending: the extension won't affect where the
566 // computation crosses zero.
567 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
568 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
573 // Otherwise, there is an index. The computation we will do will be modulo
574 // the pointer size, so get it.
575 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
577 Offset &= PtrSizeMask;
578 VariableScale &= PtrSizeMask;
580 // To do this transformation, any constant index must be a multiple of the
581 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
582 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
583 // multiple of the variable scale.
584 int64_t NewOffs = Offset / (int64_t)VariableScale;
585 if (Offset != NewOffs*(int64_t)VariableScale)
588 // Okay, we can do this evaluation. Start by converting the index to intptr.
589 if (VariableIdx->getType() != IntPtrTy)
590 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
592 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
593 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
596 /// Returns true if we can rewrite Start as a GEP with pointer Base
597 /// and some integer offset. The nodes that need to be re-written
598 /// for this transformation will be added to Explored.
599 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
600 const DataLayout &DL,
601 SetVector<Value *> &Explored) {
602 SmallVector<Value *, 16> WorkList(1, Start);
603 Explored.insert(Base);
605 // The following traversal gives us an order which can be used
606 // when doing the final transformation. Since in the final
607 // transformation we create the PHI replacement instructions first,
608 // we don't have to get them in any particular order.
610 // However, for other instructions we will have to traverse the
611 // operands of an instruction first, which means that we have to
612 // do a post-order traversal.
613 while (!WorkList.empty()) {
614 SetVector<PHINode *> PHIs;
616 while (!WorkList.empty()) {
617 if (Explored.size() >= 100)
620 Value *V = WorkList.back();
622 if (Explored.count(V) != 0) {
627 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
628 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
629 // We've found some value that we can't explore which is different from
630 // the base. Therefore we can't do this transformation.
633 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
634 auto *CI = dyn_cast<CastInst>(V);
635 if (!CI->isNoopCast(DL))
638 if (Explored.count(CI->getOperand(0)) == 0)
639 WorkList.push_back(CI->getOperand(0));
642 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
643 // We're limiting the GEP to having one index. This will preserve
644 // the original pointer type. We could handle more cases in the
646 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
647 GEP->getType() != Start->getType())
650 if (Explored.count(GEP->getOperand(0)) == 0)
651 WorkList.push_back(GEP->getOperand(0));
654 if (WorkList.back() == V) {
656 // We've finished visiting this node, mark it as such.
660 if (auto *PN = dyn_cast<PHINode>(V)) {
661 // We cannot transform PHIs on unsplittable basic blocks.
662 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
669 // Explore the PHI nodes further.
670 for (auto *PN : PHIs)
671 for (Value *Op : PN->incoming_values())
672 if (Explored.count(Op) == 0)
673 WorkList.push_back(Op);
676 // Make sure that we can do this. Since we can't insert GEPs in a basic
677 // block before a PHI node, we can't easily do this transformation if
678 // we have PHI node users of transformed instructions.
679 for (Value *Val : Explored) {
680 for (Value *Use : Val->uses()) {
682 auto *PHI = dyn_cast<PHINode>(Use);
683 auto *Inst = dyn_cast<Instruction>(Val);
685 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
686 Explored.count(PHI) == 0)
689 if (PHI->getParent() == Inst->getParent())
696 // Sets the appropriate insert point on Builder where we can add
697 // a replacement Instruction for V (if that is possible).
698 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
699 bool Before = true) {
700 if (auto *PHI = dyn_cast<PHINode>(V)) {
701 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
704 if (auto *I = dyn_cast<Instruction>(V)) {
706 I = &*std::next(I->getIterator());
707 Builder.SetInsertPoint(I);
710 if (auto *A = dyn_cast<Argument>(V)) {
711 // Set the insertion point in the entry block.
712 BasicBlock &Entry = A->getParent()->getEntryBlock();
713 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
716 // Otherwise, this is a constant and we don't need to set a new
718 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
721 /// Returns a re-written value of Start as an indexed GEP using Base as a
723 static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
724 const DataLayout &DL,
725 SetVector<Value *> &Explored) {
726 // Perform all the substitutions. This is a bit tricky because we can
727 // have cycles in our use-def chains.
728 // 1. Create the PHI nodes without any incoming values.
729 // 2. Create all the other values.
730 // 3. Add the edges for the PHI nodes.
731 // 4. Emit GEPs to get the original pointers.
732 // 5. Remove the original instructions.
733 Type *IndexType = IntegerType::get(
734 Base->getContext(), DL.getPointerTypeSizeInBits(Start->getType()));
736 DenseMap<Value *, Value *> NewInsts;
737 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
739 // Create the new PHI nodes, without adding any incoming values.
740 for (Value *Val : Explored) {
743 // Create empty phi nodes. This avoids cyclic dependencies when creating
744 // the remaining instructions.
745 if (auto *PHI = dyn_cast<PHINode>(Val))
746 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
747 PHI->getName() + ".idx", PHI);
749 IRBuilder<> Builder(Base->getContext());
751 // Create all the other instructions.
752 for (Value *Val : Explored) {
754 if (NewInsts.find(Val) != NewInsts.end())
757 if (auto *CI = dyn_cast<CastInst>(Val)) {
758 NewInsts[CI] = NewInsts[CI->getOperand(0)];
761 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
762 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
763 : GEP->getOperand(1);
764 setInsertionPoint(Builder, GEP);
765 // Indices might need to be sign extended. GEPs will magically do
766 // this, but we need to do it ourselves here.
767 if (Index->getType()->getScalarSizeInBits() !=
768 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
769 Index = Builder.CreateSExtOrTrunc(
770 Index, NewInsts[GEP->getOperand(0)]->getType(),
771 GEP->getOperand(0)->getName() + ".sext");
774 auto *Op = NewInsts[GEP->getOperand(0)];
775 if (isa<ConstantInt>(Op) && dyn_cast<ConstantInt>(Op)->isZero())
776 NewInsts[GEP] = Index;
778 NewInsts[GEP] = Builder.CreateNSWAdd(
779 Op, Index, GEP->getOperand(0)->getName() + ".add");
782 if (isa<PHINode>(Val))
785 llvm_unreachable("Unexpected instruction type");
788 // Add the incoming values to the PHI nodes.
789 for (Value *Val : Explored) {
792 // All the instructions have been created, we can now add edges to the
794 if (auto *PHI = dyn_cast<PHINode>(Val)) {
795 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
796 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
797 Value *NewIncoming = PHI->getIncomingValue(I);
799 if (NewInsts.find(NewIncoming) != NewInsts.end())
800 NewIncoming = NewInsts[NewIncoming];
802 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
807 for (Value *Val : Explored) {
811 // Depending on the type, for external users we have to emit
812 // a GEP or a GEP + ptrtoint.
813 setInsertionPoint(Builder, Val, false);
815 // If required, create an inttoptr instruction for Base.
816 Value *NewBase = Base;
817 if (!Base->getType()->isPointerTy())
818 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
819 Start->getName() + "to.ptr");
821 Value *GEP = Builder.CreateInBoundsGEP(
822 Start->getType()->getPointerElementType(), NewBase,
823 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
825 if (!Val->getType()->isPointerTy()) {
826 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
827 Val->getName() + ".conv");
830 Val->replaceAllUsesWith(GEP);
833 return NewInsts[Start];
836 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
837 /// the input Value as a constant indexed GEP. Returns a pair containing
838 /// the GEPs Pointer and Index.
839 static std::pair<Value *, Value *>
840 getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
841 Type *IndexType = IntegerType::get(V->getContext(),
842 DL.getPointerTypeSizeInBits(V->getType()));
844 Constant *Index = ConstantInt::getNullValue(IndexType);
846 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
847 // We accept only inbouds GEPs here to exclude the possibility of
849 if (!GEP->isInBounds())
851 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
852 GEP->getType() == V->getType()) {
853 V = GEP->getOperand(0);
854 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
855 Index = ConstantExpr::getAdd(
856 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
861 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
862 if (!CI->isNoopCast(DL))
864 V = CI->getOperand(0);
867 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
868 if (!CI->isNoopCast(DL))
870 V = CI->getOperand(0);
878 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
879 /// We can look through PHIs, GEPs and casts in order to determine a common base
880 /// between GEPLHS and RHS.
881 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
882 ICmpInst::Predicate Cond,
883 const DataLayout &DL) {
884 if (!GEPLHS->hasAllConstantIndices())
887 // Make sure the pointers have the same type.
888 if (GEPLHS->getType() != RHS->getType())
891 Value *PtrBase, *Index;
892 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
894 // The set of nodes that will take part in this transformation.
895 SetVector<Value *> Nodes;
897 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
900 // We know we can re-write this as
901 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
902 // Since we've only looked through inbouds GEPs we know that we
903 // can't have overflow on either side. We can therefore re-write
905 // OFFSET1 cmp OFFSET2
906 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
908 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
909 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
910 // offset. Since Index is the offset of LHS to the base pointer, we will now
911 // compare the offsets instead of comparing the pointers.
912 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
915 /// Fold comparisons between a GEP instruction and something else. At this point
916 /// we know that the GEP is on the LHS of the comparison.
917 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
918 ICmpInst::Predicate Cond,
920 // Don't transform signed compares of GEPs into index compares. Even if the
921 // GEP is inbounds, the final add of the base pointer can have signed overflow
922 // and would change the result of the icmp.
923 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
924 // the maximum signed value for the pointer type.
925 if (ICmpInst::isSigned(Cond))
928 // Look through bitcasts and addrspacecasts. We do not however want to remove
930 if (!isa<GetElementPtrInst>(RHS))
931 RHS = RHS->stripPointerCasts();
933 Value *PtrBase = GEPLHS->getOperand(0);
934 if (PtrBase == RHS && GEPLHS->isInBounds()) {
935 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
936 // This transformation (ignoring the base and scales) is valid because we
937 // know pointers can't overflow since the gep is inbounds. See if we can
938 // output an optimized form.
939 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
941 // If not, synthesize the offset the hard way.
943 Offset = EmitGEPOffset(GEPLHS);
944 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
945 Constant::getNullValue(Offset->getType()));
946 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
947 // If the base pointers are different, but the indices are the same, just
948 // compare the base pointer.
949 if (PtrBase != GEPRHS->getOperand(0)) {
950 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
951 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
952 GEPRHS->getOperand(0)->getType();
954 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
955 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
956 IndicesTheSame = false;
960 // If all indices are the same, just compare the base pointers.
962 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
964 // If we're comparing GEPs with two base pointers that only differ in type
965 // and both GEPs have only constant indices or just one use, then fold
966 // the compare with the adjusted indices.
967 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
968 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
969 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
970 PtrBase->stripPointerCasts() ==
971 GEPRHS->getOperand(0)->stripPointerCasts()) {
972 Value *LOffset = EmitGEPOffset(GEPLHS);
973 Value *ROffset = EmitGEPOffset(GEPRHS);
975 // If we looked through an addrspacecast between different sized address
976 // spaces, the LHS and RHS pointers are different sized
977 // integers. Truncate to the smaller one.
978 Type *LHSIndexTy = LOffset->getType();
979 Type *RHSIndexTy = ROffset->getType();
980 if (LHSIndexTy != RHSIndexTy) {
981 if (LHSIndexTy->getPrimitiveSizeInBits() <
982 RHSIndexTy->getPrimitiveSizeInBits()) {
983 ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy);
985 LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy);
988 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
990 return replaceInstUsesWith(I, Cmp);
993 // Otherwise, the base pointers are different and the indices are
994 // different. Try convert this to an indexed compare by looking through
996 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
999 // If one of the GEPs has all zero indices, recurse.
1000 if (GEPLHS->hasAllZeroIndices())
1001 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
1002 ICmpInst::getSwappedPredicate(Cond), I);
1004 // If the other GEP has all zero indices, recurse.
1005 if (GEPRHS->hasAllZeroIndices())
1006 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
1008 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
1009 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
1010 // If the GEPs only differ by one index, compare it.
1011 unsigned NumDifferences = 0; // Keep track of # differences.
1012 unsigned DiffOperand = 0; // The operand that differs.
1013 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
1014 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
1015 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
1016 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
1017 // Irreconcilable differences.
1021 if (NumDifferences++) break;
1026 if (NumDifferences == 0) // SAME GEP?
1027 return replaceInstUsesWith(I, // No comparison is needed here.
1028 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
1030 else if (NumDifferences == 1 && GEPsInBounds) {
1031 Value *LHSV = GEPLHS->getOperand(DiffOperand);
1032 Value *RHSV = GEPRHS->getOperand(DiffOperand);
1033 // Make sure we do a signed comparison here.
1034 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1038 // Only lower this if the icmp is the only user of the GEP or if we expect
1039 // the result to fold to a constant!
1040 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1041 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1042 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1043 Value *L = EmitGEPOffset(GEPLHS);
1044 Value *R = EmitGEPOffset(GEPRHS);
1045 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1049 // Try convert this to an indexed compare by looking through PHIs/casts as a
1051 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1054 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1055 const AllocaInst *Alloca,
1056 const Value *Other) {
1057 assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1059 // It would be tempting to fold away comparisons between allocas and any
1060 // pointer not based on that alloca (e.g. an argument). However, even
1061 // though such pointers cannot alias, they can still compare equal.
1063 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1064 // doesn't escape we can argue that it's impossible to guess its value, and we
1065 // can therefore act as if any such guesses are wrong.
1067 // The code below checks that the alloca doesn't escape, and that it's only
1068 // used in a comparison once (the current instruction). The
1069 // single-comparison-use condition ensures that we're trivially folding all
1070 // comparisons against the alloca consistently, and avoids the risk of
1071 // erroneously folding a comparison of the pointer with itself.
1073 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1075 SmallVector<const Use *, 32> Worklist;
1076 for (const Use &U : Alloca->uses()) {
1077 if (Worklist.size() >= MaxIter)
1079 Worklist.push_back(&U);
1082 unsigned NumCmps = 0;
1083 while (!Worklist.empty()) {
1084 assert(Worklist.size() <= MaxIter);
1085 const Use *U = Worklist.pop_back_val();
1086 const Value *V = U->getUser();
1089 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1090 isa<SelectInst>(V)) {
1092 } else if (isa<LoadInst>(V)) {
1093 // Loading from the pointer doesn't escape it.
1095 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1096 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1097 if (SI->getValueOperand() == U->get())
1100 } else if (isa<ICmpInst>(V)) {
1102 return nullptr; // Found more than one cmp.
1104 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1105 switch (Intrin->getIntrinsicID()) {
1106 // These intrinsics don't escape or compare the pointer. Memset is safe
1107 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1108 // we don't allow stores, so src cannot point to V.
1109 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1110 case Intrinsic::dbg_declare: case Intrinsic::dbg_value:
1111 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1119 for (const Use &U : V->uses()) {
1120 if (Worklist.size() >= MaxIter)
1122 Worklist.push_back(&U);
1126 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1127 return replaceInstUsesWith(
1129 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1132 /// Fold "icmp pred (X+CI), X".
