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/Statistic.h"
17 #include "llvm/Analysis/ConstantFolding.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/MemoryBuiltins.h"
20 #include "llvm/IR/ConstantRange.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/GetElementPtrTypeIterator.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/PatternMatch.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Analysis/TargetLibraryInfo.h"
30 using namespace PatternMatch;
32 #define DEBUG_TYPE "instcombine"
34 // How many times is a select replaced by one of its operands?
35 STATISTIC(NumSel, "Number of select opts");
37 // Initialization Routines
39 static ConstantInt *getOne(Constant *C) {
40 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
43 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
44 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
47 static bool HasAddOverflow(ConstantInt *Result,
48 ConstantInt *In1, ConstantInt *In2,
51 return Result->getValue().ult(In1->getValue());
53 if (In2->isNegative())
54 return Result->getValue().sgt(In1->getValue());
55 return Result->getValue().slt(In1->getValue());
58 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
59 /// overflowed for this type.
60 static bool AddWithOverflow(Constant *&Result, Constant *In1,
61 Constant *In2, bool IsSigned = false) {
62 Result = ConstantExpr::getAdd(In1, In2);
64 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
65 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
66 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
67 if (HasAddOverflow(ExtractElement(Result, Idx),
68 ExtractElement(In1, Idx),
69 ExtractElement(In2, Idx),
76 return HasAddOverflow(cast<ConstantInt>(Result),
77 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
81 static bool HasSubOverflow(ConstantInt *Result,
82 ConstantInt *In1, ConstantInt *In2,
85 return Result->getValue().ugt(In1->getValue());
87 if (In2->isNegative())
88 return Result->getValue().slt(In1->getValue());
90 return Result->getValue().sgt(In1->getValue());
93 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
94 /// overflowed for this type.
95 static bool SubWithOverflow(Constant *&Result, Constant *In1,
96 Constant *In2, bool IsSigned = false) {
97 Result = ConstantExpr::getSub(In1, In2);
99 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
100 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
101 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
102 if (HasSubOverflow(ExtractElement(Result, Idx),
103 ExtractElement(In1, Idx),
104 ExtractElement(In2, Idx),
111 return HasSubOverflow(cast<ConstantInt>(Result),
112 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
116 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
117 /// comparison only checks the sign bit. If it only checks the sign bit, set
118 /// TrueIfSigned if the result of the comparison is true when the input value is
120 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
121 bool &TrueIfSigned) {
123 case ICmpInst::ICMP_SLT: // True if LHS s< 0
125 return RHS->isZero();
126 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
128 return RHS->isAllOnesValue();
129 case ICmpInst::ICMP_SGT: // True if LHS s> -1
130 TrueIfSigned = false;
131 return RHS->isAllOnesValue();
132 case ICmpInst::ICMP_UGT:
133 // True if LHS u> RHS and RHS == high-bit-mask - 1
135 return RHS->isMaxValue(true);
136 case ICmpInst::ICMP_UGE:
137 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
139 return RHS->getValue().isSignBit();
145 /// Returns true if the exploded icmp can be expressed as a signed comparison
146 /// to zero and updates the predicate accordingly.
147 /// The signedness of the comparison is preserved.
148 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
149 if (!ICmpInst::isSigned(pred))
153 return ICmpInst::isRelational(pred);
156 if (pred == ICmpInst::ICMP_SLT) {
157 pred = ICmpInst::ICMP_SLE;
160 } else if (RHS->isAllOnesValue()) {
161 if (pred == ICmpInst::ICMP_SGT) {
162 pred = ICmpInst::ICMP_SGE;
170 // isHighOnes - Return true if the constant is of the form 1+0+.
171 // This is the same as lowones(~X).
172 static bool isHighOnes(const ConstantInt *CI) {
173 return (~CI->getValue() + 1).isPowerOf2();
176 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
177 /// set of known zero and one bits, compute the maximum and minimum values that
178 /// could have the specified known zero and known one bits, returning them in
180 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
181 const APInt& KnownOne,
182 APInt& Min, APInt& Max) {
183 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
184 KnownZero.getBitWidth() == Min.getBitWidth() &&
185 KnownZero.getBitWidth() == Max.getBitWidth() &&
186 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
187 APInt UnknownBits = ~(KnownZero|KnownOne);
189 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
190 // bit if it is unknown.
192 Max = KnownOne|UnknownBits;
194 if (UnknownBits.isNegative()) { // Sign bit is unknown
195 Min.setBit(Min.getBitWidth()-1);
196 Max.clearBit(Max.getBitWidth()-1);
200 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
201 // a set of known zero and one bits, compute the maximum and minimum values that
202 // could have the specified known zero and known one bits, returning them in
204 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
205 const APInt &KnownOne,
206 APInt &Min, APInt &Max) {
207 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
208 KnownZero.getBitWidth() == Min.getBitWidth() &&
209 KnownZero.getBitWidth() == Max.getBitWidth() &&
210 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
211 APInt UnknownBits = ~(KnownZero|KnownOne);
213 // The minimum value is when the unknown bits are all zeros.
215 // The maximum value is when the unknown bits are all ones.
216 Max = KnownOne|UnknownBits;
221 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
222 /// cmp pred (load (gep GV, ...)), cmpcst
223 /// where GV is a global variable with a constant initializer. Try to simplify
224 /// this into some simple computation that does not need the load. For example
225 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
227 /// If AndCst is non-null, then the loaded value is masked with that constant
228 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
229 Instruction *InstCombiner::
230 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
231 CmpInst &ICI, ConstantInt *AndCst) {
232 Constant *Init = GV->getInitializer();
233 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
236 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
237 if (ArrayElementCount > 1024) return nullptr; // Don't blow up on huge arrays.
239 // There are many forms of this optimization we can handle, for now, just do
240 // the simple index into a single-dimensional array.
242 // Require: GEP GV, 0, i {{, constant indices}}
243 if (GEP->getNumOperands() < 3 ||
244 !isa<ConstantInt>(GEP->getOperand(1)) ||
245 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
246 isa<Constant>(GEP->getOperand(2)))
249 // Check that indices after the variable are constants and in-range for the
250 // type they index. Collect the indices. This is typically for arrays of
252 SmallVector<unsigned, 4> LaterIndices;
254 Type *EltTy = Init->getType()->getArrayElementType();
255 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
256 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
257 if (!Idx) return nullptr; // Variable index.
259 uint64_t IdxVal = Idx->getZExtValue();
260 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
262 if (StructType *STy = dyn_cast<StructType>(EltTy))
263 EltTy = STy->getElementType(IdxVal);
264 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
265 if (IdxVal >= ATy->getNumElements()) return nullptr;
266 EltTy = ATy->getElementType();
268 return nullptr; // Unknown type.
271 LaterIndices.push_back(IdxVal);
274 enum { Overdefined = -3, Undefined = -2 };
276 // Variables for our state machines.
278 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
279 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
280 // and 87 is the second (and last) index. FirstTrueElement is -2 when
281 // undefined, otherwise set to the first true element. SecondTrueElement is
282 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
283 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
285 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
286 // form "i != 47 & i != 87". Same state transitions as for true elements.
287 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
289 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
290 /// define a state machine that triggers for ranges of values that the index
291 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
292 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
293 /// index in the range (inclusive). We use -2 for undefined here because we
294 /// use relative comparisons and don't want 0-1 to match -1.
295 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
297 // MagicBitvector - This is a magic bitvector where we set a bit if the
298 // comparison is true for element 'i'. If there are 64 elements or less in
299 // the array, this will fully represent all the comparison results.