1133 Instruction *InstCombiner::foldICmpAddOpConst(Instruction &ICI,
1134 Value *X, ConstantInt *CI,
1135 ICmpInst::Predicate Pred) {
1136 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1137 // so the values can never be equal. Similarly for all other "or equals"
1140 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1141 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1142 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1143 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1145 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
1146 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1149 // (X+1) >u X --> X <u (0-1) --> X != 255
1150 // (X+2) >u X --> X <u (0-2) --> X <u 254
1151 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1152 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1153 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
1155 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
1156 ConstantInt *SMax = ConstantInt::get(X->getContext(),
1157 APInt::getSignedMaxValue(BitWidth));
1159 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1160 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1161 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1162 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1163 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1164 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1165 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1166 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
1168 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1169 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1170 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1171 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1172 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1173 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1175 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1176 Constant *C = Builder->getInt(CI->getValue()-1);
1177 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
1180 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1181 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1182 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1183 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1186 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1188 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1189 if (I.getPredicate() == I.ICMP_NE)
1190 Pred = CmpInst::getInversePredicate(Pred);
1191 return new ICmpInst(Pred, LHS, RHS);
1194 // Don't bother doing any work for cases which InstSimplify handles.
1198 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1200 if (AP2.isAllOnesValue())
1202 if (AP2.isNegative() != AP1.isNegative())
1209 // 'A' must be large enough to shift out the highest set bit.
1210 return getICmp(I.ICMP_UGT, A,
1211 ConstantInt::get(A->getType(), AP2.logBase2()));
1214 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1217 if (IsAShr && AP1.isNegative())
1218 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1220 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1223 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1224 // There are multiple solutions if we are comparing against -1 and the LHS
1225 // of the ashr is not a power of two.
1226 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1227 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1228 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1229 } else if (AP1 == AP2.lshr(Shift)) {
1230 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1234 // Shifting const2 will never be equal to const1.
1235 // FIXME: This should always be handled by InstSimplify?
1236 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1237 return replaceInstUsesWith(I, TorF);
1240 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1241 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1242 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1245 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1247 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1248 if (I.getPredicate() == I.ICMP_NE)
1249 Pred = CmpInst::getInversePredicate(Pred);
1250 return new ICmpInst(Pred, LHS, RHS);
1253 // Don't bother doing any work for cases which InstSimplify handles.
1257 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1259 if (!AP1 && AP2TrailingZeros != 0)
1262 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1265 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1267 // Get the distance between the lowest bits that are set.
1268 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1270 if (Shift > 0 && AP2.shl(Shift) == AP1)
1271 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1273 // Shifting const2 will never be equal to const1.
1274 // FIXME: This should always be handled by InstSimplify?
1275 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1276 return replaceInstUsesWith(I, TorF);
1279 /// The caller has matched a pattern of the form:
1280 /// I = icmp ugt (add (add A, B), CI2), CI1
1281 /// If this is of the form:
1283 /// if (sum+128 >u 255)
1284 /// Then replace it with llvm.sadd.with.overflow.i8.
1286 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1287 ConstantInt *CI2, ConstantInt *CI1,
1289 // The transformation we're trying to do here is to transform this into an
1290 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1291 // with a narrower add, and discard the add-with-constant that is part of the
1292 // range check (if we can't eliminate it, this isn't profitable).
1294 // In order to eliminate the add-with-constant, the compare can be its only
1296 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1297 if (!AddWithCst->hasOneUse())
1300 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1301 if (!CI2->getValue().isPowerOf2())
1303 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1304 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1307 // The width of the new add formed is 1 more than the bias.
1310 // Check to see that CI1 is an all-ones value with NewWidth bits.
1311 if (CI1->getBitWidth() == NewWidth ||
1312 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1315 // This is only really a signed overflow check if the inputs have been
1316 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1317 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1318 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1319 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1320 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1323 // In order to replace the original add with a narrower
1324 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1325 // and truncates that discard the high bits of the add. Verify that this is
1327 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1328 for (User *U : OrigAdd->users()) {
1329 if (U == AddWithCst)
1332 // Only accept truncates for now. We would really like a nice recursive
1333 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1334 // chain to see which bits of a value are actually demanded. If the
1335 // original add had another add which was then immediately truncated, we
1336 // could still do the transformation.
1337 TruncInst *TI = dyn_cast<TruncInst>(U);
1338 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1342 // If the pattern matches, truncate the inputs to the narrower type and
1343 // use the sadd_with_overflow intrinsic to efficiently compute both the
1344 // result and the overflow bit.
1345 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1346 Value *F = Intrinsic::getDeclaration(I.getModule(),
1347 Intrinsic::sadd_with_overflow, NewType);
1349 InstCombiner::BuilderTy *Builder = IC.Builder;
1351 // Put the new code above the original add, in case there are any uses of the
1352 // add between the add and the compare.
1353 Builder->SetInsertPoint(OrigAdd);
1355 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName() + ".trunc");
1356 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName() + ".trunc");
1357 CallInst *Call = Builder->CreateCall(F, {TruncA, TruncB}, "sadd");
1358 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1359 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1361 // The inner add was the result of the narrow add, zero extended to the
1362 // wider type. Replace it with the result computed by the intrinsic.
1363 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1365 // The original icmp gets replaced with the overflow value.
1366 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1369 // Fold icmp Pred X, C.
1370 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1371 CmpInst::Predicate Pred = Cmp.getPredicate();
1372 Value *X = Cmp.getOperand(0);
1375 if (!match(Cmp.getOperand(1), m_APInt(C)))
1378 Value *A = nullptr, *B = nullptr;
1380 // Match the following pattern, which is a common idiom when writing
1381 // overflow-safe integer arithmetic functions. The source performs an addition
1382 // in wider type and explicitly checks for overflow using comparisons against
1383 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1385 // TODO: This could probably be generalized to handle other overflow-safe
1386 // operations if we worked out the formulas to compute the appropriate magic
1390 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1392 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1393 if (Pred == ICmpInst::ICMP_UGT &&
1394 match(X, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1395 if (Instruction *Res = processUGT_ADDCST_ADD(
1396 Cmp, A, B, CI2, cast<ConstantInt>(Cmp.getOperand(1)), *this))
1400 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1401 if (*C == 0 && Pred == ICmpInst::ICMP_SGT) {
1402 SelectPatternResult SPR = matchSelectPattern(X, A, B);
1403 if (SPR.Flavor == SPF_SMIN) {
1404 if (isKnownPositive(A, DL))
1405 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1406 if (isKnownPositive(B, DL))
1407 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1411 // FIXME: Use m_APInt to allow folds for splat constants.
1412 ConstantInt *CI = dyn_cast<ConstantInt>(Cmp.getOperand(1));
1416 // Canonicalize icmp instructions based on dominating conditions.
1417 BasicBlock *Parent = Cmp.getParent();
1418 BasicBlock *Dom = Parent->getSinglePredecessor();
1419 auto *BI = Dom ? dyn_cast<BranchInst>(Dom->getTerminator()) : nullptr;
1420 ICmpInst::Predicate Pred2;
1421 BasicBlock *TrueBB, *FalseBB;
1423 if (BI && match(BI, m_Br(m_ICmp(Pred2, m_Specific(X), m_ConstantInt(CI2)),
1424 TrueBB, FalseBB)) &&
1425 TrueBB != FalseBB) {
1427 ConstantRange::makeAllowedICmpRegion(Pred, CI->getValue());
1428 ConstantRange DominatingCR =
1430 ? ConstantRange::makeExactICmpRegion(Pred2, CI2->getValue())
1431 : ConstantRange::makeExactICmpRegion(
1432 CmpInst::getInversePredicate(Pred2), CI2->getValue());
1433 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1434 ConstantRange Difference = DominatingCR.difference(CR);
1435 if (Intersection.isEmptySet())
1436 return replaceInstUsesWith(Cmp, Builder->getFalse());
1437 if (Difference.isEmptySet())
1438 return replaceInstUsesWith(Cmp, Builder->getTrue());
1440 // If this is a normal comparison, it demands all bits. If it is a sign
1441 // bit comparison, it only demands the sign bit.
1443 bool IsSignBit = isSignBitCheck(Pred, CI->getValue(), UnusedBit);
1445 // Canonicalizing a sign bit comparison that gets used in a branch,
1446 // pessimizes codegen by generating branch on zero instruction instead
1447 // of a test and branch. So we avoid canonicalizing in such situations
1448 // because test and branch instruction has better branch displacement
1449 // than compare and branch instruction.
1450 if (!isBranchOnSignBitCheck(Cmp, IsSignBit) && !Cmp.isEquality()) {
1451 if (auto *AI = Intersection.getSingleElement())
1452 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder->getInt(*AI));
1453 if (auto *AD = Difference.getSingleElement())
1454 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder->getInt(*AD));
1461 /// Fold icmp (trunc X, Y), C.
1462 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1465 ICmpInst::Predicate Pred = Cmp.getPredicate();
1466 Value *X = Trunc->getOperand(0);
1467 if (*C == 1 && C->getBitWidth() > 1) {
1468 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1470 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1471 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1472 ConstantInt::get(V->getType(), 1));
1475 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1476 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1477 // of the high bits truncated out of x are known.
1478 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1479 SrcBits = X->getType()->getScalarSizeInBits();
1480 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1481 computeKnownBits(X, KnownZero, KnownOne, 0, &Cmp);
1483 // If all the high bits are known, we can do this xform.
1484 if ((KnownZero | KnownOne).countLeadingOnes() >= SrcBits - DstBits) {
1485 // Pull in the high bits from known-ones set.
1486 APInt NewRHS = C->zext(SrcBits);
1487 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1488 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1495 /// Fold icmp (xor X, Y), C.
1496 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1497 BinaryOperator *Xor,
1499 Value *X = Xor->getOperand(0);
1500 Value *Y = Xor->getOperand(1);
1502 if (!match(Y, m_APInt(XorC)))
1505 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1507 ICmpInst::Predicate Pred = Cmp.getPredicate();
1508 if ((Pred == ICmpInst::ICMP_SLT && *C == 0) ||
1509 (Pred == ICmpInst::ICMP_SGT && C->isAllOnesValue())) {
1511 // If the sign bit of the XorCst is not set, there is no change to
1512 // the operation, just stop using the Xor.
1513 if (!XorC->isNegative()) {
1514 Cmp.setOperand(0, X);
1519 // Was the old condition true if the operand is positive?
1520 bool isTrueIfPositive = Pred == ICmpInst::ICMP_SGT;
1522 // If so, the new one isn't.
1523 isTrueIfPositive ^= true;
1525 Constant *CmpConstant = cast<Constant>(Cmp.getOperand(1));
1526 if (isTrueIfPositive)
1527 return new ICmpInst(ICmpInst::ICMP_SGT, X, SubOne(CmpConstant));
1529 return new ICmpInst(ICmpInst::ICMP_SLT, X, AddOne(CmpConstant));
1532 if (Xor->hasOneUse()) {
1533 // (icmp u/s (xor X SignBit), C) -> (icmp s/u X, (xor C SignBit))
1534 if (!Cmp.isEquality() && XorC->isSignBit()) {
1535 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1536 : Cmp.getSignedPredicate();
1537 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), *C ^ *XorC));
1540 // (icmp u/s (xor X ~SignBit), C) -> (icmp s/u X, (xor C ~SignBit))
1541 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1542 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1543 : Cmp.getSignedPredicate();
1544 Pred = Cmp.getSwappedPredicate(Pred);
1545 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), *C ^ *XorC));
1549 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1550 // iff -C is a power of 2
1551 if (Pred == ICmpInst::ICMP_UGT && *XorC == ~(*C) && (*C + 1).isPowerOf2())
1552 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1554 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1555 // iff -C is a power of 2
1556 if (Pred == ICmpInst::ICMP_ULT && *XorC == -(*C) && C->isPowerOf2())
1557 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
1562 /// Fold icmp (and (sh X, Y), C2), C1.
1563 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1564 const APInt *C1, const APInt *C2) {
1565 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1566 if (!Shift || !Shift->isShift())
1569 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1570 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1571 // code produced by the clang front-end, for bitfield access.
1572 // This seemingly simple opportunity to fold away a shift turns out to be
1573 // rather complicated. See PR17827 for details.
1574 unsigned ShiftOpcode = Shift->getOpcode();
1575 bool IsShl = ShiftOpcode == Instruction::Shl;
1577 if (match(Shift->getOperand(1), m_APInt(C3))) {
1578 bool CanFold = false;
1579 if (ShiftOpcode == Instruction::AShr) {
1580 // There may be some constraints that make this possible, but nothing
1581 // simple has been discovered yet.
1583 } else if (ShiftOpcode == Instruction::Shl) {
1584 // For a left shift, we can fold if the comparison is not signed. We can
1585 // also fold a signed comparison if the mask value and comparison value
1586 // are not negative. These constraints may not be obvious, but we can
1587 // prove that they are correct using an SMT solver.
1588 if (!Cmp.isSigned() || (!C2->isNegative() && !C1->isNegative()))
1590 } else if (ShiftOpcode == Instruction::LShr) {
1591 // For a logical right shift, we can fold if the comparison is not signed.
1592 // We can also fold a signed comparison if the shifted mask value and the
1593 // shifted comparison value are not negative. These constraints may not be
1594 // obvious, but we can prove that they are correct using an SMT solver.
1595 if (!Cmp.isSigned() ||
1596 (!C2->shl(*C3).isNegative() && !C1->shl(*C3).isNegative()))
1601 APInt NewCst = IsShl ? C1->lshr(*C3) : C1->shl(*C3);
1602 APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1603 // Check to see if we are shifting out any of the bits being compared.
1604 if (SameAsC1 != *C1) {
1605 // If we shifted bits out, the fold is not going to work out. As a
1606 // special case, check to see if this means that the result is always
1607 // true or false now.
1608 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1609 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1610 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1611 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1613 Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1614 APInt NewAndCst = IsShl ? C2->lshr(*C3) : C2->shl(*C3);
1615 And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1616 And->setOperand(0, Shift->getOperand(0));
1617 Worklist.Add(Shift); // Shift is dead.
1623 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1624 // preferable because it allows the C2 << Y expression to be hoisted out of a
1625 // loop if Y is invariant and X is not.
1626 if (Shift->hasOneUse() && *C1 == 0 && Cmp.isEquality() &&
1627 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1630 IsShl ? Builder->CreateLShr(And->getOperand(1), Shift->getOperand(1))
1631 : Builder->CreateShl(And->getOperand(1), Shift->getOperand(1));
1633 // Compute X & (C2 << Y).
1634 Value *NewAnd = Builder->CreateAnd(Shift->getOperand(0), NewShift);
1635 Cmp.setOperand(0, NewAnd);
1642 /// Fold icmp (and X, C2), C1.
1643 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1644 BinaryOperator *And,
1647 if (!match(And->getOperand(1), m_APInt(C2)))
1650 if (!And->hasOneUse() || !And->getOperand(0)->hasOneUse())
1653 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1654 // the input width without changing the value produced, eliminate the cast:
1656 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1658 // We can do this transformation if the constants do not have their sign bits
1659 // set or if it is an equality comparison. Extending a relational comparison
1660 // when we're checking the sign bit would not work.
1662 if (match(And->getOperand(0), m_Trunc(m_Value(W))) &&
1663 (Cmp.isEquality() || (!C1->isNegative() && !C2->isNegative()))) {
1664 // TODO: Is this a good transform for vectors? Wider types may reduce
1665 // throughput. Should this transform be limited (even for scalars) by using
1666 // ShouldChangeType()?
1667 if (!Cmp.getType()->isVectorTy()) {
1668 Type *WideType = W->getType();
1669 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1670 Constant *ZextC1 = ConstantInt::get(WideType, C1->zext(WideScalarBits));
1671 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1672 Value *NewAnd = Builder->CreateAnd(W, ZextC2, And->getName());
1673 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1677 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, C2))
1680 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1681 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1683 // iff pred isn't signed
1684 if (!Cmp.isSigned() && *C1 == 0 && match(And->getOperand(1), m_One())) {
1685 Constant *One = cast<Constant>(And->getOperand(1));
1686 Value *Or = And->getOperand(0);
1687 Value *A, *B, *LShr;
1688 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1689 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1690 unsigned UsesRemoved = 0;
1691 if (And->hasOneUse())
1693 if (Or->hasOneUse())
1695 if (LShr->hasOneUse())
1698 // Compute A & ((1 << B) | 1)
1699 Value *NewOr = nullptr;
1700 if (auto *C = dyn_cast<Constant>(B)) {
1701 if (UsesRemoved >= 1)
1702 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1704 if (UsesRemoved >= 3)
1705 NewOr = Builder->CreateOr(Builder->CreateShl(One, B, LShr->getName(),
1707 One, Or->getName());
1710 Value *NewAnd = Builder->CreateAnd(A, NewOr, And->getName());
1711 Cmp.setOperand(0, NewAnd);
1717 // (X & C2) > C1 --> (X & C2) != 0, if any bit set in (X & C2) will produce a
1718 // result greater than C1.
1719 unsigned NumTZ = C2->countTrailingZeros();
1720 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && NumTZ < C2->getBitWidth() &&
1721 APInt::getOneBitSet(C2->getBitWidth(), NumTZ).ugt(*C1)) {
1722 Constant *Zero = Constant::getNullValue(And->getType());
1723 return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
1729 /// Fold icmp (and X, Y), C.
1730 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1731 BinaryOperator *And,
1733 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1736 // TODO: These all require that Y is constant too, so refactor with the above.
1738 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1739 Value *X = And->getOperand(0);
1740 Value *Y = And->getOperand(1);
1741 if (auto *LI = dyn_cast<LoadInst>(X))
1742 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1743 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1744 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1745 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1746 ConstantInt *C2 = cast<ConstantInt>(Y);
1747 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1751 if (!Cmp.isEquality())
1754 // X & -C == -C -> X > u ~C
1755 // X & -C != -C -> X <= u ~C
1756 // iff C is a power of 2
1757 if (Cmp.getOperand(1) == Y && (-(*C)).isPowerOf2()) {
1758 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1759 : CmpInst::ICMP_ULE;
1760 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1763 // (X & C2) == 0 -> (trunc X) >= 0
1764 // (X & C2) != 0 -> (trunc X) < 0
1765 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1767 if (And->hasOneUse() && *C == 0 && match(Y, m_APInt(C2))) {
1768 int32_t ExactLogBase2 = C2->exactLogBase2();
1769 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1770 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1771 if (And->getType()->isVectorTy())
1772 NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1773 Value *Trunc = Builder->CreateTrunc(X, NTy);
1774 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1775 : CmpInst::ICMP_SLT;
1776 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1783 /// Fold icmp (or X, Y), C.
1784 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1786 ICmpInst::Predicate Pred = Cmp.getPredicate();
1788 // icmp slt signum(V) 1 --> icmp slt V, 1
1790 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1791 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1792 ConstantInt::get(V->getType(), 1));
1795 if (!Cmp.isEquality() || *C != 0 || !Or->hasOneUse())
1799 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1800 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1801 // -> and (icmp eq P, null), (icmp eq Q, null).
1803 Builder->CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1805 Builder->CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1806 auto LogicOpc = Pred == ICmpInst::Predicate::ICMP_EQ ? Instruction::And
1808 return BinaryOperator::Create(LogicOpc, CmpP, CmpQ);
1814 /// Fold icmp (mul X, Y), C.
1815 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1816 BinaryOperator *Mul,
1819 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1822 // If this is a test of the sign bit and the multiply is sign-preserving with
1823 // a constant operand, use the multiply LHS operand instead.
1824 ICmpInst::Predicate Pred = Cmp.getPredicate();
1825 if (isSignTest(Pred, *C) && Mul->hasNoSignedWrap()) {
1826 if (MulC->isNegative())
1827 Pred = ICmpInst::getSwappedPredicate(Pred);
1828 return new ICmpInst(Pred, Mul->getOperand(0),
1829 Constant::getNullValue(Mul->getType()));
1835 /// Fold icmp (shl 1, Y), C.
1836 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
1839 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1842 Type *ShiftType = Shl->getType();
1843 uint32_t TypeBits = C->getBitWidth();
1844 bool CIsPowerOf2 = C->isPowerOf2();
1845 ICmpInst::Predicate Pred = Cmp.getPredicate();
1846 if (Cmp.isUnsigned()) {
1847 // (1 << Y) pred C -> Y pred Log2(C)
1849 // (1 << Y) < 30 -> Y <= 4
1850 // (1 << Y) <= 30 -> Y <= 4
1851 // (1 << Y) >= 30 -> Y > 4
1852 // (1 << Y) > 30 -> Y > 4
1853 if (Pred == ICmpInst::ICMP_ULT)
1854 Pred = ICmpInst::ICMP_ULE;
1855 else if (Pred == ICmpInst::ICMP_UGE)
1856 Pred = ICmpInst::ICMP_UGT;
1859 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1860 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1861 unsigned CLog2 = C->logBase2();
1862 if (CLog2 == TypeBits - 1) {
1863 if (Pred == ICmpInst::ICMP_UGE)
1864 Pred = ICmpInst::ICMP_EQ;
1865 else if (Pred == ICmpInst::ICMP_ULT)
1866 Pred = ICmpInst::ICMP_NE;
1868 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1869 } else if (Cmp.isSigned()) {
1870 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1871 if (C->isAllOnesValue()) {
1872 // (1 << Y) <= -1 -> Y == 31
1873 if (Pred == ICmpInst::ICMP_SLE)
1874 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1876 // (1 << Y) > -1 -> Y != 31
1877 if (Pred == ICmpInst::ICMP_SGT)
1878 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1880 // (1 << Y) < 0 -> Y == 31
1881 // (1 << Y) <= 0 -> Y == 31
1882 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1883 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1885 // (1 << Y) >= 0 -> Y != 31
1886 // (1 << Y) > 0 -> Y != 31
1887 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1888 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1890 } else if (Cmp.isEquality() && CIsPowerOf2) {
1891 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C->logBase2()));
1897 /// Fold icmp (shl X, Y), C.
1898 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1899 BinaryOperator *Shl,
1901 const APInt *ShiftVal;
1902 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1903 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), *C, *ShiftVal);
1905 const APInt *ShiftAmt;
1906 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1907 return foldICmpShlOne(Cmp, Shl, C);
1909 // Check that the shift amount is in range. If not, don't perform undefined
1910 // shifts. When the shift is visited, it will be simplified.
1911 unsigned TypeBits = C->getBitWidth();
1912 if (ShiftAmt->uge(TypeBits))
1915 ICmpInst::Predicate Pred = Cmp.getPredicate();
1916 Value *X = Shl->getOperand(0);
1917 if (Cmp.isEquality()) {
1918 // If the shift is NUW, then it is just shifting out zeros, no need for an
1920 Constant *LShrC = ConstantInt::get(Shl->getType(), C->lshr(*ShiftAmt));
1921 if (Shl->hasNoUnsignedWrap())
1922 return new ICmpInst(Pred, X, LShrC);
1924 // If the shift is NSW and we compare to 0, then it is just shifting out
1925 // sign bits, no need for an AND either.
1926 if (Shl->hasNoSignedWrap() && *C == 0)
1927 return new ICmpInst(Pred, X, LShrC);
1929 if (Shl->hasOneUse()) {
1930 // Otherwise, strength reduce the shift into an and.
1931 Constant *Mask = ConstantInt::get(Shl->getType(),
1932 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
1934 Value *And = Builder->CreateAnd(X, Mask, Shl->getName() + ".mask");
1935 return new ICmpInst(Pred, And, LShrC);
1939 // If this is a signed comparison to 0 and the shift is sign preserving,
1940 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
1941 // do that if we're sure to not continue on in this function.
1942 if (Shl->hasNoSignedWrap() && isSignTest(Pred, *C))
1943 return new ICmpInst(Pred, X, Constant::getNullValue(X->getType()));
1945 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1946 bool TrueIfSigned = false;
1947 if (Shl->hasOneUse() && isSignBitCheck(Pred, *C, TrueIfSigned)) {
1948 // (X << 31) <s 0 --> (X & 1) != 0
1949 Constant *Mask = ConstantInt::get(
1951 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
1952 Value *And = Builder->CreateAnd(X, Mask, Shl->getName() + ".mask");
1953 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1954 And, Constant::getNullValue(And->getType()));
1957 // When the shift is nuw and pred is >u or <=u, comparison only really happens
1958 // in the pre-shifted bits. Since InstSimplify canonicalizes <=u into <u, the
1959 // <=u case can be further converted to match <u (see below).
1960 if (Shl->hasNoUnsignedWrap() &&
1961 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT)) {
1962 // Derivation for the ult case:
1963 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
1964 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
1965 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
1966 assert((Pred != ICmpInst::ICMP_ULT || C->ugt(0)) &&
1967 "Encountered `ult 0` that should have been eliminated by "
1969 APInt ShiftedC = Pred == ICmpInst::ICMP_ULT ? (*C - 1).lshr(*ShiftAmt) + 1
1970 : C->lshr(*ShiftAmt);
1971 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), ShiftedC));
1974 // Transform (icmp pred iM (shl iM %v, N), C)
1975 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
1976 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
1977 // This enables us to get rid of the shift in favor of a trunc that may be
1978 // free on the target. It has the additional benefit of comparing to a
1979 // smaller constant that may be more target-friendly.
1980 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
1981 if (Shl->hasOneUse() && Amt != 0 && C->countTrailingZeros() >= Amt &&
1982 DL.isLegalInteger(TypeBits - Amt)) {
1983 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
1984 if (X->getType()->isVectorTy())
1985 TruncTy = VectorType::get(TruncTy, X->getType()->getVectorNumElements());
1987 ConstantInt::get(TruncTy, C->ashr(*ShiftAmt).trunc(TypeBits - Amt));
1988 return new ICmpInst(Pred, Builder->CreateTrunc(X, TruncTy), NewC);
1994 /// Fold icmp ({al}shr X, Y), C.
1995 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
1996 BinaryOperator *Shr,
1998 // An exact shr only shifts out zero bits, so:
1999 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2000 Value *X = Shr->getOperand(0);
2001 CmpInst::Predicate Pred = Cmp.getPredicate();
2002 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() && *C == 0)
2003 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2005 const APInt *ShiftVal;
2006 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2007 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), *C, *ShiftVal);
2009 const APInt *ShiftAmt;
2010 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2013 // Check that the shift amount is in range. If not, don't perform undefined
2014 // shifts. When the shift is visited it will be simplified.
2015 unsigned TypeBits = C->getBitWidth();
2016 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2017 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2020 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2021 if (!Cmp.isEquality()) {
2022 // If we have an unsigned comparison and an ashr, we can't simplify this.
2023 // Similarly for signed comparisons with lshr.
2024 if (Cmp.isSigned() != IsAShr)
2027 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
2028 // by a power of 2. Since we already have logic to simplify these,
2029 // transform to div and then simplify the resultant comparison.
2030 if (IsAShr && (!Shr->isExact() || ShAmtVal == TypeBits - 1))
2033 // Revisit the shift (to delete it).
2036 Constant *DivCst = ConstantInt::get(
2037 Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
2039 Value *Tmp = IsAShr ? Builder->CreateSDiv(X, DivCst, "", Shr->isExact())
2040 : Builder->CreateUDiv(X, DivCst, "", Shr->isExact());
2042 Cmp.setOperand(0, Tmp);
2044 // If the builder folded the binop, just return it.
2045 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
2049 // Otherwise, fold this div/compare.
2050 assert(TheDiv->getOpcode() == Instruction::SDiv ||
2051 TheDiv->getOpcode() == Instruction::UDiv);
2053 Instruction *Res = foldICmpDivConstant(Cmp, TheDiv, C);
2054 assert(Res && "This div/cst should have folded!");
2058 // Handle equality comparisons of shift-by-constant.
2060 // If the comparison constant changes with the shift, the comparison cannot
2061 // succeed (bits of the comparison constant cannot match the shifted value).
2062 // This should be known by InstSimplify and already be folded to true/false.
2063 assert(((IsAShr && C->shl(ShAmtVal).ashr(ShAmtVal) == *C) ||
2064 (!IsAShr && C->shl(ShAmtVal).lshr(ShAmtVal) == *C)) &&
2065 "Expected icmp+shr simplify did not occur.");
2067 // Check if the bits shifted out are known to be zero. If so, we can compare
2068 // against the unshifted value:
2069 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2070 Constant *ShiftedCmpRHS = ConstantInt::get(Shr->getType(), *C << ShAmtVal);
2071 if (Shr->hasOneUse()) {
2073 return new ICmpInst(Pred, X, ShiftedCmpRHS);
2075 // Otherwise strength reduce the shift into an 'and'.
2076 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2077 Constant *Mask = ConstantInt::get(Shr->getType(), Val);
2078 Value *And = Builder->CreateAnd(X, Mask, Shr->getName() + ".mask");
2079 return new ICmpInst(Pred, And, ShiftedCmpRHS);
2085 /// Fold icmp (udiv X, Y), C.
2086 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2087 BinaryOperator *UDiv,
2090 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2093 assert(C2 != 0 && "udiv 0, X should have been simplified already.");
2095 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2096 Value *Y = UDiv->getOperand(1);
2097 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2098 assert(!C->isMaxValue() &&
2099 "icmp ugt X, UINT_MAX should have been simplified already.");
2100 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2101 ConstantInt::get(Y->getType(), C2->udiv(*C + 1)));
2104 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2105 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2106 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2107 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2108 ConstantInt::get(Y->getType(), C2->udiv(*C)));
2114 /// Fold icmp ({su}div X, Y), C.
2115 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2116 BinaryOperator *Div,
2118 // Fold: icmp pred ([us]div X, C2), C -> range test
2119 // Fold this div into the comparison, producing a range check.
2120 // Determine, based on the divide type, what the range is being
2121 // checked. If there is an overflow on the low or high side, remember
2122 // it, otherwise compute the range [low, hi) bounding the new value.
2123 // See: InsertRangeTest above for the kinds of replacements possible.
2125 if (!match(Div->getOperand(1), m_APInt(C2)))
2128 // FIXME: If the operand types don't match the type of the divide
2129 // then don't attempt this transform. The code below doesn't have the
2130 // logic to deal with a signed divide and an unsigned compare (and
2131 // vice versa). This is because (x /s C2) <s C produces different
2132 // results than (x /s C2) <u C or (x /u C2) <s C or even
2133 // (x /u C2) <u C. Simply casting the operands and result won't
2134 // work. :( The if statement below tests that condition and bails
2136 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2137 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2140 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2141 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2142 // division-by-constant cases should be present, we can not assert that they
2143 // have happened before we reach this icmp instruction.
2144 if (*C2 == 0 || *C2 == 1 || (DivIsSigned && C2->isAllOnesValue()))
2147 // TODO: We could do all of the computations below using APInt.