300 uint64_t MagicBitvector = 0;
302 // Scan the array and see if one of our patterns matches.
303 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
304 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
305 Constant *Elt = Init->getAggregateElement(i);
306 if (!Elt) return nullptr;
308 // If this is indexing an array of structures, get the structure element.
309 if (!LaterIndices.empty())
310 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
312 // If the element is masked, handle it.
313 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
315 // Find out if the comparison would be true or false for the i'th element.
316 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
317 CompareRHS, DL, TLI);
318 // If the result is undef for this element, ignore it.
319 if (isa<UndefValue>(C)) {
320 // Extend range state machines to cover this element in case there is an
321 // undef in the middle of the range.
322 if (TrueRangeEnd == (int)i-1)
324 if (FalseRangeEnd == (int)i-1)
329 // If we can't compute the result for any of the elements, we have to give
330 // up evaluating the entire conditional.
331 if (!isa<ConstantInt>(C)) return nullptr;
333 // Otherwise, we know if the comparison is true or false for this element,
334 // update our state machines.
335 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
337 // State machine for single/double/range index comparison.
339 // Update the TrueElement state machine.
340 if (FirstTrueElement == Undefined)
341 FirstTrueElement = TrueRangeEnd = i; // First true element.
343 // Update double-compare state machine.
344 if (SecondTrueElement == Undefined)
345 SecondTrueElement = i;
347 SecondTrueElement = Overdefined;
349 // Update range state machine.
350 if (TrueRangeEnd == (int)i-1)
353 TrueRangeEnd = Overdefined;
356 // Update the FalseElement state machine.
357 if (FirstFalseElement == Undefined)
358 FirstFalseElement = FalseRangeEnd = i; // First false element.
360 // Update double-compare state machine.
361 if (SecondFalseElement == Undefined)
362 SecondFalseElement = i;
364 SecondFalseElement = Overdefined;
366 // Update range state machine.
367 if (FalseRangeEnd == (int)i-1)
370 FalseRangeEnd = Overdefined;
375 // If this element is in range, update our magic bitvector.
376 if (i < 64 && IsTrueForElt)
377 MagicBitvector |= 1ULL << i;
379 // If all of our states become overdefined, bail out early. Since the
380 // predicate is expensive, only check it every 8 elements. This is only
381 // really useful for really huge arrays.
382 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
383 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
384 FalseRangeEnd == Overdefined)
388 // Now that we've scanned the entire array, emit our new comparison(s). We
389 // order the state machines in complexity of the generated code.
390 Value *Idx = GEP->getOperand(2);
392 // If the index is larger than the pointer size of the target, truncate the
393 // index down like the GEP would do implicitly. We don't have to do this for
394 // an inbounds GEP because the index can't be out of range.
395 if (!GEP->isInBounds()) {
396 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
397 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
398 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
399 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
402 // If the comparison is only true for one or two elements, emit direct
404 if (SecondTrueElement != Overdefined) {
405 // None true -> false.
406 if (FirstTrueElement == Undefined)
407 return ReplaceInstUsesWith(ICI, Builder->getFalse());
409 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
411 // True for one element -> 'i == 47'.
412 if (SecondTrueElement == Undefined)
413 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
415 // True for two elements -> 'i == 47 | i == 72'.
416 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
417 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
418 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
419 return BinaryOperator::CreateOr(C1, C2);
422 // If the comparison is only false for one or two elements, emit direct
424 if (SecondFalseElement != Overdefined) {
425 // None false -> true.
426 if (FirstFalseElement == Undefined)
427 return ReplaceInstUsesWith(ICI, Builder->getTrue());
429 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
431 // False for one element -> 'i != 47'.
432 if (SecondFalseElement == Undefined)
433 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
435 // False for two elements -> 'i != 47 & i != 72'.
436 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
437 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
438 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
439 return BinaryOperator::CreateAnd(C1, C2);
442 // If the comparison can be replaced with a range comparison for the elements
443 // where it is true, emit the range check.
444 if (TrueRangeEnd != Overdefined) {
445 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
447 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
448 if (FirstTrueElement) {
449 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
450 Idx = Builder->CreateAdd(Idx, Offs);
453 Value *End = ConstantInt::get(Idx->getType(),
454 TrueRangeEnd-FirstTrueElement+1);
455 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
458 // False range check.
459 if (FalseRangeEnd != Overdefined) {
460 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
461 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
462 if (FirstFalseElement) {
463 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
464 Idx = Builder->CreateAdd(Idx, Offs);
467 Value *End = ConstantInt::get(Idx->getType(),
468 FalseRangeEnd-FirstFalseElement);
469 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
473 // If a magic bitvector captures the entire comparison state
474 // of this load, replace it with computation that does:
475 // ((magic_cst >> i) & 1) != 0
479 // Look for an appropriate type:
480 // - The type of Idx if the magic fits
481 // - The smallest fitting legal type if we have a DataLayout
483 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
486 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
489 Value *V = Builder->CreateIntCast(Idx, Ty, false);
490 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
491 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
492 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
500 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
501 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
502 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
503 /// be complex, and scales are involved. The above expression would also be
504 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
505 /// This later form is less amenable to optimization though, and we are allowed
506 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
508 /// If we can't emit an optimized form for this expression, this returns null.
510 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC,
511 const DataLayout &DL) {
512 gep_type_iterator GTI = gep_type_begin(GEP);
514 // Check to see if this gep only has a single variable index. If so, and if
515 // any constant indices are a multiple of its scale, then we can compute this
516 // in terms of the scale of the variable index. For example, if the GEP
517 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
518 // because the expression will cross zero at the same point.
519 unsigned i, e = GEP->getNumOperands();
521 for (i = 1; i != e; ++i, ++GTI) {
522 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
523 // Compute the aggregate offset of constant indices.
524 if (CI->isZero()) continue;
526 // Handle a struct index, which adds its field offset to the pointer.
527 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
528 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
530 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
531 Offset += Size*CI->getSExtValue();
534 // Found our variable index.
539 // If there are no variable indices, we must have a constant offset, just
540 // evaluate it the general way.
541 if (i == e) return nullptr;
543 Value *VariableIdx = GEP->getOperand(i);
544 // Determine the scale factor of the variable element. For example, this is
545 // 4 if the variable index is into an array of i32.
546 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
548 // Verify that there are no other variable indices. If so, emit the hard way.
549 for (++i, ++GTI; i != e; ++i, ++GTI) {
550 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
551 if (!CI) return nullptr;
553 // Compute the aggregate offset of constant indices.
554 if (CI->isZero()) continue;
556 // Handle a struct index, which adds its field offset to the pointer.
557 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
558 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
560 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
561 Offset += Size*CI->getSExtValue();
567 // Okay, we know we have a single variable index, which must be a
568 // pointer/array/vector index. If there is no offset, life is simple, return
570 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
571 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
573 // Cast to intptrty in case a truncation occurs. If an extension is needed,
574 // we don't need to bother extending: the extension won't affect where the
575 // computation crosses zero.
576 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
577 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
582 // Otherwise, there is an index. The computation we will do will be modulo
583 // the pointer size, so get it.
584 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
586 Offset &= PtrSizeMask;
587 VariableScale &= PtrSizeMask;
589 // To do this transformation, any constant index must be a multiple of the
590 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
591 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
592 // multiple of the variable scale.