2148 Constant *CmpRHS = cast<Constant>(Cmp.getOperand(1));
2149 Constant *DivRHS = cast<Constant>(Div->getOperand(1));
2151 // Compute Prod = CmpRHS * DivRHS. We are essentially solving an equation of
2152 // form X / C2 = C. We solve for X by multiplying C2 (DivRHS) and C (CmpRHS).
2153 // By solving for X, we can turn this into a range check instead of computing
2155 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
2157 // Determine if the product overflows by seeing if the product is not equal to
2158 // the divide. Make sure we do the same kind of divide as in the LHS
2159 // instruction that we're folding.
2160 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS)
2161 : ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
2163 ICmpInst::Predicate Pred = Cmp.getPredicate();
2165 // If the division is known to be exact, then there is no remainder from the
2166 // divide, so the covered range size is unit, otherwise it is the divisor.
2167 Constant *RangeSize =
2168 Div->isExact() ? ConstantInt::get(Div->getType(), 1) : DivRHS;
2170 // Figure out the interval that is being checked. For example, a comparison
2171 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2172 // Compute this interval based on the constants involved and the signedness of
2173 // the compare/divide. This computes a half-open interval, keeping track of
2174 // whether either value in the interval overflows. After analysis each
2175 // overflow variable is set to 0 if it's corresponding bound variable is valid
2176 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2177 int LoOverflow = 0, HiOverflow = 0;
2178 Constant *LoBound = nullptr, *HiBound = nullptr;
2180 if (!DivIsSigned) { // udiv
2181 // e.g. X/5 op 3 --> [15, 20)
2183 HiOverflow = LoOverflow = ProdOV;
2185 // If this is not an exact divide, then many values in the range collapse
2186 // to the same result value.
2187 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2189 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2190 if (*C == 0) { // (X / pos) op 0
2191 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2192 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
2193 HiBound = RangeSize;
2194 } else if (C->isStrictlyPositive()) { // (X / pos) op pos
2195 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2196 HiOverflow = LoOverflow = ProdOV;
2198 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2199 } else { // (X / pos) op neg
2200 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2201 HiBound = AddOne(Prod);
2202 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2204 Constant *DivNeg = ConstantExpr::getNeg(RangeSize);
2205 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2208 } else if (C2->isNegative()) { // Divisor is < 0.
2210 RangeSize = ConstantExpr::getNeg(RangeSize);
2211 if (*C == 0) { // (X / neg) op 0
2212 // e.g. X/-5 op 0 --> [-4, 5)
2213 LoBound = AddOne(RangeSize);
2214 HiBound = ConstantExpr::getNeg(RangeSize);
2215 if (HiBound == DivRHS) { // -INTMIN = INTMIN
2216 HiOverflow = 1; // [INTMIN+1, overflow)
2217 HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
2219 } else if (C->isStrictlyPositive()) { // (X / neg) op pos
2220 // e.g. X/-5 op 3 --> [-19, -14)
2221 HiBound = AddOne(Prod);
2222 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2224 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2225 } else { // (X / neg) op neg
2226 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2227 LoOverflow = HiOverflow = ProdOV;
2229 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2232 // Dividing by a negative swaps the condition. LT <-> GT
2233 Pred = ICmpInst::getSwappedPredicate(Pred);
2236 Value *X = Div->getOperand(0);
2238 default: llvm_unreachable("Unhandled icmp opcode!");
2239 case ICmpInst::ICMP_EQ:
2240 if (LoOverflow && HiOverflow)
2241 return replaceInstUsesWith(Cmp, Builder->getFalse());
2243 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2244 ICmpInst::ICMP_UGE, X, LoBound);
2246 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2247 ICmpInst::ICMP_ULT, X, HiBound);
2248 return replaceInstUsesWith(
2249 Cmp, insertRangeTest(X, LoBound->getUniqueInteger(),
2250 HiBound->getUniqueInteger(), DivIsSigned, true));
2251 case ICmpInst::ICMP_NE:
2252 if (LoOverflow && HiOverflow)
2253 return replaceInstUsesWith(Cmp, Builder->getTrue());
2255 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2256 ICmpInst::ICMP_ULT, X, LoBound);
2258 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2259 ICmpInst::ICMP_UGE, X, HiBound);
2260 return replaceInstUsesWith(Cmp,
2261 insertRangeTest(X, LoBound->getUniqueInteger(),
2262 HiBound->getUniqueInteger(),
2263 DivIsSigned, false));
2264 case ICmpInst::ICMP_ULT:
2265 case ICmpInst::ICMP_SLT:
2266 if (LoOverflow == +1) // Low bound is greater than input range.
2267 return replaceInstUsesWith(Cmp, Builder->getTrue());
2268 if (LoOverflow == -1) // Low bound is less than input range.
2269 return replaceInstUsesWith(Cmp, Builder->getFalse());
2270 return new ICmpInst(Pred, X, LoBound);
2271 case ICmpInst::ICMP_UGT:
2272 case ICmpInst::ICMP_SGT:
2273 if (HiOverflow == +1) // High bound greater than input range.
2274 return replaceInstUsesWith(Cmp, Builder->getFalse());
2275 if (HiOverflow == -1) // High bound less than input range.
2276 return replaceInstUsesWith(Cmp, Builder->getTrue());
2277 if (Pred == ICmpInst::ICMP_UGT)
2278 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
2279 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
2285 /// Fold icmp (sub X, Y), C.
2286 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2287 BinaryOperator *Sub,
2289 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2290 ICmpInst::Predicate Pred = Cmp.getPredicate();
2292 // The following transforms are only worth it if the only user of the subtract
2294 if (!Sub->hasOneUse())
2297 if (Sub->hasNoSignedWrap()) {
2298 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2299 if (Pred == ICmpInst::ICMP_SGT && C->isAllOnesValue())
2300 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2302 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2303 if (Pred == ICmpInst::ICMP_SGT && *C == 0)
2304 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2306 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2307 if (Pred == ICmpInst::ICMP_SLT && *C == 0)
2308 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2310 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2311 if (Pred == ICmpInst::ICMP_SLT && *C == 1)
2312 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2316 if (!match(X, m_APInt(C2)))
2319 // C2 - Y <u C -> (Y | (C - 1)) == C2
2320 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2321 if (Pred == ICmpInst::ICMP_ULT && C->isPowerOf2() &&
2322 (*C2 & (*C - 1)) == (*C - 1))
2323 return new ICmpInst(ICmpInst::ICMP_EQ, Builder->CreateOr(Y, *C - 1), X);
2325 // C2 - Y >u C -> (Y | C) != C2
2326 // iff C2 & C == C and C + 1 is a power of 2
2327 if (Pred == ICmpInst::ICMP_UGT && (*C + 1).isPowerOf2() && (*C2 & *C) == *C)
2328 return new ICmpInst(ICmpInst::ICMP_NE, Builder->CreateOr(Y, *C), X);
2333 /// Fold icmp (add X, Y), C.
2334 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2335 BinaryOperator *Add,
2337 Value *Y = Add->getOperand(1);
2339 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2342 // Fold icmp pred (add X, C2), C.
2343 Value *X = Add->getOperand(0);
2344 Type *Ty = Add->getType();
2346 ConstantRange::makeExactICmpRegion(Cmp.getPredicate(), *C).subtract(*C2);
2347 const APInt &Upper = CR.getUpper();
2348 const APInt &Lower = CR.getLower();
2349 if (Cmp.isSigned()) {
2350 if (Lower.isSignBit())
2351 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2352 if (Upper.isSignBit())
2353 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2355 if (Lower.isMinValue())
2356 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2357 if (Upper.isMinValue())
2358 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2361 if (!Add->hasOneUse())
2364 // X+C <u C2 -> (X & -C2) == C
2365 // iff C & (C2-1) == 0
2366 // C2 is a power of 2
2367 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT && C->isPowerOf2() &&
2368 (*C2 & (*C - 1)) == 0)
2369 return new ICmpInst(ICmpInst::ICMP_EQ, Builder->CreateAnd(X, -(*C)),
2370 ConstantExpr::getNeg(cast<Constant>(Y)));
2372 // X+C >u C2 -> (X & ~C2) != C
2374 // C2+1 is a power of 2
2375 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && (*C + 1).isPowerOf2() &&
2377 return new ICmpInst(ICmpInst::ICMP_NE, Builder->CreateAnd(X, ~(*C)),
2378 ConstantExpr::getNeg(cast<Constant>(Y)));
2383 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2384 /// where X is some kind of instruction.
2385 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2387 if (!match(Cmp.getOperand(1), m_APInt(C)))
2391 if (match(Cmp.getOperand(0), m_BinOp(BO))) {
2392 switch (BO->getOpcode()) {
2393 case Instruction::Xor:
2394 if (Instruction *I = foldICmpXorConstant(Cmp, BO, C))
2397 case Instruction::And:
2398 if (Instruction *I = foldICmpAndConstant(Cmp, BO, C))
2401 case Instruction::Or:
2402 if (Instruction *I = foldICmpOrConstant(Cmp, BO, C))
2405 case Instruction::Mul:
2406 if (Instruction *I = foldICmpMulConstant(Cmp, BO, C))
2409 case Instruction::Shl:
2410 if (Instruction *I = foldICmpShlConstant(Cmp, BO, C))
2413 case Instruction::LShr:
2414 case Instruction::AShr:
2415 if (Instruction *I = foldICmpShrConstant(Cmp, BO, C))
2418 case Instruction::UDiv:
2419 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, C))
2422 case Instruction::SDiv:
2423 if (Instruction *I = foldICmpDivConstant(Cmp, BO, C))
2426 case Instruction::Sub:
2427 if (Instruction *I = foldICmpSubConstant(Cmp, BO, C))
2430 case Instruction::Add:
2431 if (Instruction *I = foldICmpAddConstant(Cmp, BO, C))
2437 // TODO: These folds could be refactored to be part of the above calls.
2438 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, C))
2443 if (match(Cmp.getOperand(0), m_Instruction(LHSI)) &&
2444 LHSI->getOpcode() == Instruction::Trunc)
2445 if (Instruction *I = foldICmpTruncConstant(Cmp, LHSI, C))
2448 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, C))
2454 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2455 /// icmp eq/ne BO, C.
2456 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2459 // TODO: Some of these folds could work with arbitrary constants, but this
2460 // function is limited to scalar and vector splat constants.
2461 if (!Cmp.isEquality())
2464 ICmpInst::Predicate Pred = Cmp.getPredicate();
2465 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2466 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2467 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2469 switch (BO->getOpcode()) {
2470 case Instruction::SRem:
2471 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2472 if (*C == 0 && BO->hasOneUse()) {
2474 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2475 Value *NewRem = Builder->CreateURem(BOp0, BOp1, BO->getName());
2476 return new ICmpInst(Pred, NewRem,
2477 Constant::getNullValue(BO->getType()));
2481 case Instruction::Add: {
2482 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2484 if (match(BOp1, m_APInt(BOC))) {
2485 if (BO->hasOneUse()) {
2486 Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2487 return new ICmpInst(Pred, BOp0, SubC);
2489 } else if (*C == 0) {
2490 // Replace ((add A, B) != 0) with (A != -B) if A or B is
2491 // efficiently invertible, or if the add has just this one use.
2492 if (Value *NegVal = dyn_castNegVal(BOp1))
2493 return new ICmpInst(Pred, BOp0, NegVal);
2494 if (Value *NegVal = dyn_castNegVal(BOp0))
2495 return new ICmpInst(Pred, NegVal, BOp1);
2496 if (BO->hasOneUse()) {
2497 Value *Neg = Builder->CreateNeg(BOp1);
2499 return new ICmpInst(Pred, BOp0, Neg);
2504 case Instruction::Xor:
2505 if (BO->hasOneUse()) {
2506 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2507 // For the xor case, we can xor two constants together, eliminating
2508 // the explicit xor.
2509 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2510 } else if (*C == 0) {
2511 // Replace ((xor A, B) != 0) with (A != B)
2512 return new ICmpInst(Pred, BOp0, BOp1);
2516 case Instruction::Sub:
2517 if (BO->hasOneUse()) {
2519 if (match(BOp0, m_APInt(BOC))) {
2520 // Replace ((sub BOC, B) != C) with (B != BOC-C).
2521 Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2522 return new ICmpInst(Pred, BOp1, SubC);
2523 } else if (*C == 0) {
2524 // Replace ((sub A, B) != 0) with (A != B).
2525 return new ICmpInst(Pred, BOp0, BOp1);
2529 case Instruction::Or: {
2531 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2532 // Comparing if all bits outside of a constant mask are set?
2533 // Replace (X | C) == -1 with (X & ~C) == ~C.
2534 // This removes the -1 constant.
2535 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2536 Value *And = Builder->CreateAnd(BOp0, NotBOC);
2537 return new ICmpInst(Pred, And, NotBOC);
2541 case Instruction::And: {
2543 if (match(BOp1, m_APInt(BOC))) {
2544 // If we have ((X & C) == C), turn it into ((X & C) != 0).
2545 if (C == BOC && C->isPowerOf2())
2546 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2547 BO, Constant::getNullValue(RHS->getType()));
2549 // Don't perform the following transforms if the AND has multiple uses
2550 if (!BO->hasOneUse())
2553 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
2554 if (BOC->isSignBit()) {
2555 Constant *Zero = Constant::getNullValue(BOp0->getType());
2556 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
2557 return new ICmpInst(NewPred, BOp0, Zero);
2560 // ((X & ~7) == 0) --> X < 8
2561 if (*C == 0 && (~(*BOC) + 1).isPowerOf2()) {
2562 Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1));
2563 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
2564 return new ICmpInst(NewPred, BOp0, NegBOC);
2569 case Instruction::Mul:
2570 if (*C == 0 && BO->hasNoSignedWrap()) {
2572 if (match(BOp1, m_APInt(BOC)) && *BOC != 0) {
2573 // The trivial case (mul X, 0) is handled by InstSimplify.
2574 // General case : (mul X, C) != 0 iff X != 0
2575 // (mul X, C) == 0 iff X == 0
2576 return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2580 case Instruction::UDiv:
2582 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2583 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2584 return new ICmpInst(NewPred, BOp1, BOp0);
2593 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
2594 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2596 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0));
2597 if (!II || !Cmp.isEquality())
2600 // Handle icmp {eq|ne} <intrinsic>, intcst.
2601 switch (II->getIntrinsicID()) {
2602 case Intrinsic::bswap:
2604 Cmp.setOperand(0, II->getArgOperand(0));
2605 Cmp.setOperand(1, Builder->getInt(C->byteSwap()));
2607 case Intrinsic::ctlz:
2608 case Intrinsic::cttz:
2609 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
2610 if (*C == C->getBitWidth()) {
2612 Cmp.setOperand(0, II->getArgOperand(0));
2613 Cmp.setOperand(1, ConstantInt::getNullValue(II->getType()));
2617 case Intrinsic::ctpop: {
2618 // popcount(A) == 0 -> A == 0 and likewise for !=
2619 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
2620 bool IsZero = *C == 0;
2621 if (IsZero || *C == C->getBitWidth()) {
2623 Cmp.setOperand(0, II->getArgOperand(0));
2624 auto *NewOp = IsZero ? Constant::getNullValue(II->getType())
2625 : Constant::getAllOnesValue(II->getType());
2626 Cmp.setOperand(1, NewOp);
2637 /// Handle icmp with constant (but not simple integer constant) RHS.
2638 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
2639 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2640 Constant *RHSC = dyn_cast<Constant>(Op1);
2641 Instruction *LHSI = dyn_cast<Instruction>(Op0);
2645 switch (LHSI->getOpcode()) {
2646 case Instruction::GetElementPtr:
2647 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2648 if (RHSC->isNullValue() &&
2649 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2650 return new ICmpInst(
2651 I.getPredicate(), LHSI->getOperand(0),
2652 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2654 case Instruction::PHI:
2655 // Only fold icmp into the PHI if the phi and icmp are in the same
2656 // block. If in the same block, we're encouraging jump threading. If
2657 // not, we are just pessimizing the code by making an i1 phi.
2658 if (LHSI->getParent() == I.getParent())
2659 if (Instruction *NV = FoldOpIntoPhi(I))
2662 case Instruction::Select: {
2663 // If either operand of the select is a constant, we can fold the
2664 // comparison into the select arms, which will cause one to be
2665 // constant folded and the select turned into a bitwise or.
2666 Value *Op1 = nullptr, *Op2 = nullptr;
2667 ConstantInt *CI = nullptr;
2668 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2669 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2670 CI = dyn_cast<ConstantInt>(Op1);
2672 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2673 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2674 CI = dyn_cast<ConstantInt>(Op2);
2677 // We only want to perform this transformation if it will not lead to
2678 // additional code. This is true if either both sides of the select
2679 // fold to a constant (in which case the icmp is replaced with a select
2680 // which will usually simplify) or this is the only user of the
2681 // select (in which case we are trading a select+icmp for a simpler
2682 // select+icmp) or all uses of the select can be replaced based on
2683 // dominance information ("Global cases").
2684 bool Transform = false;
2687 else if (Op1 || Op2) {
2689 if (LHSI->hasOneUse())
2692 else if (CI && !CI->isZero())
2693 // When Op1 is constant try replacing select with second operand.
2694 // Otherwise Op2 is constant and try replacing select with first
2697 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
2701 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
2704 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
2706 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2710 case Instruction::IntToPtr:
2711 // icmp pred inttoptr(X), null -> icmp pred X, 0
2712 if (RHSC->isNullValue() &&
2713 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
2714 return new ICmpInst(
2715 I.getPredicate(), LHSI->getOperand(0),
2716 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2719 case Instruction::Load:
2720 // Try to optimize things like "A[i] > 4" to index computations.
2721 if (GetElementPtrInst *GEP =
2722 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2723 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2724 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2725 !cast<LoadInst>(LHSI)->isVolatile())
2726 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
2735 /// Try to fold icmp (binop), X or icmp X, (binop).
2736 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
2737 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2739 // Special logic for binary operators.
2740 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2741 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2745 CmpInst::Predicate Pred = I.getPredicate();
2746 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2747 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2749 ICmpInst::isEquality(Pred) ||
2750 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2751 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2752 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2754 ICmpInst::isEquality(Pred) ||
2755 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2756 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2758 // Analyze the case when either Op0 or Op1 is an add instruction.
2759 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2760 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2761 if (BO0 && BO0->getOpcode() == Instruction::Add) {
2762 A = BO0->getOperand(0);
2763 B = BO0->getOperand(1);
2765 if (BO1 && BO1->getOpcode() == Instruction::Add) {
2766 C = BO1->getOperand(0);
2767 D = BO1->getOperand(1);
2770 // icmp (X+cst) < 0 --> X < -cst
2771 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred) && match(Op1, m_Zero()))
2772 if (ConstantInt *RHSC = dyn_cast_or_null<ConstantInt>(B))
2773 if (!RHSC->isMinValue(/*isSigned=*/true))
2774 return new ICmpInst(Pred, A, ConstantExpr::getNeg(RHSC));
2776 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2777 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2778 return new ICmpInst(Pred, A == Op1 ? B : A,
2779 Constant::getNullValue(Op1->getType()));
2781 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2782 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2783 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2786 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2787 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
2789 // Try not to increase register pressure.
2790 BO0->hasOneUse() && BO1->hasOneUse()) {
2791 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2794 // C + B == C + D -> B == D
2797 } else if (A == D) {
2798 // D + B == C + D -> B == C
2801 } else if (B == C) {
2802 // A + C == C + D -> A == D
2807 // A + D == C + D -> A == C
2811 return new ICmpInst(Pred, Y, Z);
2814 // icmp slt (X + -1), Y -> icmp sle X, Y
2815 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2816 match(B, m_AllOnes()))
2817 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2819 // icmp sge (X + -1), Y -> icmp sgt X, Y
2820 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2821 match(B, m_AllOnes()))
2822 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2824 // icmp sle (X + 1), Y -> icmp slt X, Y
2825 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
2826 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2828 // icmp sgt (X + 1), Y -> icmp sge X, Y
2829 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
2830 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2832 // icmp sgt X, (Y + -1) -> icmp sge X, Y
2833 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
2834 match(D, m_AllOnes()))
2835 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
2837 // icmp sle X, (Y + -1) -> icmp slt X, Y
2838 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
2839 match(D, m_AllOnes()))
2840 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
2842 // icmp sge X, (Y + 1) -> icmp sgt X, Y
2843 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
2844 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
2846 // icmp slt X, (Y + 1) -> icmp sle X, Y
2847 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
2848 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
2850 // if C1 has greater magnitude than C2:
2851 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2852 // s.t. C3 = C1 - C2
2854 // if C2 has greater magnitude than C1:
2855 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2856 // s.t. C3 = C2 - C1
2857 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2858 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2859 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2860 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2861 const APInt &AP1 = C1->getValue();
2862 const APInt &AP2 = C2->getValue();
2863 if (AP1.isNegative() == AP2.isNegative()) {
2864 APInt AP1Abs = C1->getValue().abs();
2865 APInt AP2Abs = C2->getValue().abs();
2866 if (AP1Abs.uge(AP2Abs)) {
2867 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2868 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2869 return new ICmpInst(Pred, NewAdd, C);
2871 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2872 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2873 return new ICmpInst(Pred, A, NewAdd);
2878 // Analyze the case when either Op0 or Op1 is a sub instruction.
2879 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2884 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
2885 A = BO0->getOperand(0);
2886 B = BO0->getOperand(1);
2888 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
2889 C = BO1->getOperand(0);
2890 D = BO1->getOperand(1);
2893 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2894 if (A == Op1 && NoOp0WrapProblem)
2895 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2897 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2898 if (C == Op0 && NoOp1WrapProblem)
2899 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2901 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2902 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2903 // Try not to increase register pressure.
2904 BO0->hasOneUse() && BO1->hasOneUse())
2905 return new ICmpInst(Pred, A, C);
2907 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2908 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2909 // Try not to increase register pressure.
2910 BO0->hasOneUse() && BO1->hasOneUse())
2911 return new ICmpInst(Pred, D, B);
2913 // icmp (0-X) < cst --> x > -cst
2914 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
2916 if (match(BO0, m_Neg(m_Value(X))))
2917 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2918 if (!RHSC->isMinValue(/*isSigned=*/true))
2919 return new ICmpInst(I.getSwappedPredicate(), X,
2920 ConstantExpr::getNeg(RHSC));
2923 BinaryOperator *SRem = nullptr;
2924 // icmp (srem X, Y), Y
2925 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
2927 // icmp Y, (srem X, Y)
2928 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2929 Op0 == BO1->getOperand(1))
2932 // We don't check hasOneUse to avoid increasing register pressure because
2933 // the value we use is the same value this instruction was already using.
2934 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2937 case ICmpInst::ICMP_EQ:
2938 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2939 case ICmpInst::ICMP_NE:
2940 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2941 case ICmpInst::ICMP_SGT:
2942 case ICmpInst::ICMP_SGE:
2943 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2944 Constant::getAllOnesValue(SRem->getType()));
2945 case ICmpInst::ICMP_SLT:
2946 case ICmpInst::ICMP_SLE:
2947 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2948 Constant::getNullValue(SRem->getType()));
2952 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
2953 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
2954 switch (BO0->getOpcode()) {
2957 case Instruction::Add:
2958 case Instruction::Sub:
2959 case Instruction::Xor:
2960 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2961 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2962 BO1->getOperand(0));
2963 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2964 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2965 if (CI->getValue().isSignBit()) {
2966 ICmpInst::Predicate Pred =
2967 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
2968 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
2971 if (BO0->getOpcode() == Instruction::Xor && CI->isMaxValue(true)) {
2972 ICmpInst::Predicate Pred =
2973 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
2974 Pred = I.getSwappedPredicate(Pred);
2975 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
2979 case Instruction::Mul:
2980 if (!I.isEquality())
2983 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2984 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2985 // Mask = -1 >> count-trailing-zeros(Cst).
2986 if (!CI->isZero() && !CI->isOne()) {
2987 const APInt &AP = CI->getValue();
2988 ConstantInt *Mask = ConstantInt::get(
2990 APInt::getLowBitsSet(AP.getBitWidth(),
2991 AP.getBitWidth() - AP.countTrailingZeros()));
2992 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2993 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2994 return new ICmpInst(I.getPredicate(), And1, And2);
2998 case Instruction::UDiv:
2999 case Instruction::LShr:
3003 case Instruction::SDiv:
3004 case Instruction::AShr:
3005 if (!BO0->isExact() || !BO1->isExact())
3007 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3008 BO1->getOperand(0));
3009 case Instruction::Shl: {
3010 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3011 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3014 if (!NSW && I.isSigned())
3016 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3017 BO1->getOperand(0));
3023 // Transform A & (L - 1) `ult` L --> L != 0
3024 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3026 m_CombineOr(m_And(m_Value(), LSubOne), m_And(LSubOne, m_Value()));
3028 if (match(BO0, BitwiseAnd) && I.getPredicate() == ICmpInst::ICMP_ULT) {
3029 auto *Zero = Constant::getNullValue(BO0->getType());
3030 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3037 /// Fold icmp Pred min|max(X, Y), X.
3038 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
3039 ICmpInst::Predicate Pred = Cmp.getPredicate();
3040 Value *Op0 = Cmp.getOperand(0);
3041 Value *X = Cmp.getOperand(1);
3043 // Canonicalize minimum or maximum operand to LHS of the icmp.
3044 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3045 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3046 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3047 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3049 Pred = Cmp.getSwappedPredicate();
3053 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3054 // smin(X, Y) == X --> X s<= Y
3055 // smin(X, Y) s>= X --> X s<= Y
3056 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3057 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3059 // smin(X, Y) != X --> X s> Y
3060 // smin(X, Y) s< X --> X s> Y
3061 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3062 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3064 // These cases should be handled in InstSimplify:
3065 // smin(X, Y) s<= X --> true
3066 // smin(X, Y) s> X --> false
3070 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3071 // smax(X, Y) == X --> X s>= Y
3072 // smax(X, Y) s<= X --> X s>= Y
3073 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3074 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3076 // smax(X, Y) != X --> X s< Y
3077 // smax(X, Y) s> X --> X s< Y
3078 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3079 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3081 // These cases should be handled in InstSimplify:
3082 // smax(X, Y) s>= X --> true
3083 // smax(X, Y) s< X --> false
3087 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3088 // umin(X, Y) == X --> X u<= Y
3089 // umin(X, Y) u>= X --> X u<= Y
3090 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3091 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3093 // umin(X, Y) != X --> X u> Y
3094 // umin(X, Y) u< X --> X u> Y
3095 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3096 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3098 // These cases should be handled in InstSimplify:
3099 // umin(X, Y) u<= X --> true
3100 // umin(X, Y) u> X --> false
3104 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3105 // umax(X, Y) == X --> X u>= Y
3106 // umax(X, Y) u<= X --> X u>= Y
3107 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3108 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3110 // umax(X, Y) != X --> X u< Y
3111 // umax(X, Y) u> X --> X u< Y
3112 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3113 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3115 // These cases should be handled in InstSimplify:
3116 // umax(X, Y) u>= X --> true
3117 // umax(X, Y) u< X --> false
3124 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3125 if (!I.isEquality())
3128 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3129 Value *A, *B, *C, *D;
3130 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3131 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3132 Value *OtherVal = A == Op1 ? B : A;
3133 return new ICmpInst(I.getPredicate(), OtherVal,
3134 Constant::getNullValue(A->getType()));
3137 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3138 // A^c1 == C^c2 --> A == C^(c1^c2)
3139 ConstantInt *C1, *C2;
3140 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3142 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3143 Value *Xor = Builder->CreateXor(C, NC);
3144 return new ICmpInst(I.getPredicate(), A, Xor);
3147 // A^B == A^D -> B == D
3149 return new ICmpInst(I.getPredicate(), B, D);
3151 return new ICmpInst(I.getPredicate(), B, C);
3153 return new ICmpInst(I.getPredicate(), A, D);
3155 return new ICmpInst(I.getPredicate(), A, C);
3159 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3160 // A == (A^B) -> B == 0
3161 Value *OtherVal = A == Op0 ? B : A;
3162 return new ICmpInst(I.getPredicate(), OtherVal,
3163 Constant::getNullValue(A->getType()));
3166 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3167 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3168 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3169 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3175 } else if (A == D) {
3179 } else if (B == C) {
3183 } else if (B == D) {
3189 if (X) { // Build (X^Y) & Z
3190 Op1 = Builder->CreateXor(X, Y);
3191 Op1 = Builder->CreateAnd(Op1, Z);
3192 I.setOperand(0, Op1);
3193 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3198 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3199 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3201 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3202 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3203 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3204 match(Op1, m_ZExt(m_Value(A))))) {
3205 APInt Pow2 = Cst1->getValue() + 1;
3206 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3207 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3208 return new ICmpInst(I.getPredicate(), A,
3209 Builder->CreateTrunc(B, A->getType()));
3212 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3213 // For lshr and ashr pairs.
3214 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3215 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3216 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3217 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3218 unsigned TypeBits = Cst1->getBitWidth();
3219 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3220 if (ShAmt < TypeBits && ShAmt != 0) {
3221 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
3222 ? ICmpInst::ICMP_UGE
3223 : ICmpInst::ICMP_ULT;
3224 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3225 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3226 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
3230 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3231 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3232 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3233 unsigned TypeBits = Cst1->getBitWidth();
3234 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3235 if (ShAmt < TypeBits && ShAmt != 0) {
3236 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3237 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3238 Value *And = Builder->CreateAnd(Xor, Builder->getInt(AndVal),
3239 I.getName() + ".mask");
3240 return new ICmpInst(I.getPredicate(), And,
3241 Constant::getNullValue(Cst1->getType()));
3245 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3246 // "icmp (and X, mask), cst"
3248 if (Op0->hasOneUse() &&
3249 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3250 match(Op1, m_ConstantInt(Cst1)) &&
3251 // Only do this when A has multiple uses. This is most important to do
3252 // when it exposes other optimizations.
3254 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3256 if (ShAmt < ASize) {
3258 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3261 APInt CmpV = Cst1->getValue().zext(ASize);
3264 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3265 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3272 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3274 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3275 const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3276 Value *LHSCIOp = LHSCI->getOperand(0);
3277 Type *SrcTy = LHSCIOp->getType();
3278 Type *DestTy = LHSCI->getType();
3281 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3282 // integer type is the same size as the pointer type.