593 int64_t NewOffs = Offset / (int64_t)VariableScale;
594 if (Offset != NewOffs*(int64_t)VariableScale)
597 // Okay, we can do this evaluation. Start by converting the index to intptr.
598 if (VariableIdx->getType() != IntPtrTy)
599 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
601 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
602 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
605 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
606 /// else. At this point we know that the GEP is on the LHS of the comparison.
607 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
608 ICmpInst::Predicate Cond,
610 // Don't transform signed compares of GEPs into index compares. Even if the
611 // GEP is inbounds, the final add of the base pointer can have signed overflow
612 // and would change the result of the icmp.
613 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
614 // the maximum signed value for the pointer type.
615 if (ICmpInst::isSigned(Cond))
618 // Look through bitcasts and addrspacecasts. We do not however want to remove
620 if (!isa<GetElementPtrInst>(RHS))
621 RHS = RHS->stripPointerCasts();
623 Value *PtrBase = GEPLHS->getOperand(0);
624 if (PtrBase == RHS && GEPLHS->isInBounds()) {
625 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
626 // This transformation (ignoring the base and scales) is valid because we
627 // know pointers can't overflow since the gep is inbounds. See if we can
628 // output an optimized form.
629 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this, DL);
631 // If not, synthesize the offset the hard way.
633 Offset = EmitGEPOffset(GEPLHS);
634 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
635 Constant::getNullValue(Offset->getType()));
636 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
637 // If the base pointers are different, but the indices are the same, just
638 // compare the base pointer.
639 if (PtrBase != GEPRHS->getOperand(0)) {
640 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
641 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
642 GEPRHS->getOperand(0)->getType();
644 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
645 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
646 IndicesTheSame = false;
650 // If all indices are the same, just compare the base pointers.
652 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
654 // If we're comparing GEPs with two base pointers that only differ in type
655 // and both GEPs have only constant indices or just one use, then fold
656 // the compare with the adjusted indices.
657 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
658 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
659 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
660 PtrBase->stripPointerCasts() ==
661 GEPRHS->getOperand(0)->stripPointerCasts()) {
662 Value *LOffset = EmitGEPOffset(GEPLHS);
663 Value *ROffset = EmitGEPOffset(GEPRHS);
665 // If we looked through an addrspacecast between different sized address
666 // spaces, the LHS and RHS pointers are different sized
667 // integers. Truncate to the smaller one.
668 Type *LHSIndexTy = LOffset->getType();
669 Type *RHSIndexTy = ROffset->getType();
670 if (LHSIndexTy != RHSIndexTy) {
671 if (LHSIndexTy->getPrimitiveSizeInBits() <
672 RHSIndexTy->getPrimitiveSizeInBits()) {
673 ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy);
675 LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy);
678 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
680 return ReplaceInstUsesWith(I, Cmp);
683 // Otherwise, the base pointers are different and the indices are
684 // different, bail out.
688 // If one of the GEPs has all zero indices, recurse.
689 if (GEPLHS->hasAllZeroIndices())
690 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
691 ICmpInst::getSwappedPredicate(Cond), I);
693 // If the other GEP has all zero indices, recurse.
694 if (GEPRHS->hasAllZeroIndices())
695 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
697 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
698 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
699 // If the GEPs only differ by one index, compare it.
700 unsigned NumDifferences = 0; // Keep track of # differences.
701 unsigned DiffOperand = 0; // The operand that differs.
702 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
703 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
704 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
705 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
706 // Irreconcilable differences.
710 if (NumDifferences++) break;
715 if (NumDifferences == 0) // SAME GEP?
716 return ReplaceInstUsesWith(I, // No comparison is needed here.
717 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
719 else if (NumDifferences == 1 && GEPsInBounds) {
720 Value *LHSV = GEPLHS->getOperand(DiffOperand);
721 Value *RHSV = GEPRHS->getOperand(DiffOperand);
722 // Make sure we do a signed comparison here.
723 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
727 // Only lower this if the icmp is the only user of the GEP or if we expect
728 // the result to fold to a constant!
729 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
730 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
731 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
732 Value *L = EmitGEPOffset(GEPLHS);
733 Value *R = EmitGEPOffset(GEPRHS);
734 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
740 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
741 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
742 Value *X, ConstantInt *CI,
743 ICmpInst::Predicate Pred) {
744 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
745 // so the values can never be equal. Similarly for all other "or equals"
748 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
749 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
750 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
751 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
753 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
754 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
757 // (X+1) >u X --> X <u (0-1) --> X != 255
758 // (X+2) >u X --> X <u (0-2) --> X <u 254
759 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
760 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
761 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
763 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
764 ConstantInt *SMax = ConstantInt::get(X->getContext(),
765 APInt::getSignedMaxValue(BitWidth));
767 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
768 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
769 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
770 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
771 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
772 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
773 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
774 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
776 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
777 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
778 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
779 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
780 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
781 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
783 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
784 Constant *C = Builder->getInt(CI->getValue()-1);
785 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
788 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
789 /// and CmpRHS are both known to be integer constants.
790 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
791 ConstantInt *DivRHS) {
792 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
793 const APInt &CmpRHSV = CmpRHS->getValue();
795 // FIXME: If the operand types don't match the type of the divide
796 // then don't attempt this transform. The code below doesn't have the
797 // logic to deal with a signed divide and an unsigned compare (and
798 // vice versa). This is because (x /s C1) <s C2 produces different
799 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
800 // (x /u C1) <u C2. Simply casting the operands and result won't
801 // work. :( The if statement below tests that condition and bails
803 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
804 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
806 if (DivRHS->isZero())
807 return nullptr; // The ProdOV computation fails on divide by zero.
808 if (DivIsSigned && DivRHS->isAllOnesValue())
809 return nullptr; // The overflow computation also screws up here
810 if (DivRHS->isOne()) {
811 // This eliminates some funny cases with INT_MIN.
812 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
816 // Compute Prod = CI * DivRHS. We are essentially solving an equation
817 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
818 // C2 (CI). By solving for X we can turn this into a range check
819 // instead of computing a divide.
820 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
822 // Determine if the product overflows by seeing if the product is
823 // not equal to the divide. Make sure we do the same kind of divide
824 // as in the LHS instruction that we're folding.
825 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
826 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
828 // Get the ICmp opcode
829 ICmpInst::Predicate Pred = ICI.getPredicate();
831 /// If the division is known to be exact, then there is no remainder from the
832 /// divide, so the covered range size is unit, otherwise it is the divisor.
833 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
835 // Figure out the interval that is being checked. For example, a comparison
836 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
837 // Compute this interval based on the constants involved and the signedness of
838 // the compare/divide. This computes a half-open interval, keeping track of
839 // whether either value in the interval overflows. After analysis each
840 // overflow variable is set to 0 if it's corresponding bound variable is valid
841 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
842 int LoOverflow = 0, HiOverflow = 0;
843 Constant *LoBound = nullptr, *HiBound = nullptr;
845 if (!DivIsSigned) { // udiv
846 // e.g. X/5 op 3 --> [15, 20)
848 HiOverflow = LoOverflow = ProdOV;
850 // If this is not an exact divide, then many values in the range collapse
851 // to the same result value.
852 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
855 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
856 if (CmpRHSV == 0) { // (X / pos) op 0
857 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
858 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
860 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
861 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
862 HiOverflow = LoOverflow = ProdOV;
864 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
865 } else { // (X / pos) op neg
866 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
867 HiBound = AddOne(Prod);
868 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
870 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
871 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
874 } else if (DivRHS->isNegative()) { // Divisor is < 0.