3283 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3284 DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
3285 Value *RHSOp = nullptr;
3286 if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3287 Value *RHSCIOp = RHSC->getOperand(0);
3288 if (RHSCIOp->getType()->getPointerAddressSpace() ==
3289 LHSCIOp->getType()->getPointerAddressSpace()) {
3290 RHSOp = RHSC->getOperand(0);
3291 // If the pointer types don't match, insert a bitcast.
3292 if (LHSCIOp->getType() != RHSOp->getType())
3293 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
3295 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
3296 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
3300 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
3303 // The code below only handles extension cast instructions, so far.
3305 if (LHSCI->getOpcode() != Instruction::ZExt &&
3306 LHSCI->getOpcode() != Instruction::SExt)
3309 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
3310 bool isSignedCmp = ICmp.isSigned();
3312 if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
3313 // Not an extension from the same type?
3314 RHSCIOp = CI->getOperand(0);
3315 if (RHSCIOp->getType() != LHSCIOp->getType())
3318 // If the signedness of the two casts doesn't agree (i.e. one is a sext
3319 // and the other is a zext), then we can't handle this.
3320 if (CI->getOpcode() != LHSCI->getOpcode())
3323 // Deal with equality cases early.
3324 if (ICmp.isEquality())
3325 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3327 // A signed comparison of sign extended values simplifies into a
3328 // signed comparison.
3329 if (isSignedCmp && isSignedExt)
3330 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3332 // The other three cases all fold into an unsigned comparison.
3333 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
3336 // If we aren't dealing with a constant on the RHS, exit early.
3337 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
3341 // Compute the constant that would happen if we truncated to SrcTy then
3342 // re-extended to DestTy.
3343 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
3344 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
3346 // If the re-extended constant didn't change...
3348 // Deal with equality cases early.
3349 if (ICmp.isEquality())
3350 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3352 // A signed comparison of sign extended values simplifies into a
3353 // signed comparison.
3354 if (isSignedExt && isSignedCmp)
3355 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3357 // The other three cases all fold into an unsigned comparison.
3358 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
3361 // The re-extended constant changed, partly changed (in the case of a vector),
3362 // or could not be determined to be equal (in the case of a constant
3363 // expression), so the constant cannot be represented in the shorter type.
3364 // Consequently, we cannot emit a simple comparison.
3365 // All the cases that fold to true or false will have already been handled
3366 // by SimplifyICmpInst, so only deal with the tricky case.
3368 if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
3371 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
3372 // should have been folded away previously and not enter in here.
3374 // We're performing an unsigned comp with a sign extended value.
3375 // This is true if the input is >= 0. [aka >s -1]
3376 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
3377 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
3379 // Finally, return the value computed.
3380 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
3381 return replaceInstUsesWith(ICmp, Result);
3383 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
3384 return BinaryOperator::CreateNot(Result);
3387 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
3388 Value *RHS, Instruction &OrigI,
3389 Value *&Result, Constant *&Overflow) {
3390 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
3391 std::swap(LHS, RHS);
3393 auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
3395 Overflow = OverflowVal;
3397 Result->takeName(&OrigI);
3401 // If the overflow check was an add followed by a compare, the insertion point
3402 // may be pointing to the compare. We want to insert the new instructions
3403 // before the add in case there are uses of the add between the add and the
3405 Builder->SetInsertPoint(&OrigI);
3409 llvm_unreachable("bad overflow check kind!");
3411 case OCF_UNSIGNED_ADD: {
3412 OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
3413 if (OR == OverflowResult::NeverOverflows)
3414 return SetResult(Builder->CreateNUWAdd(LHS, RHS), Builder->getFalse(),
3417 if (OR == OverflowResult::AlwaysOverflows)
3418 return SetResult(Builder->CreateAdd(LHS, RHS), Builder->getTrue(), true);
3420 // Fall through uadd into sadd
3423 case OCF_SIGNED_ADD: {
3424 // X + 0 -> {X, false}
3425 if (match(RHS, m_Zero()))
3426 return SetResult(LHS, Builder->getFalse(), false);
3428 // We can strength reduce this signed add into a regular add if we can prove
3429 // that it will never overflow.
3430 if (OCF == OCF_SIGNED_ADD)
3431 if (WillNotOverflowSignedAdd(LHS, RHS, OrigI))
3432 return SetResult(Builder->CreateNSWAdd(LHS, RHS), Builder->getFalse(),
3437 case OCF_UNSIGNED_SUB:
3438 case OCF_SIGNED_SUB: {
3439 // X - 0 -> {X, false}
3440 if (match(RHS, m_Zero()))
3441 return SetResult(LHS, Builder->getFalse(), false);
3443 if (OCF == OCF_SIGNED_SUB) {
3444 if (WillNotOverflowSignedSub(LHS, RHS, OrigI))
3445 return SetResult(Builder->CreateNSWSub(LHS, RHS), Builder->getFalse(),
3448 if (WillNotOverflowUnsignedSub(LHS, RHS, OrigI))
3449 return SetResult(Builder->CreateNUWSub(LHS, RHS), Builder->getFalse(),
3455 case OCF_UNSIGNED_MUL: {
3456 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
3457 if (OR == OverflowResult::NeverOverflows)
3458 return SetResult(Builder->CreateNUWMul(LHS, RHS), Builder->getFalse(),
3460 if (OR == OverflowResult::AlwaysOverflows)
3461 return SetResult(Builder->CreateMul(LHS, RHS), Builder->getTrue(), true);
3464 case OCF_SIGNED_MUL:
3465 // X * undef -> undef
3466 if (isa<UndefValue>(RHS))
3467 return SetResult(RHS, UndefValue::get(Builder->getInt1Ty()), false);
3469 // X * 0 -> {0, false}
3470 if (match(RHS, m_Zero()))
3471 return SetResult(RHS, Builder->getFalse(), false);
3473 // X * 1 -> {X, false}
3474 if (match(RHS, m_One()))
3475 return SetResult(LHS, Builder->getFalse(), false);
3477 if (OCF == OCF_SIGNED_MUL)
3478 if (WillNotOverflowSignedMul(LHS, RHS, OrigI))
3479 return SetResult(Builder->CreateNSWMul(LHS, RHS), Builder->getFalse(),
3487 /// \brief Recognize and process idiom involving test for multiplication
3490 /// The caller has matched a pattern of the form:
3491 /// I = cmp u (mul(zext A, zext B), V
3492 /// The function checks if this is a test for overflow and if so replaces
3493 /// multiplication with call to 'mul.with.overflow' intrinsic.
3495 /// \param I Compare instruction.
3496 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
3497 /// the compare instruction. Must be of integer type.
3498 /// \param OtherVal The other argument of compare instruction.
3499 /// \returns Instruction which must replace the compare instruction, NULL if no
3500 /// replacement required.
3501 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
3502 Value *OtherVal, InstCombiner &IC) {
3503 // Don't bother doing this transformation for pointers, don't do it for
3505 if (!isa<IntegerType>(MulVal->getType()))
3508 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
3509 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
3510 auto *MulInstr = dyn_cast<Instruction>(MulVal);
3513 assert(MulInstr->getOpcode() == Instruction::Mul);
3515 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
3516 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
3517 assert(LHS->getOpcode() == Instruction::ZExt);
3518 assert(RHS->getOpcode() == Instruction::ZExt);
3519 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
3521 // Calculate type and width of the result produced by mul.with.overflow.
3522 Type *TyA = A->getType(), *TyB = B->getType();
3523 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
3524 WidthB = TyB->getPrimitiveSizeInBits();
3527 if (WidthB > WidthA) {
3535 // In order to replace the original mul with a narrower mul.with.overflow,
3536 // all uses must ignore upper bits of the product. The number of used low
3537 // bits must be not greater than the width of mul.with.overflow.
3538 if (MulVal->hasNUsesOrMore(2))
3539 for (User *U : MulVal->users()) {
3542 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3543 // Check if truncation ignores bits above MulWidth.
3544 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
3545 if (TruncWidth > MulWidth)
3547 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3548 // Check if AND ignores bits above MulWidth.
3549 if (BO->getOpcode() != Instruction::And)
3551 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3552 const APInt &CVal = CI->getValue();
3553 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
3557 // Other uses prohibit this transformation.
3562 // Recognize patterns
3563 switch (I.getPredicate()) {
3564 case ICmpInst::ICMP_EQ:
3565 case ICmpInst::ICMP_NE:
3566 // Recognize pattern:
3567 // mulval = mul(zext A, zext B)
3568 // cmp eq/neq mulval, zext trunc mulval
3569 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
3570 if (Zext->hasOneUse()) {
3571 Value *ZextArg = Zext->getOperand(0);
3572 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
3573 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
3577 // Recognize pattern:
3578 // mulval = mul(zext A, zext B)
3579 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
3582 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
3583 if (ValToMask != MulVal)
3585 const APInt &CVal = CI->getValue() + 1;
3586 if (CVal.isPowerOf2()) {
3587 unsigned MaskWidth = CVal.logBase2();
3588 if (MaskWidth == MulWidth)
3589 break; // Recognized
3594 case ICmpInst::ICMP_UGT:
3595 // Recognize pattern:
3596 // mulval = mul(zext A, zext B)
3597 // cmp ugt mulval, max
3598 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3599 APInt MaxVal = APInt::getMaxValue(MulWidth);
3600 MaxVal = MaxVal.zext(CI->getBitWidth());
3601 if (MaxVal.eq(CI->getValue()))
3602 break; // Recognized
3606 case ICmpInst::ICMP_UGE:
3607 // Recognize pattern:
3608 // mulval = mul(zext A, zext B)
3609 // cmp uge mulval, max+1
3610 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3611 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
3612 if (MaxVal.eq(CI->getValue()))
3613 break; // Recognized
3617 case ICmpInst::ICMP_ULE:
3618 // Recognize pattern:
3619 // mulval = mul(zext A, zext B)
3620 // cmp ule mulval, max
3621 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3622 APInt MaxVal = APInt::getMaxValue(MulWidth);
3623 MaxVal = MaxVal.zext(CI->getBitWidth());
3624 if (MaxVal.eq(CI->getValue()))
3625 break; // Recognized
3629 case ICmpInst::ICMP_ULT:
3630 // Recognize pattern:
3631 // mulval = mul(zext A, zext B)
3632 // cmp ule mulval, max + 1
3633 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
3634 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
3635 if (MaxVal.eq(CI->getValue()))
3636 break; // Recognized
3644 InstCombiner::BuilderTy *Builder = IC.Builder;
3645 Builder->SetInsertPoint(MulInstr);
3647 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
3648 Value *MulA = A, *MulB = B;
3649 if (WidthA < MulWidth)
3650 MulA = Builder->CreateZExt(A, MulType);
3651 if (WidthB < MulWidth)
3652 MulB = Builder->CreateZExt(B, MulType);
3653 Value *F = Intrinsic::getDeclaration(I.getModule(),
3654 Intrinsic::umul_with_overflow, MulType);
3655 CallInst *Call = Builder->CreateCall(F, {MulA, MulB}, "umul");
3656 IC.Worklist.Add(MulInstr);
3658 // If there are uses of mul result other than the comparison, we know that
3659 // they are truncation or binary AND. Change them to use result of
3660 // mul.with.overflow and adjust properly mask/size.
3661 if (MulVal->hasNUsesOrMore(2)) {
3662 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
3663 for (User *U : MulVal->users()) {
3664 if (U == &I || U == OtherVal)
3666 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3667 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
3668 IC.replaceInstUsesWith(*TI, Mul);
3670 TI->setOperand(0, Mul);
3671 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3672 assert(BO->getOpcode() == Instruction::And);
3673 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
3674 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
3675 APInt ShortMask = CI->getValue().trunc(MulWidth);
3676 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
3678 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
3679 IC.Worklist.Add(Zext);
3680 IC.replaceInstUsesWith(*BO, Zext);
3682 llvm_unreachable("Unexpected Binary operation");
3684 IC.Worklist.Add(cast<Instruction>(U));
3687 if (isa<Instruction>(OtherVal))
3688 IC.Worklist.Add(cast<Instruction>(OtherVal));
3690 // The original icmp gets replaced with the overflow value, maybe inverted
3691 // depending on predicate.
3692 bool Inverse = false;
3693 switch (I.getPredicate()) {
3694 case ICmpInst::ICMP_NE:
3696 case ICmpInst::ICMP_EQ:
3699 case ICmpInst::ICMP_UGT:
3700 case ICmpInst::ICMP_UGE:
3701 if (I.getOperand(0) == MulVal)
3705 case ICmpInst::ICMP_ULT:
3706 case ICmpInst::ICMP_ULE:
3707 if (I.getOperand(1) == MulVal)
3712 llvm_unreachable("Unexpected predicate");
3715 Value *Res = Builder->CreateExtractValue(Call, 1);
3716 return BinaryOperator::CreateNot(Res);
3719 return ExtractValueInst::Create(Call, 1);
3722 /// When performing a comparison against a constant, it is possible that not all
3723 /// the bits in the LHS are demanded. This helper method computes the mask that
3725 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth,
3728 return APInt::getSignBit(BitWidth);
3730 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
3731 if (!CI) return APInt::getAllOnesValue(BitWidth);
3732 const APInt &RHS = CI->getValue();
3734 switch (I.getPredicate()) {
3735 // For a UGT comparison, we don't care about any bits that
3736 // correspond to the trailing ones of the comparand. The value of these
3737 // bits doesn't impact the outcome of the comparison, because any value
3738 // greater than the RHS must differ in a bit higher than these due to carry.
3739 case ICmpInst::ICMP_UGT: {
3740 unsigned trailingOnes = RHS.countTrailingOnes();
3741 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
3745 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
3746 // Any value less than the RHS must differ in a higher bit because of carries.
3747 case ICmpInst::ICMP_ULT: {
3748 unsigned trailingZeros = RHS.countTrailingZeros();
3749 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
3754 return APInt::getAllOnesValue(BitWidth);
3758 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
3759 /// should be swapped.
3760 /// The decision is based on how many times these two operands are reused
3761 /// as subtract operands and their positions in those instructions.
3762 /// The rational is that several architectures use the same instruction for
3763 /// both subtract and cmp, thus it is better if the order of those operands
3765 /// \return true if Op0 and Op1 should be swapped.
3766 static bool swapMayExposeCSEOpportunities(const Value * Op0,
3767 const Value * Op1) {
3768 // Filter out pointer value as those cannot appears directly in subtract.
3769 // FIXME: we may want to go through inttoptrs or bitcasts.
3770 if (Op0->getType()->isPointerTy())
3772 // Count every uses of both Op0 and Op1 in a subtract.
3773 // Each time Op0 is the first operand, count -1: swapping is bad, the
3774 // subtract has already the same layout as the compare.
3775 // Each time Op0 is the second operand, count +1: swapping is good, the
3776 // subtract has a different layout as the compare.
3777 // At the end, if the benefit is greater than 0, Op0 should come second to
3778 // expose more CSE opportunities.
3779 int GlobalSwapBenefits = 0;
3780 for (const User *U : Op0->users()) {
3781 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
3782 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
3784 // If Op0 is the first argument, this is not beneficial to swap the
3786 int LocalSwapBenefits = -1;
3787 unsigned Op1Idx = 1;
3788 if (BinOp->getOperand(Op1Idx) == Op0) {
3790 LocalSwapBenefits = 1;
3792 if (BinOp->getOperand(Op1Idx) != Op1)
3794 GlobalSwapBenefits += LocalSwapBenefits;
3796 return GlobalSwapBenefits > 0;
3799 /// \brief Check that one use is in the same block as the definition and all
3800 /// other uses are in blocks dominated by a given block.