876 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
877 if (CmpRHSV == 0) { // (X / neg) op 0
878 // e.g. X/-5 op 0 --> [-4, 5)
879 LoBound = AddOne(RangeSize);
880 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
881 if (HiBound == DivRHS) { // -INTMIN = INTMIN
882 HiOverflow = 1; // [INTMIN+1, overflow)
883 HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
885 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
886 // e.g. X/-5 op 3 --> [-19, -14)
887 HiBound = AddOne(Prod);
888 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
890 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
891 } else { // (X / neg) op neg
892 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
893 LoOverflow = HiOverflow = ProdOV;
895 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
898 // Dividing by a negative swaps the condition. LT <-> GT
899 Pred = ICmpInst::getSwappedPredicate(Pred);
902 Value *X = DivI->getOperand(0);
904 default: llvm_unreachable("Unhandled icmp opcode!");
905 case ICmpInst::ICMP_EQ:
906 if (LoOverflow && HiOverflow)
907 return ReplaceInstUsesWith(ICI, Builder->getFalse());
909 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
910 ICmpInst::ICMP_UGE, X, LoBound);
912 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
913 ICmpInst::ICMP_ULT, X, HiBound);
914 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
916 case ICmpInst::ICMP_NE:
917 if (LoOverflow && HiOverflow)
918 return ReplaceInstUsesWith(ICI, Builder->getTrue());
920 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
921 ICmpInst::ICMP_ULT, X, LoBound);
923 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
924 ICmpInst::ICMP_UGE, X, HiBound);
925 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
926 DivIsSigned, false));
927 case ICmpInst::ICMP_ULT:
928 case ICmpInst::ICMP_SLT:
929 if (LoOverflow == +1) // Low bound is greater than input range.
930 return ReplaceInstUsesWith(ICI, Builder->getTrue());
931 if (LoOverflow == -1) // Low bound is less than input range.
932 return ReplaceInstUsesWith(ICI, Builder->getFalse());
933 return new ICmpInst(Pred, X, LoBound);
934 case ICmpInst::ICMP_UGT:
935 case ICmpInst::ICMP_SGT:
936 if (HiOverflow == +1) // High bound greater than input range.
937 return ReplaceInstUsesWith(ICI, Builder->getFalse());
938 if (HiOverflow == -1) // High bound less than input range.
939 return ReplaceInstUsesWith(ICI, Builder->getTrue());
940 if (Pred == ICmpInst::ICMP_UGT)
941 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
942 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
946 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
947 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
948 ConstantInt *ShAmt) {
949 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
951 // Check that the shift amount is in range. If not, don't perform
952 // undefined shifts. When the shift is visited it will be
954 uint32_t TypeBits = CmpRHSV.getBitWidth();
955 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
956 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
959 if (!ICI.isEquality()) {
960 // If we have an unsigned comparison and an ashr, we can't simplify this.
961 // Similarly for signed comparisons with lshr.
962 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
965 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
966 // by a power of 2. Since we already have logic to simplify these,
967 // transform to div and then simplify the resultant comparison.
968 if (Shr->getOpcode() == Instruction::AShr &&
969 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
972 // Revisit the shift (to delete it).
976 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
979 Shr->getOpcode() == Instruction::AShr ?
980 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
981 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
983 ICI.setOperand(0, Tmp);
985 // If the builder folded the binop, just return it.
986 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
990 // Otherwise, fold this div/compare.
991 assert(TheDiv->getOpcode() == Instruction::SDiv ||
992 TheDiv->getOpcode() == Instruction::UDiv);
994 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
995 assert(Res && "This div/cst should have folded!");
1000 // If we are comparing against bits always shifted out, the
1001 // comparison cannot succeed.
1002 APInt Comp = CmpRHSV << ShAmtVal;
1003 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1004 if (Shr->getOpcode() == Instruction::LShr)
1005 Comp = Comp.lshr(ShAmtVal);
1007 Comp = Comp.ashr(ShAmtVal);
1009 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1010 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1011 Constant *Cst = Builder->getInt1(IsICMP_NE);
1012 return ReplaceInstUsesWith(ICI, Cst);
1015 // Otherwise, check to see if the bits shifted out are known to be zero.
1016 // If so, we can compare against the unshifted value:
1017 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1018 if (Shr->hasOneUse() && Shr->isExact())
1019 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1021 if (Shr->hasOneUse()) {
1022 // Otherwise strength reduce the shift into an and.
1023 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1024 Constant *Mask = Builder->getInt(Val);
1026 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1027 Mask, Shr->getName()+".mask");
1028 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1033 /// FoldICmpCstShrCst - Handle "(icmp eq/ne (ashr/lshr const2, A), const1)" ->
1034 /// (icmp eq/ne A, Log2(const2/const1)) ->
1035 /// (icmp eq/ne A, Log2(const2) - Log2(const1)).
1036 Instruction *InstCombiner::FoldICmpCstShrCst(ICmpInst &I, Value *Op, Value *A,
1039 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1041 auto getConstant = [&I, this](bool IsTrue) {
1042 if (I.getPredicate() == I.ICMP_NE)
1044 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1047 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1048 if (I.getPredicate() == I.ICMP_NE)
1049 Pred = CmpInst::getInversePredicate(Pred);
1050 return new ICmpInst(Pred, LHS, RHS);
1053 APInt AP1 = CI1->getValue();
1054 APInt AP2 = CI2->getValue();
1056 // Don't bother doing any work for cases which InstSimplify handles.
1059 bool IsAShr = isa<AShrOperator>(Op);
1061 if (AP2.isAllOnesValue())
1063 if (AP2.isNegative() != AP1.isNegative())
1070 // 'A' must be large enough to shift out the highest set bit.
1071 return getICmp(I.ICMP_UGT, A,
1072 ConstantInt::get(A->getType(), AP2.logBase2()));
1075 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1077 // Get the distance between the highest bit that's set.
1079 // Both the constants are negative, take their positive to calculate log.
1080 if (IsAShr && AP1.isNegative())
1081 // Get the ones' complement of AP2 and AP1 when computing the distance.
1082 Shift = (~AP2).logBase2() - (~AP1).logBase2();
1084 Shift = AP2.logBase2() - AP1.logBase2();
1087 if (IsAShr ? AP1 == AP2.ashr(Shift) : AP1 == AP2.lshr(Shift))
1088 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1090 // Shifting const2 will never be equal to const1.
1091 return getConstant(false);
1094 /// FoldICmpCstShlCst - Handle "(icmp eq/ne (shl const2, A), const1)" ->
1095 /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)).
1096 Instruction *InstCombiner::FoldICmpCstShlCst(ICmpInst &I, Value *Op, Value *A,
1099 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1101 auto getConstant = [&I, this](bool IsTrue) {
1102 if (I.getPredicate() == I.ICMP_NE)
1104 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1107 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1108 if (I.getPredicate() == I.ICMP_NE)
1109 Pred = CmpInst::getInversePredicate(Pred);
1110 return new ICmpInst(Pred, LHS, RHS);
1113 APInt AP1 = CI1->getValue();
1114 APInt AP2 = CI2->getValue();
1116 // Don't bother doing any work for cases which InstSimplify handles.
1120 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1122 if (!AP1 && AP2TrailingZeros != 0)
1123 return getICmp(I.ICMP_UGE, A,
1124 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1127 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1129 // Get the distance between the lowest bits that are set.