3802 /// \param DI Definition
3804 /// \param DB Block that must dominate all uses of \p DI outside
3805 /// the parent block
3806 /// \return true when \p UI is the only use of \p DI in the parent block
3807 /// and all other uses of \p DI are in blocks dominated by \p DB.
3809 bool InstCombiner::dominatesAllUses(const Instruction *DI,
3810 const Instruction *UI,
3811 const BasicBlock *DB) const {
3812 assert(DI && UI && "Instruction not defined\n");
3813 // Ignore incomplete definitions.
3814 if (!DI->getParent())
3816 // DI and UI must be in the same block.
3817 if (DI->getParent() != UI->getParent())
3819 // Protect from self-referencing blocks.
3820 if (DI->getParent() == DB)
3822 for (const User *U : DI->users()) {
3823 auto *Usr = cast<Instruction>(U);
3824 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
3830 /// Return true when the instruction sequence within a block is select-cmp-br.
3831 static bool isChainSelectCmpBranch(const SelectInst *SI) {
3832 const BasicBlock *BB = SI->getParent();
3835 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
3836 if (!BI || BI->getNumSuccessors() != 2)
3838 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
3839 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
3844 /// \brief True when a select result is replaced by one of its operands
3845 /// in select-icmp sequence. This will eventually result in the elimination
3848 /// \param SI Select instruction
3849 /// \param Icmp Compare instruction
3850 /// \param SIOpd Operand that replaces the select
3853 /// - The replacement is global and requires dominator information
3854 /// - The caller is responsible for the actual replacement
3859 /// %4 = select i1 %3, %C* %0, %C* null
3860 /// %5 = icmp eq %C* %4, null
3861 /// br i1 %5, label %9, label %7
3863 /// ; <label>:7 ; preds = %entry
3864 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
3867 /// can be transformed to
3869 /// %5 = icmp eq %C* %0, null
3870 /// %6 = select i1 %3, i1 %5, i1 true
3871 /// br i1 %6, label %9, label %7
3873 /// ; <label>:7 ; preds = %entry
3874 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
3876 /// Similar when the first operand of the select is a constant or/and
3877 /// the compare is for not equal rather than equal.
3879 /// NOTE: The function is only called when the select and compare constants
3880 /// are equal, the optimization can work only for EQ predicates. This is not a
3881 /// major restriction since a NE compare should be 'normalized' to an equal
3882 /// compare, which usually happens in the combiner and test case
3883 /// select-cmp-br.ll checks for it.
3884 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
3885 const ICmpInst *Icmp,
3886 const unsigned SIOpd) {
3887 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
3888 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
3889 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
3890 // The check for the unique predecessor is not the best that can be
3891 // done. But it protects efficiently against cases like when SI's
3892 // home block has two successors, Succ and Succ1, and Succ1 predecessor
3893 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
3894 // replaced can be reached on either path. So the uniqueness check
3895 // guarantees that the path all uses of SI (outside SI's parent) are on
3896 // is disjoint from all other paths out of SI. But that information
3897 // is more expensive to compute, and the trade-off here is in favor
3899 if (Succ->getUniquePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
3901 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
3908 /// Try to fold the comparison based on range information we can get by checking
3909 /// whether bits are known to be zero or one in the inputs.
3910 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
3911 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3912 Type *Ty = Op0->getType();
3913 ICmpInst::Predicate Pred = I.getPredicate();
3915 // Get scalar or pointer size.
3916 unsigned BitWidth = Ty->isIntOrIntVectorTy()
3917 ? Ty->getScalarSizeInBits()
3918 : DL.getTypeSizeInBits(Ty->getScalarType());
3923 // If this is a normal comparison, it demands all bits. If it is a sign bit
3924 // comparison, it only demands the sign bit.
3925 bool IsSignBit = false;
3927 if (match(Op1, m_APInt(CmpC))) {
3929 IsSignBit = isSignBitCheck(Pred, *CmpC, UnusedBit);
3932 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
3933 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
3935 if (SimplifyDemandedBits(I.getOperandUse(0),
3936 getDemandedBitsLHSMask(I, BitWidth, IsSignBit),
3937 Op0KnownZero, Op0KnownOne, 0))
3940 if (SimplifyDemandedBits(I.getOperandUse(1), APInt::getAllOnesValue(BitWidth),
3941 Op1KnownZero, Op1KnownOne, 0))
3944 // Given the known and unknown bits, compute a range that the LHS could be
3945 // in. Compute the Min, Max and RHS values based on the known bits. For the
3946 // EQ and NE we use unsigned values.
3947 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
3948 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
3950 computeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, Op0Min,
3952 computeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, Op1Min,
3955 computeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, Op0Min,
3957 computeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, Op1Min,
3961 // If Min and Max are known to be the same, then SimplifyDemandedBits
3962 // figured out that the LHS is a constant. Constant fold this now, so that
3963 // code below can assume that Min != Max.
3964 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
3965 return new ICmpInst(Pred, ConstantInt::get(Op0->getType(), Op0Min), Op1);
3966 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
3967 return new ICmpInst(Pred, Op0, ConstantInt::get(Op1->getType(), Op1Min));
3969 // Based on the range information we know about the LHS, see if we can
3970 // simplify this comparison. For example, (x&4) < 8 is always true.
3973 llvm_unreachable("Unknown icmp opcode!");
3974 case ICmpInst::ICMP_EQ:
3975 case ICmpInst::ICMP_NE: {
3976 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
3977 return Pred == CmpInst::ICMP_EQ
3978 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
3979 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3982 // If all bits are known zero except for one, then we know at most one bit
3983 // is set. If the comparison is against zero, then this is a check to see if
3984 // *that* bit is set.
3985 APInt Op0KnownZeroInverted = ~Op0KnownZero;
3986 if (~Op1KnownZero == 0) {
3987 // If the LHS is an AND with the same constant, look through it.
3988 Value *LHS = nullptr;
3990 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
3991 *LHSC != Op0KnownZeroInverted)
3995 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
3996 APInt ValToCheck = Op0KnownZeroInverted;
3997 Type *XTy = X->getType();
3998 if (ValToCheck.isPowerOf2()) {
3999 // ((1 << X) & 8) == 0 -> X != 3
4000 // ((1 << X) & 8) != 0 -> X == 3
4001 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4002 auto NewPred = ICmpInst::getInversePredicate(Pred);
4003 return new ICmpInst(NewPred, X, CmpC);
4004 } else if ((++ValToCheck).isPowerOf2()) {
4005 // ((1 << X) & 7) == 0 -> X >= 3
4006 // ((1 << X) & 7) != 0 -> X < 3
4007 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4009 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
4010 return new ICmpInst(NewPred, X, CmpC);
4014 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4016 if (Op0KnownZeroInverted == 1 &&
4017 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4018 // ((8 >>u X) & 1) == 0 -> X != 3
4019 // ((8 >>u X) & 1) != 0 -> X == 3
4020 unsigned CmpVal = CI->countTrailingZeros();
4021 auto NewPred = ICmpInst::getInversePredicate(Pred);
4022 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4027 case ICmpInst::ICMP_ULT: {
4028 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4029 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4030 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4031 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4032 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4033 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4036 if (match(Op1, m_APInt(CmpC))) {
4037 // A <u C -> A == C-1 if min(A)+1 == C
4038 if (Op1Max == Op0Min + 1) {
4039 Constant *CMinus1 = ConstantInt::get(Op0->getType(), *CmpC - 1);
4040 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, CMinus1);
4042 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4043 if (CmpC->isMinSignedValue()) {
4044 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
4045 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
4050 case ICmpInst::ICMP_UGT: {
4051 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4052 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4054 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4055 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4057 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4058 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4061 if (match(Op1, m_APInt(CmpC))) {
4062 // A >u C -> A == C+1 if max(a)-1 == C
4063 if (*CmpC == Op0Max - 1)
4064 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4065 ConstantInt::get(Op1->getType(), *CmpC + 1));
4067 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4068 if (CmpC->isMaxSignedValue())
4069 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4070 Constant::getNullValue(Op0->getType()));
4074 case ICmpInst::ICMP_SLT:
4075 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4076 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4077 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4078 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4079 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4080 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4081 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4082 if (Op1Max == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4083 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4084 Builder->getInt(CI->getValue() - 1));
4087 case ICmpInst::ICMP_SGT:
4088 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4089 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4090 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4091 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4093 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4094 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4095 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4096 if (Op1Min == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4097 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4098 Builder->getInt(CI->getValue() + 1));
4101 case ICmpInst::ICMP_SGE:
4102 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4103 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4104 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4105 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4106 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4108 case ICmpInst::ICMP_SLE:
4109 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4110 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4111 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4112 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4113 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4115 case ICmpInst::ICMP_UGE:
4116 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4117 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4118 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4119 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4120 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4122 case ICmpInst::ICMP_ULE:
4123 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4124 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4125 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4126 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4127 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4131 // Turn a signed comparison into an unsigned one if both operands are known to
4132 // have the same sign.
4134 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
4135 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
4136 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4141 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4142 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4143 /// allows them to be folded in visitICmpInst.
4144 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
4145 ICmpInst::Predicate Pred = I.getPredicate();
4146 if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4147 Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4150 Value *Op0 = I.getOperand(0);
4151 Value *Op1 = I.getOperand(1);
4152 auto *Op1C = dyn_cast<Constant>(Op1);
4156 // Check if the constant operand can be safely incremented/decremented without
4157 // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4158 // the edge cases for us, so we just assert on them. For vectors, we must
4159 // handle the edge cases.
4160 Type *Op1Type = Op1->getType();
4161 bool IsSigned = I.isSigned();
4162 bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4163 auto *CI = dyn_cast<ConstantInt>(Op1C);
4165 // A <= MAX -> TRUE ; A >= MIN -> TRUE
4166 assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4167 } else if (Op1Type->isVectorTy()) {
4168 // TODO? If the edge cases for vectors were guaranteed to be handled as they
4169 // are for scalar, we could remove the min/max checks. However, to do that,
4170 // we would have to use insertelement/shufflevector to replace edge values.
4171 unsigned NumElts = Op1Type->getVectorNumElements();
4172 for (unsigned i = 0; i != NumElts; ++i) {
4173 Constant *Elt = Op1C->getAggregateElement(i);
4177 if (isa<UndefValue>(Elt))
4180 // Bail out if we can't determine if this constant is min/max or if we
4181 // know that this constant is min/max.
4182 auto *CI = dyn_cast<ConstantInt>(Elt);
4183 if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4191 // Increment or decrement the constant and set the new comparison predicate:
4192 // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4193 Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4194 CmpInst::Predicate NewPred = IsLE ? ICmpInst::ICMP_ULT: ICmpInst::ICMP_UGT;
4195 NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4196 return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4199 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4200 bool Changed = false;
4201 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4202 unsigned Op0Cplxity = getComplexity(Op0);
4203 unsigned Op1Cplxity = getComplexity(Op1);
4205 /// Orders the operands of the compare so that they are listed from most
4206 /// complex to least complex. This puts constants before unary operators,
4207 /// before binary operators.
4208 if (Op0Cplxity < Op1Cplxity ||
4209 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
4211 std::swap(Op0, Op1);
4216 SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, &TLI, &DT, &AC, &I))
4217 return replaceInstUsesWith(I, V);
4219 // comparing -val or val with non-zero is the same as just comparing val
4220 // ie, abs(val) != 0 -> val != 0
4221 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
4222 Value *Cond, *SelectTrue, *SelectFalse;
4223 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
4224 m_Value(SelectFalse)))) {
4225 if (Value *V = dyn_castNegVal(SelectTrue)) {
4226 if (V == SelectFalse)
4227 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4229 else if (Value *V = dyn_castNegVal(SelectFalse)) {
4230 if (V == SelectTrue)
4231 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4236 Type *Ty = Op0->getType();
4238 // icmp's with boolean values can always be turned into bitwise operations
4239 if (Ty->getScalarType()->isIntegerTy(1)) {
4240 switch (I.getPredicate()) {
4241 default: llvm_unreachable("Invalid icmp instruction!");
4242 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
4243 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName() + "tmp");
4244 return BinaryOperator::CreateNot(Xor);
4246 case ICmpInst::ICMP_NE: // icmp ne i1 A, B -> A^B
4247 return BinaryOperator::CreateXor(Op0, Op1);
4249 case ICmpInst::ICMP_UGT:
4250 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
4252 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
4253 Value *Not = Builder->CreateNot(Op0, I.getName() + "tmp");
4254 return BinaryOperator::CreateAnd(Not, Op1);
4256 case ICmpInst::ICMP_SGT:
4257 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
4259 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
4260 Value *Not = Builder->CreateNot(Op1, I.getName() + "tmp");
4261 return BinaryOperator::CreateAnd(Not, Op0);
4263 case ICmpInst::ICMP_UGE:
4264 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
4266 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
4267 Value *Not = Builder->CreateNot(Op0, I.getName() + "tmp");
4268 return BinaryOperator::CreateOr(Not, Op1);
4270 case ICmpInst::ICMP_SGE:
4271 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
4273 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
4274 Value *Not = Builder->CreateNot(Op1, I.getName() + "tmp");
4275 return BinaryOperator::CreateOr(Not, Op0);
4280 if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
4283 if (Instruction *Res = foldICmpWithConstant(I))
4286 if (Instruction *Res = foldICmpUsingKnownBits(I))
4289 // Test if the ICmpInst instruction is used exclusively by a select as
4290 // part of a minimum or maximum operation. If so, refrain from doing
4291 // any other folding. This helps out other analyses which understand
4292 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4293 // and CodeGen. And in this case, at least one of the comparison
4294 // operands has at least one user besides the compare (the select),
4295 // which would often largely negate the benefit of folding anyway.
4297 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
4298 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
4299 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
4302 if (Instruction *Res = foldICmpInstWithConstant(I))
4305 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
4308 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4309 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
4310 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
4312 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
4313 if (Instruction *NI = foldGEPICmp(GEP, Op0,
4314 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
4317 // Try to optimize equality comparisons against alloca-based pointers.
4318 if (Op0->getType()->isPointerTy() && I.isEquality()) {
4319 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
4320 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
4321 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
4323 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
4324 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
4328 // Test to see if the operands of the icmp are casted versions of other
4329 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4331 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4332 if (Op0->getType()->isPointerTy() &&
4333 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4334 // We keep moving the cast from the left operand over to the right
4335 // operand, where it can often be eliminated completely.
4336 Op0 = CI->getOperand(0);
4338 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
4339 // so eliminate it as well.
4340 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
4341 Op1 = CI2->getOperand(0);
4343 // If Op1 is a constant, we can fold the cast into the constant.
4344 if (Op0->getType() != Op1->getType()) {
4345 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4346 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4348 // Otherwise, cast the RHS right before the icmp
4349 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
4352 return new ICmpInst(I.getPredicate(), Op0, Op1);
4356 if (isa<CastInst>(Op0)) {
4357 // Handle the special case of: icmp (cast bool to X), <cst>
4358 // This comes up when you have code like
4361 // For generality, we handle any zero-extension of any operand comparison
4362 // with a constant or another cast from the same type.