1130 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1132 if (Shift > 0 && AP2.shl(Shift) == AP1)
1133 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1135 // Shifting const2 will never be equal to const1.
1136 return getConstant(false);
1139 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1141 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1144 const APInt &RHSV = RHS->getValue();
1146 switch (LHSI->getOpcode()) {
1147 case Instruction::Trunc:
1148 if (RHS->isOne() && RHSV.getBitWidth() > 1) {
1149 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1151 if (ICI.getPredicate() == ICmpInst::ICMP_SLT &&
1152 match(LHSI->getOperand(0), m_Signum(m_Value(V))))
1153 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1154 ConstantInt::get(V->getType(), 1));
1156 if (ICI.isEquality() && LHSI->hasOneUse()) {
1157 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1158 // of the high bits truncated out of x are known.
1159 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1160 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1161 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1162 computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne, 0, &ICI);
1164 // If all the high bits are known, we can do this xform.
1165 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1166 // Pull in the high bits from known-ones set.
1167 APInt NewRHS = RHS->getValue().zext(SrcBits);
1168 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1169 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1170 Builder->getInt(NewRHS));
1175 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1176 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1177 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1179 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1180 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1181 Value *CompareVal = LHSI->getOperand(0);
1183 // If the sign bit of the XorCst is not set, there is no change to
1184 // the operation, just stop using the Xor.
1185 if (!XorCst->isNegative()) {
1186 ICI.setOperand(0, CompareVal);
1191 // Was the old condition true if the operand is positive?
1192 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1194 // If so, the new one isn't.
1195 isTrueIfPositive ^= true;
1197 if (isTrueIfPositive)
1198 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1201 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1205 if (LHSI->hasOneUse()) {
1206 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1207 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1208 const APInt &SignBit = XorCst->getValue();
1209 ICmpInst::Predicate Pred = ICI.isSigned()
1210 ? ICI.getUnsignedPredicate()
1211 : ICI.getSignedPredicate();
1212 return new ICmpInst(Pred, LHSI->getOperand(0),
1213 Builder->getInt(RHSV ^ SignBit));
1216 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1217 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1218 const APInt &NotSignBit = XorCst->getValue();
1219 ICmpInst::Predicate Pred = ICI.isSigned()
1220 ? ICI.getUnsignedPredicate()
1221 : ICI.getSignedPredicate();
1222 Pred = ICI.getSwappedPredicate(Pred);
1223 return new ICmpInst(Pred, LHSI->getOperand(0),
1224 Builder->getInt(RHSV ^ NotSignBit));
1228 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1229 // iff -C is a power of 2
1230 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1231 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1232 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1234 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1235 // iff -C is a power of 2
1236 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1237 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1238 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1241 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1242 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1243 LHSI->getOperand(0)->hasOneUse()) {
1244 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1246 // If the LHS is an AND of a truncating cast, we can widen the
1247 // and/compare to be the input width without changing the value
1248 // produced, eliminating a cast.
1249 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1250 // We can do this transformation if either the AND constant does not
1251 // have its sign bit set or if it is an equality comparison.
1252 // Extending a relational comparison when we're checking the sign
1253 // bit would not work.
1254 if (ICI.isEquality() ||
1255 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1257 Builder->CreateAnd(Cast->getOperand(0),
1258 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1259 NewAnd->takeName(LHSI);
1260 return new ICmpInst(ICI.getPredicate(), NewAnd,
1261 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1265 // If the LHS is an AND of a zext, and we have an equality compare, we can
1266 // shrink the and/compare to the smaller type, eliminating the cast.
1267 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1268 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1269 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1270 // should fold the icmp to true/false in that case.
1271 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1273 Builder->CreateAnd(Cast->getOperand(0),
1274 ConstantExpr::getTrunc(AndCst, Ty));
1275 NewAnd->takeName(LHSI);
1276 return new ICmpInst(ICI.getPredicate(), NewAnd,
1277 ConstantExpr::getTrunc(RHS, Ty));
1281 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1282 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1283 // happens a LOT in code produced by the C front-end, for bitfield
1285 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1286 if (Shift && !Shift->isShift())
1290 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
1292 // This seemingly simple opportunity to fold away a shift turns out to
1293 // be rather complicated. See PR17827
1294 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1296 bool CanFold = false;
1297 unsigned ShiftOpcode = Shift->getOpcode();
1298 if (ShiftOpcode == Instruction::AShr) {
1299 // There may be some constraints that make this possible,
1300 // but nothing simple has been discovered yet.
1302 } else if (ShiftOpcode == Instruction::Shl) {
1303 // For a left shift, we can fold if the comparison is not signed.
1304 // We can also fold a signed comparison if the mask value and
1305 // comparison value are not negative. These constraints may not be
1306 // obvious, but we can prove that they are correct using an SMT
1308 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1310 } else if (ShiftOpcode == Instruction::LShr) {
1311 // For a logical right shift, we can fold if the comparison is not
1312 // signed. We can also fold a signed comparison if the shifted mask
1313 // value and the shifted comparison value are not negative.
1314 // These constraints may not be obvious, but we can prove that they
1315 // are correct using an SMT solver.
1316 if (!ICI.isSigned())
1319 ConstantInt *ShiftedAndCst =
1320 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1321 ConstantInt *ShiftedRHSCst =
1322 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1324 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1331 if (ShiftOpcode == Instruction::Shl)
1332 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1334 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1336 // Check to see if we are shifting out any of the bits being
1338 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1339 // If we shifted bits out, the fold is not going to work out.
1340 // As a special case, check to see if this means that the
1341 // result is always true or false now.
1342 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1343 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1344 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1345 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1347 ICI.setOperand(1, NewCst);
1348 Constant *NewAndCst;
1349 if (ShiftOpcode == Instruction::Shl)
1350 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1352 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1353 LHSI->setOperand(1, NewAndCst);
1354 LHSI->setOperand(0, Shift->getOperand(0));
1355 Worklist.Add(Shift); // Shift is dead.
1361 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1362 // preferable because it allows the C<<Y expression to be hoisted out
1363 // of a loop if Y is invariant and X is not.
1364 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1365 ICI.isEquality() && !Shift->isArithmeticShift() &&
1366 !isa<Constant>(Shift->getOperand(0))) {
1369 if (Shift->getOpcode() == Instruction::LShr) {
1370 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1372 // Insert a logical shift.
1373 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1376 // Compute X & (C << Y).
1378 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1380 ICI.setOperand(0, NewAnd);
1384 // (icmp pred (and (or (lshr X, Y), X), 1), 0) -->
1385 // (icmp pred (and X, (or (shl 1, Y), 1), 0))
1387 // iff pred isn't signed
1389 Value *X, *Y, *LShr;
1390 if (!ICI.isSigned() && RHSV == 0) {
1391 if (match(LHSI->getOperand(1), m_One())) {
1392 Constant *One = cast<Constant>(LHSI->getOperand(1));
1393 Value *Or = LHSI->getOperand(0);
1394 if (match(Or, m_Or(m_Value(LShr), m_Value(X))) &&
1395 match(LShr, m_LShr(m_Specific(X), m_Value(Y)))) {
1396 unsigned UsesRemoved = 0;
1397 if (LHSI->hasOneUse())
1399 if (Or->hasOneUse())
1401 if (LShr->hasOneUse())
1403 Value *NewOr = nullptr;
1404 // Compute X & ((1 << Y) | 1)
1405 if (auto *C = dyn_cast<Constant>(Y)) {
1406 if (UsesRemoved >= 1)
1408 ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1410 if (UsesRemoved >= 3)
1411 NewOr = Builder->CreateOr(Builder->CreateShl(One, Y,
1414 One, Or->getName());
1417 Value *NewAnd = Builder->CreateAnd(X, NewOr, LHSI->getName());
1418 ICI.setOperand(0, NewAnd);
1426 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1427 // bit set in (X & AndCst) will produce a result greater than RHSV.
1428 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1429 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1430 if ((NTZ < AndCst->getBitWidth()) &&
1431 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1432 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1433 Constant::getNullValue(RHS->getType()));
1437 // Try to optimize things like "A[i]&42 == 0" to index computations.
1438 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1439 if (GetElementPtrInst *GEP =
1440 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1441 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1442 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1443 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1444 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1445 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1450 // X & -C == -C -> X > u ~C
1451 // X & -C != -C -> X <= u ~C
1452 // iff C is a power of 2
1453 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1454 return new ICmpInst(
1455 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1456 : ICmpInst::ICMP_ULE,
1457 LHSI->getOperand(0), SubOne(RHS));
1459 // (icmp eq (and %A, C), 0) -> (icmp sgt (trunc %A), -1)
1460 // iff C is a power of 2
1461 if (ICI.isEquality() && LHSI->hasOneUse() && match(RHS, m_Zero())) {
1462 if (auto *CI = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1463 const APInt &AI = CI->getValue();
1464 int32_t ExactLogBase2 = AI.exactLogBase2();
1465 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1466 Type *NTy = IntegerType::get(ICI.getContext(), ExactLogBase2 + 1);
1467 Value *Trunc = Builder->CreateTrunc(LHSI->getOperand(0), NTy);
1468 return new ICmpInst(ICI.getPredicate() == ICmpInst::ICMP_EQ
1469 ? ICmpInst::ICMP_SGE
1470 : ICmpInst::ICMP_SLT,
1471 Trunc, Constant::getNullValue(NTy));
1477 case Instruction::Or: {
1479 // icmp slt signum(V) 1 --> icmp slt V, 1
1481 if (ICI.getPredicate() == ICmpInst::ICMP_SLT &&
1482 match(LHSI, m_Signum(m_Value(V))))
1483 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1484 ConstantInt::get(V->getType(), 1));
1487 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1490 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1491 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1492 // -> and (icmp eq P, null), (icmp eq Q, null).
1493 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1494 Constant::getNullValue(P->getType()));
1495 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1496 Constant::getNullValue(Q->getType()));
1498 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1499 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1501 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1507 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1508 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1511 // If this is a signed comparison to 0 and the mul is sign preserving,
1512 // use the mul LHS operand instead.
1513 ICmpInst::Predicate pred = ICI.getPredicate();
1514 if (isSignTest(pred, RHS) && !Val->isZero() &&
1515 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1516 return new ICmpInst(Val->isNegative() ?
1517 ICmpInst::getSwappedPredicate(pred) : pred,
1518 LHSI->getOperand(0),
1519 Constant::getNullValue(RHS->getType()));
1524 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1525 uint32_t TypeBits = RHSV.getBitWidth();
1526 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1529 // (1 << X) pred P2 -> X pred Log2(P2)
1530 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1531 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1532 ICmpInst::Predicate Pred = ICI.getPredicate();
1533 if (ICI.isUnsigned()) {
1534 if (!RHSVIsPowerOf2) {
1535 // (1 << X) < 30 -> X <= 4
1536 // (1 << X) <= 30 -> X <= 4
1537 // (1 << X) >= 30 -> X > 4
1538 // (1 << X) > 30 -> X > 4
1539 if (Pred == ICmpInst::ICMP_ULT)
1540 Pred = ICmpInst::ICMP_ULE;
1541 else if (Pred == ICmpInst::ICMP_UGE)
1542 Pred = ICmpInst::ICMP_UGT;
1544 unsigned RHSLog2 = RHSV.logBase2();
1546 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1547 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1548 if (RHSLog2 == TypeBits-1) {
1549 if (Pred == ICmpInst::ICMP_UGE)
1550 Pred = ICmpInst::ICMP_EQ;
1551 else if (Pred == ICmpInst::ICMP_ULT)
1552 Pred = ICmpInst::ICMP_NE;
1555 return new ICmpInst(Pred, X,
1556 ConstantInt::get(RHS->getType(), RHSLog2));
1557 } else if (ICI.isSigned()) {
1558 if (RHSV.isAllOnesValue()) {
1559 // (1 << X) <= -1 -> X == 31
1560 if (Pred == ICmpInst::ICMP_SLE)
1561 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1562 ConstantInt::get(RHS->getType(), TypeBits-1));
1564 // (1 << X) > -1 -> X != 31
1565 if (Pred == ICmpInst::ICMP_SGT)
1566 return new ICmpInst(ICmpInst::ICMP_NE, X,
1567 ConstantInt::get(RHS->getType(), TypeBits-1));
1569 // (1 << X) < 0 -> X == 31
1570 // (1 << X) <= 0 -> X == 31
1571 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1572 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1573 ConstantInt::get(RHS->getType(), TypeBits-1));
1575 // (1 << X) >= 0 -> X != 31
1576 // (1 << X) > 0 -> X != 31
1577 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1578 return new ICmpInst(ICmpInst::ICMP_NE, X,
1579 ConstantInt::get(RHS->getType(), TypeBits-1));
1581 } else if (ICI.isEquality()) {
1583 return new ICmpInst(
1584 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1590 // Check that the shift amount is in range. If not, don't perform
1591 // undefined shifts. When the shift is visited it will be
1593 if (ShAmt->uge(TypeBits))
1596 if (ICI.isEquality()) {
1597 // If we are comparing against bits always shifted out, the
1598 // comparison cannot succeed.
1600 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1602 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1603 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1604 Constant *Cst = Builder->getInt1(IsICMP_NE);
1605 return ReplaceInstUsesWith(ICI, Cst);
1608 // If the shift is NUW, then it is just shifting out zeros, no need for an
1610 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1611 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1612 ConstantExpr::getLShr(RHS, ShAmt));
1614 // If the shift is NSW and we compare to 0, then it is just shifting out
1615 // sign bits, no need for an AND either.
1616 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1617 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1618 ConstantExpr::getLShr(RHS, ShAmt));
1620 if (LHSI->hasOneUse()) {
1621 // Otherwise strength reduce the shift into an and.
1622 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1623 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1624 TypeBits - ShAmtVal));
1627 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1628 return new ICmpInst(ICI.getPredicate(), And,
1629 ConstantExpr::getLShr(RHS, ShAmt));
1633 // If this is a signed comparison to 0 and the shift is sign preserving,
1634 // use the shift LHS operand instead.
1635 ICmpInst::Predicate pred = ICI.getPredicate();
1636 if (isSignTest(pred, RHS) &&
1637 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1638 return new ICmpInst(pred,
1639 LHSI->getOperand(0),
1640 Constant::getNullValue(RHS->getType()));
1642 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1643 bool TrueIfSigned = false;
1644 if (LHSI->hasOneUse() &&
1645 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1646 // (X << 31) <s 0 --> (X&1) != 0
1647 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1648 APInt::getOneBitSet(TypeBits,
1649 TypeBits-ShAmt->getZExtValue()-1));
1651 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1652 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1653 And, Constant::getNullValue(And->getType()));
1656 // Transform (icmp pred iM (shl iM %v, N), CI)
1657 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1658 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1659 // This enables to get rid of the shift in favor of a trunc which can be
1660 // free on the target. It has the additional benefit of comparing to a
1661 // smaller constant, which will be target friendly.