4363 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
4364 if (Instruction *R = foldICmpWithCastAndCast(I))
4368 if (Instruction *Res = foldICmpBinOp(I))
4371 if (Instruction *Res = foldICmpWithMinMax(I))
4376 // Transform (A & ~B) == 0 --> (A & B) != 0
4377 // and (A & ~B) != 0 --> (A & B) == 0
4378 // if A is a power of 2.
4379 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
4380 match(Op1, m_Zero()) &&
4381 isKnownToBeAPowerOfTwo(A, DL, false, 0, &AC, &I, &DT) && I.isEquality())
4382 return new ICmpInst(I.getInversePredicate(),
4383 Builder->CreateAnd(A, B),
4386 // ~x < ~y --> y < x
4387 // ~x < cst --> ~cst < x
4388 if (match(Op0, m_Not(m_Value(A)))) {
4389 if (match(Op1, m_Not(m_Value(B))))
4390 return new ICmpInst(I.getPredicate(), B, A);
4391 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
4392 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
4395 Instruction *AddI = nullptr;
4396 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
4397 m_Instruction(AddI))) &&
4398 isa<IntegerType>(A->getType())) {
4401 if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result,
4403 replaceInstUsesWith(*AddI, Result);
4404 return replaceInstUsesWith(I, Overflow);
4408 // (zext a) * (zext b) --> llvm.umul.with.overflow.
4409 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4410 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
4413 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4414 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
4419 if (Instruction *Res = foldICmpEquality(I))
4422 // The 'cmpxchg' instruction returns an aggregate containing the old value and
4423 // an i1 which indicates whether or not we successfully did the swap.
4425 // Replace comparisons between the old value and the expected value with the
4426 // indicator that 'cmpxchg' returns.
4428 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
4429 // spuriously fail. In those cases, the old value may equal the expected
4430 // value but it is possible for the swap to not occur.
4431 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4432 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
4433 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
4434 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
4436 return ExtractValueInst::Create(ACXI, 1);
4439 Value *X; ConstantInt *Cst;
4441 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
4442 return foldICmpAddOpConst(I, X, Cst, I.getPredicate());
4445 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
4446 return foldICmpAddOpConst(I, X, Cst, I.getSwappedPredicate());
4448 return Changed ? &I : nullptr;
4451 /// Fold fcmp ([us]itofp x, cst) if possible.
4452 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
4454 if (!isa<ConstantFP>(RHSC)) return nullptr;
4455 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4457 // Get the width of the mantissa. We don't want to hack on conversions that
4458 // might lose information from the integer, e.g. "i64 -> float"
4459 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4460 if (MantissaWidth == -1) return nullptr; // Unknown.
4462 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4464 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
4466 if (I.isEquality()) {
4467 FCmpInst::Predicate P = I.getPredicate();
4468 bool IsExact = false;
4469 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
4470 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
4472 // If the floating point constant isn't an integer value, we know if we will
4473 // ever compare equal / not equal to it.
4475 // TODO: Can never be -0.0 and other non-representable values
4476 APFloat RHSRoundInt(RHS);
4477 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
4478 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
4479 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
4480 return replaceInstUsesWith(I, Builder->getFalse());
4482 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
4483 return replaceInstUsesWith(I, Builder->getTrue());
4487 // TODO: If the constant is exactly representable, is it always OK to do
4488 // equality compares as integer?
4491 // Check to see that the input is converted from an integer type that is small
4492 // enough that preserves all bits. TODO: check here for "known" sign bits.
4493 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4494 unsigned InputSize = IntTy->getScalarSizeInBits();
4496 // Following test does NOT adjust InputSize downwards for signed inputs,
4497 // because the most negative value still requires all the mantissa bits
4498 // to distinguish it from one less than that value.
4499 if ((int)InputSize > MantissaWidth) {
4500 // Conversion would lose accuracy. Check if loss can impact comparison.
4501 int Exp = ilogb(RHS);
4502 if (Exp == APFloat::IEK_Inf) {
4503 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
4504 if (MaxExponent < (int)InputSize - !LHSUnsigned)
4505 // Conversion could create infinity.
4508 // Note that if RHS is zero or NaN, then Exp is negative
4509 // and first condition is trivially false.
4510 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
4511 // Conversion could affect comparison.
4516 // Otherwise, we can potentially simplify the comparison. We know that it
4517 // will always come through as an integer value and we know the constant is
4518 // not a NAN (it would have been previously simplified).
4519 assert(!RHS.isNaN() && "NaN comparison not already folded!");
4521 ICmpInst::Predicate Pred;
4522 switch (I.getPredicate()) {
4523 default: llvm_unreachable("Unexpected predicate!");
4524 case FCmpInst::FCMP_UEQ:
4525 case FCmpInst::FCMP_OEQ:
4526 Pred = ICmpInst::ICMP_EQ;
4528 case FCmpInst::FCMP_UGT:
4529 case FCmpInst::FCMP_OGT:
4530 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
4532 case FCmpInst::FCMP_UGE:
4533 case FCmpInst::FCMP_OGE:
4534 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
4536 case FCmpInst::FCMP_ULT:
4537 case FCmpInst::FCMP_OLT:
4538 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
4540 case FCmpInst::FCMP_ULE:
4541 case FCmpInst::FCMP_OLE:
4542 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
4544 case FCmpInst::FCMP_UNE:
4545 case FCmpInst::FCMP_ONE:
4546 Pred = ICmpInst::ICMP_NE;
4548 case FCmpInst::FCMP_ORD:
4549 return replaceInstUsesWith(I, Builder->getTrue());
4550 case FCmpInst::FCMP_UNO:
4551 return replaceInstUsesWith(I, Builder->getFalse());
4554 // Now we know that the APFloat is a normal number, zero or inf.
4556 // See if the FP constant is too large for the integer. For example,
4557 // comparing an i8 to 300.0.
4558 unsigned IntWidth = IntTy->getScalarSizeInBits();
4561 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4562 // and large values.
4563 APFloat SMax(RHS.getSemantics());
4564 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4565 APFloat::rmNearestTiesToEven);
4566 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4567 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4568 Pred == ICmpInst::ICMP_SLE)
4569 return replaceInstUsesWith(I, Builder->getTrue());
4570 return replaceInstUsesWith(I, Builder->getFalse());
4573 // If the RHS value is > UnsignedMax, fold the comparison. This handles
4574 // +INF and large values.
4575 APFloat UMax(RHS.getSemantics());
4576 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
4577 APFloat::rmNearestTiesToEven);
4578 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
4579 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
4580 Pred == ICmpInst::ICMP_ULE)
4581 return replaceInstUsesWith(I, Builder->getTrue());
4582 return replaceInstUsesWith(I, Builder->getFalse());
4587 // See if the RHS value is < SignedMin.
4588 APFloat SMin(RHS.getSemantics());
4589 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
4590 APFloat::rmNearestTiesToEven);
4591 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
4592 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
4593 Pred == ICmpInst::ICMP_SGE)
4594 return replaceInstUsesWith(I, Builder->getTrue());
4595 return replaceInstUsesWith(I, Builder->getFalse());
4598 // See if the RHS value is < UnsignedMin.
4599 APFloat SMin(RHS.getSemantics());
4600 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
4601 APFloat::rmNearestTiesToEven);
4602 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
4603 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
4604 Pred == ICmpInst::ICMP_UGE)
4605 return replaceInstUsesWith(I, Builder->getTrue());
4606 return replaceInstUsesWith(I, Builder->getFalse());
4610 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
4611 // [0, UMAX], but it may still be fractional. See if it is fractional by
4612 // casting the FP value to the integer value and back, checking for equality.
4613 // Don't do this for zero, because -0.0 is not fractional.
4614 Constant *RHSInt = LHSUnsigned
4615 ? ConstantExpr::getFPToUI(RHSC, IntTy)
4616 : ConstantExpr::getFPToSI(RHSC, IntTy);
4617 if (!RHS.isZero()) {
4618 bool Equal = LHSUnsigned
4619 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
4620 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
4622 // If we had a comparison against a fractional value, we have to adjust
4623 // the compare predicate and sometimes the value. RHSC is rounded towards
4624 // zero at this point.
4626 default: llvm_unreachable("Unexpected integer comparison!");
4627 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
4628 return replaceInstUsesWith(I, Builder->getTrue());
4629 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
4630 return replaceInstUsesWith(I, Builder->getFalse());
4631 case ICmpInst::ICMP_ULE:
4632 // (float)int <= 4.4 --> int <= 4
4633 // (float)int <= -4.4 --> false
4634 if (RHS.isNegative())
4635 return replaceInstUsesWith(I, Builder->getFalse());
4637 case ICmpInst::ICMP_SLE:
4638 // (float)int <= 4.4 --> int <= 4
4639 // (float)int <= -4.4 --> int < -4
4640 if (RHS.isNegative())
4641 Pred = ICmpInst::ICMP_SLT;
4643 case ICmpInst::ICMP_ULT:
4644 // (float)int < -4.4 --> false
4645 // (float)int < 4.4 --> int <= 4
4646 if (RHS.isNegative())
4647 return replaceInstUsesWith(I, Builder->getFalse());
4648 Pred = ICmpInst::ICMP_ULE;
4650 case ICmpInst::ICMP_SLT:
4651 // (float)int < -4.4 --> int < -4
4652 // (float)int < 4.4 --> int <= 4
4653 if (!RHS.isNegative())
4654 Pred = ICmpInst::ICMP_SLE;
4656 case ICmpInst::ICMP_UGT:
4657 // (float)int > 4.4 --> int > 4
4658 // (float)int > -4.4 --> true
4659 if (RHS.isNegative())
4660 return replaceInstUsesWith(I, Builder->getTrue());
4662 case ICmpInst::ICMP_SGT:
4663 // (float)int > 4.4 --> int > 4
4664 // (float)int > -4.4 --> int >= -4
4665 if (RHS.isNegative())
4666 Pred = ICmpInst::ICMP_SGE;
4668 case ICmpInst::ICMP_UGE:
4669 // (float)int >= -4.4 --> true
4670 // (float)int >= 4.4 --> int > 4
4671 if (RHS.isNegative())
4672 return replaceInstUsesWith(I, Builder->getTrue());
4673 Pred = ICmpInst::ICMP_UGT;
4675 case ICmpInst::ICMP_SGE:
4676 // (float)int >= -4.4 --> int >= -4
4677 // (float)int >= 4.4 --> int > 4
4678 if (!RHS.isNegative())
4679 Pred = ICmpInst::ICMP_SGT;
4685 // Lower this FP comparison into an appropriate integer version of the
4687 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
4690 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4691 bool Changed = false;
4693 /// Orders the operands of the compare so that they are listed from most
4694 /// complex to least complex. This puts constants before unary operators,
4695 /// before binary operators.
4696 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
4701 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4703 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1,
4704 I.getFastMathFlags(), DL, &TLI, &DT, &AC, &I))
4705 return replaceInstUsesWith(I, V);
4707 // Simplify 'fcmp pred X, X'
4709 switch (I.getPredicate()) {
4710 default: llvm_unreachable("Unknown predicate!");
4711 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4712 case FCmpInst::FCMP_ULT: // True if unordered or less than
4713 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4714 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4715 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4716 I.setPredicate(FCmpInst::FCMP_UNO);
4717 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4720 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4721 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4722 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4723 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4724 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4725 I.setPredicate(FCmpInst::FCMP_ORD);
4726 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4731 // Test if the FCmpInst instruction is used exclusively by a select as
4732 // part of a minimum or maximum operation. If so, refrain from doing
4733 // any other folding. This helps out other analyses which understand
4734 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4735 // and CodeGen. And in this case, at least one of the comparison
4736 // operands has at least one user besides the compare (the select),
4737 // which would often largely negate the benefit of folding anyway.
4739 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
4740 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
4741 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
4744 // Handle fcmp with constant RHS
4745 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4746 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4747 switch (LHSI->getOpcode()) {
4748 case Instruction::FPExt: {
4749 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
4750 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
4751 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
4755 const fltSemantics *Sem;
4756 // FIXME: This shouldn't be here.
4757 if (LHSExt->getSrcTy()->isHalfTy())
4758 Sem = &APFloat::IEEEhalf();
4759 else if (LHSExt->getSrcTy()->isFloatTy())
4760 Sem = &APFloat::IEEEsingle();
4761 else if (LHSExt->getSrcTy()->isDoubleTy())
4762 Sem = &APFloat::IEEEdouble();
4763 else if (LHSExt->getSrcTy()->isFP128Ty())
4764 Sem = &APFloat::IEEEquad();
4765 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
4766 Sem = &APFloat::x87DoubleExtended();
4767 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
4768 Sem = &APFloat::PPCDoubleDouble();
4773 APFloat F = RHSF->getValueAPF();
4774 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
4776 // Avoid lossy conversions and denormals. Zero is a special case
4777 // that's OK to convert.
4781 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
4782 APFloat::cmpLessThan) || Fabs.isZero()))
4784 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
4785 ConstantFP::get(RHSC->getContext(), F));
4788 case Instruction::PHI:
4789 // Only fold fcmp into the PHI if the phi and fcmp are in the same
4790 // block. If in the same block, we're encouraging jump threading. If
4791 // not, we are just pessimizing the code by making an i1 phi.
4792 if (LHSI->getParent() == I.getParent())
4793 if (Instruction *NV = FoldOpIntoPhi(I))
4796 case Instruction::SIToFP:
4797 case Instruction::UIToFP:
4798 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
4801 case Instruction::FSub: {
4802 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
4804 if (match(LHSI, m_FNeg(m_Value(Op))))
4805 return new FCmpInst(I.getSwappedPredicate(), Op,
4806 ConstantExpr::getFNeg(RHSC));
4809 case Instruction::Load:
4810 if (GetElementPtrInst *GEP =
4811 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
4812 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
4813 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
4814 !cast<LoadInst>(LHSI)->isVolatile())
4815 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
4819 case Instruction::Call: {
4820 if (!RHSC->isNullValue())
4823 CallInst *CI = cast<CallInst>(LHSI);
4824 Intrinsic::ID IID = getIntrinsicForCallSite(CI, &TLI);
4825 if (IID != Intrinsic::fabs)
4828 // Various optimization for fabs compared with zero.
4829 switch (I.getPredicate()) {
4832 // fabs(x) < 0 --> false
4833 case FCmpInst::FCMP_OLT:
4834 llvm_unreachable("handled by SimplifyFCmpInst");
4835 // fabs(x) > 0 --> x != 0
4836 case FCmpInst::FCMP_OGT:
4837 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), RHSC);
4838 // fabs(x) <= 0 --> x == 0
4839 case FCmpInst::FCMP_OLE:
4840 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), RHSC);
4841 // fabs(x) >= 0 --> !isnan(x)
4842 case FCmpInst::FCMP_OGE:
4843 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), RHSC);
4844 // fabs(x) == 0 --> x == 0
4845 // fabs(x) != 0 --> x != 0
4846 case FCmpInst::FCMP_OEQ:
4847 case FCmpInst::FCMP_UEQ:
4848 case FCmpInst::FCMP_ONE:
4849 case FCmpInst::FCMP_UNE:
4850 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), RHSC);
4856 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
4858 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
4859 return new FCmpInst(I.getSwappedPredicate(), X, Y);
4861 // fcmp (fpext x), (fpext y) -> fcmp x, y
4862 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
4863 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
4864 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
4865 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
4866 RHSExt->getOperand(0));
4868 return Changed ? &I : nullptr;