1662 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1663 if (LHSI->hasOneUse() &&
1664 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1665 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1666 Constant *NCI = ConstantExpr::getTrunc(
1667 ConstantExpr::getAShr(RHS,
1668 ConstantInt::get(RHS->getType(), Amt)),
1670 return new ICmpInst(ICI.getPredicate(),
1671 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1678 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1679 case Instruction::AShr: {
1680 // Handle equality comparisons of shift-by-constant.
1681 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1682 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1683 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1687 // Handle exact shr's.
1688 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1689 if (RHSV.isMinValue())
1690 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1695 case Instruction::SDiv:
1696 case Instruction::UDiv:
1697 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1698 // Fold this div into the comparison, producing a range check.
1699 // Determine, based on the divide type, what the range is being
1700 // checked. If there is an overflow on the low or high side, remember
1701 // it, otherwise compute the range [low, hi) bounding the new value.
1702 // See: InsertRangeTest above for the kinds of replacements possible.
1703 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1704 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1709 case Instruction::Sub: {
1710 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1712 const APInt &LHSV = LHSC->getValue();
1714 // C1-X <u C2 -> (X|(C2-1)) == C1
1715 // iff C1 & (C2-1) == C2-1
1716 // C2 is a power of 2
1717 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1718 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1719 return new ICmpInst(ICmpInst::ICMP_EQ,
1720 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1723 // C1-X >u C2 -> (X|C2) != C1
1724 // iff C1 & C2 == C2
1725 // C2+1 is a power of 2
1726 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1727 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1728 return new ICmpInst(ICmpInst::ICMP_NE,
1729 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1733 case Instruction::Add:
1734 // Fold: icmp pred (add X, C1), C2
1735 if (!ICI.isEquality()) {
1736 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1738 const APInt &LHSV = LHSC->getValue();
1740 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1743 if (ICI.isSigned()) {
1744 if (CR.getLower().isSignBit()) {
1745 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1746 Builder->getInt(CR.getUpper()));
1747 } else if (CR.getUpper().isSignBit()) {
1748 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1749 Builder->getInt(CR.getLower()));
1752 if (CR.getLower().isMinValue()) {
1753 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1754 Builder->getInt(CR.getUpper()));
1755 } else if (CR.getUpper().isMinValue()) {
1756 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1757 Builder->getInt(CR.getLower()));
1761 // X-C1 <u C2 -> (X & -C2) == C1
1762 // iff C1 & (C2-1) == 0
1763 // C2 is a power of 2
1764 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1765 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1766 return new ICmpInst(ICmpInst::ICMP_EQ,
1767 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1768 ConstantExpr::getNeg(LHSC));
1770 // X-C1 >u C2 -> (X & ~C2) != C1
1772 // C2+1 is a power of 2
1773 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1774 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1775 return new ICmpInst(ICmpInst::ICMP_NE,
1776 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1777 ConstantExpr::getNeg(LHSC));
1782 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1783 if (ICI.isEquality()) {
1784 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1786 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1787 // the second operand is a constant, simplify a bit.
1788 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1789 switch (BO->getOpcode()) {
1790 case Instruction::SRem:
1791 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1792 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1793 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1794 if (V.sgt(1) && V.isPowerOf2()) {
1796 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1798 return new ICmpInst(ICI.getPredicate(), NewRem,
1799 Constant::getNullValue(BO->getType()));
1803 case Instruction::Add:
1804 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1805 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1806 if (BO->hasOneUse())
1807 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1808 ConstantExpr::getSub(RHS, BOp1C));
1809 } else if (RHSV == 0) {
1810 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1811 // efficiently invertible, or if the add has just this one use.
1812 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1814 if (Value *NegVal = dyn_castNegVal(BOp1))
1815 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1816 if (Value *NegVal = dyn_castNegVal(BOp0))
1817 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1818 if (BO->hasOneUse()) {
1819 Value *Neg = Builder->CreateNeg(BOp1);
1821 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1825 case Instruction::Xor:
1826 // For the xor case, we can xor two constants together, eliminating
1827 // the explicit xor.
1828 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1829 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1830 ConstantExpr::getXor(RHS, BOC));
1831 } else if (RHSV == 0) {
1832 // Replace ((xor A, B) != 0) with (A != B)
1833 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1837 case Instruction::Sub:
1838 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1839 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1840 if (BO->hasOneUse())
1841 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1842 ConstantExpr::getSub(BOp0C, RHS));
1843 } else if (RHSV == 0) {
1844 // Replace ((sub A, B) != 0) with (A != B)
1845 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1849 case Instruction::Or:
1850 // If bits are being or'd in that are not present in the constant we
1851 // are comparing against, then the comparison could never succeed!
1852 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1853 Constant *NotCI = ConstantExpr::getNot(RHS);
1854 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1855 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1859 case Instruction::And:
1860 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1861 // If bits are being compared against that are and'd out, then the
1862 // comparison can never succeed!
1863 if ((RHSV & ~BOC->getValue()) != 0)
1864 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1866 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1867 if (RHS == BOC && RHSV.isPowerOf2())
1868 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1869 ICmpInst::ICMP_NE, LHSI,
1870 Constant::getNullValue(RHS->getType()));
1872 // Don't perform the following transforms if the AND has multiple uses
1873 if (!BO->hasOneUse())
1876 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1877 if (BOC->getValue().isSignBit()) {
1878 Value *X = BO->getOperand(0);
1879 Constant *Zero = Constant::getNullValue(X->getType());
1880 ICmpInst::Predicate pred = isICMP_NE ?
1881 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1882 return new ICmpInst(pred, X, Zero);
1885 // ((X & ~7) == 0) --> X < 8
1886 if (RHSV == 0 && isHighOnes(BOC)) {
1887 Value *X = BO->getOperand(0);
1888 Constant *NegX = ConstantExpr::getNeg(BOC);
1889 ICmpInst::Predicate pred = isICMP_NE ?
1890 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1891 return new ICmpInst(pred, X, NegX);
1895 case Instruction::Mul:
1896 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1897 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1898 // The trivial case (mul X, 0) is handled by InstSimplify
1899 // General case : (mul X, C) != 0 iff X != 0
1900 // (mul X, C) == 0 iff X == 0
1902 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1903 Constant::getNullValue(RHS->getType()));
1909 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1910 // Handle icmp {eq|ne} <intrinsic>, intcst.
1911 switch (II->getIntrinsicID()) {
1912 case Intrinsic::bswap:
1914 ICI.setOperand(0, II->getArgOperand(0));
1915 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1917 case Intrinsic::ctlz:
1918 case Intrinsic::cttz:
1919 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1920 if (RHSV == RHS->getType()->getBitWidth()) {
1922 ICI.setOperand(0, II->getArgOperand(0));
1923 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1927 case Intrinsic::ctpop:
1928 // popcount(A) == 0 -> A == 0 and likewise for !=
1929 if (RHS->isZero()) {
1931 ICI.setOperand(0, II->getArgOperand(0));
1932 ICI.setOperand(1, RHS);
1944 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1945 /// We only handle extending casts so far.
1947 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1948 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1949 Value *LHSCIOp = LHSCI->getOperand(0);
1950 Type *SrcTy = LHSCIOp->getType();
1951 Type *DestTy = LHSCI->getType();
1954 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1955 // integer type is the same size as the pointer type.
1956 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
1957 DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1958 Value *RHSOp = nullptr;
1959 if (PtrToIntOperator *RHSC = dyn_cast<PtrToIntOperator>(ICI.getOperand(1))) {
1960 Value *RHSCIOp = RHSC->getOperand(0);
1961 if (RHSCIOp->getType()->getPointerAddressSpace() ==
1962 LHSCIOp->getType()->getPointerAddressSpace()) {
1963 RHSOp = RHSC->getOperand(0);
1964 // If the pointer types don't match, insert a bitcast.
1965 if (LHSCIOp->getType() != RHSOp->getType())
1966 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1968 } else if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1)))
1969 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1972 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1975 // The code below only handles extension cast instructions, so far.
1977 if (LHSCI->getOpcode() != Instruction::ZExt &&
1978 LHSCI->getOpcode() != Instruction::SExt)
1981 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1982 bool isSignedCmp = ICI.isSigned();
1984 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1985 // Not an extension from the same type?
1986 RHSCIOp = CI->getOperand(0);
1987 if (RHSCIOp->getType() != LHSCIOp->getType())
1990 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1991 // and the other is a zext), then we can't handle this.
1992 if (CI->getOpcode() != LHSCI->getOpcode())
1995 // Deal with equality cases early.
1996 if (ICI.isEquality())
1997 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1999 // A signed comparison of sign extended values simplifies into a
2000 // signed comparison.
2001 if (isSignedCmp && isSignedExt)
2002 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
2004 // The other three cases all fold into an unsigned comparison.
2005 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
2008 // If we aren't dealing with a constant on the RHS, exit early
2009 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
2013 // Compute the constant that would happen if we truncated to SrcTy then
2014 // reextended to DestTy.
2015 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
2016 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
2019 // If the re-extended constant didn't change...
2021 // Deal with equality cases early.
2022 if (ICI.isEquality())
2023 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
2025 // A signed comparison of sign extended values simplifies into a
2026 // signed comparison.
2027 if (isSignedExt && isSignedCmp)
2028 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
2030 // The other three cases all fold into an unsigned comparison.
2031 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
2034 // The re-extended constant changed so the constant cannot be represented
2035 // in the shorter type. Consequently, we cannot emit a simple comparison.
2036 // All the cases that fold to true or false will have already been handled
2037 // by SimplifyICmpInst, so only deal with the tricky case.
2039 if (isSignedCmp || !isSignedExt)
2042 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
2043 // should have been folded away previously and not enter in here.
2045 // We're performing an unsigned comp with a sign extended value.
2046 // This is true if the input is >= 0. [aka >s -1]
2047 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
2048 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
2050 // Finally, return the value computed.
2051 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
2052 return ReplaceInstUsesWith(ICI, Result);
2054 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
2055 return BinaryOperator::CreateNot(Result);
2058 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
2059 /// I = icmp ugt (add (add A, B), CI2), CI1
2060 /// If this is of the form:
2062 /// if (sum+128 >u 255)
2063 /// Then replace it with llvm.sadd.with.overflow.i8.
2065 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
2066 ConstantInt *CI2, ConstantInt *CI1,
2068 // The transformation we're trying to do here is to transform this into an
2069 // llvm.sadd.with.overflow. To do this, we have to replace the original add
2070 // with a narrower add, and discard the add-with-constant that is part of the
2071 // range check (if we can't eliminate it, this isn't profitable).
2073 // In order to eliminate the add-with-constant, the compare can be its only
2075 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
2076 if (!AddWithCst->hasOneUse()) return nullptr;
2078 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
2079 if (!CI2->getValue().isPowerOf2()) return nullptr;
2080 unsigned NewWidth = CI2->getValue().countTrailingZeros();
2081 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
2083 // The width of the new add formed is 1 more than the bias.
2086 // Check to see that CI1 is an all-ones value with NewWidth bits.
2087 if (CI1->getBitWidth() == NewWidth ||
2088 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
2091 // This is only really a signed overflow check if the inputs have been
2092 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
2093 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
2094 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
2095 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
2096 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
2099 // In order to replace the original add with a narrower
2100 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
2101 // and truncates that discard the high bits of the add. Verify that this is
2103 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
2104 for (User *U : OrigAdd->users()) {
2105 if (U == AddWithCst) continue;
2107 // Only accept truncates for now. We would really like a nice recursive
2108 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
2109 // chain to see which bits of a value are actually demanded. If the
2110 // original add had another add which was then immediately truncated, we
2111 // could still do the transformation.
2112 TruncInst *TI = dyn_cast<TruncInst>(U);
2113 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
2117 // If the pattern matches, truncate the inputs to the narrower type and
2118 // use the sadd_with_overflow intrinsic to efficiently compute both the
2119 // result and the overflow bit.
2120 Module *M = I.getParent()->getParent()->getParent();
2122 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
2123 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
2126 InstCombiner::BuilderTy *Builder = IC.Builder;
2128 // Put the new code above the original add, in case there are any uses of the
2129 // add between the add and the compare.
2130 Builder->SetInsertPoint(OrigAdd);
2132 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
2133 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
2134 CallInst *Call = Builder->CreateCall(F, {TruncA, TruncB}, "sadd");
2135 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
2136 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
2138 // The inner add was the result of the narrow add, zero extended to the
2139 // wider type. Replace it with the result computed by the intrinsic.
2140 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
2142 // The original icmp gets replaced with the overflow value.
2143 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
2146 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
2147 Value *RHS, Instruction &OrigI,
2148 Value *&Result, Constant *&Overflow) {
2149 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
2150 std::swap(LHS, RHS);
2152 auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
2154 Overflow = OverflowVal;
2156 Result->takeName(&OrigI);
2160 // If the overflow check was an add followed by a compare, the insertion point
2161 // may be pointing to the compare. We want to insert the new instructions
2162 // before the add in case there are uses of the add between the add and the
2164 Builder->SetInsertPoint(&OrigI);
2168 llvm_unreachable("bad overflow check kind!");
2170 case OCF_UNSIGNED_ADD: {
2171 OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
2172 if (OR == OverflowResult::NeverOverflows)
2173 return SetResult(Builder->CreateNUWAdd(LHS, RHS), Builder->getFalse(),
2176 if (OR == OverflowResult::AlwaysOverflows)
2177 return SetResult(Builder->CreateAdd(LHS, RHS), Builder->getTrue(), true);
2179 // FALL THROUGH uadd into sadd
2180 case OCF_SIGNED_ADD: {
2181 // X + 0 -> {X, false}
2182 if (match(RHS, m_Zero()))
2183 return SetResult(LHS, Builder->getFalse(), false);
2185 // We can strength reduce this signed add into a regular add if we can prove
2186 // that it will never overflow.
2187 if (OCF == OCF_SIGNED_ADD)
2188 if (WillNotOverflowSignedAdd(LHS, RHS, OrigI))
2189 return SetResult(Builder->CreateNSWAdd(LHS, RHS), Builder->getFalse(),
2194 case OCF_UNSIGNED_SUB:
2195 case OCF_SIGNED_SUB: {
2196 // X - 0 -> {X, false}
2197 if (match(RHS, m_Zero()))
2198 return SetResult(LHS, Builder->getFalse(), false);
2200 if (OCF == OCF_SIGNED_SUB) {
2201 if (WillNotOverflowSignedSub(LHS, RHS, OrigI))
2202 return SetResult(Builder->CreateNSWSub(LHS, RHS), Builder->getFalse(),
2205 if (WillNotOverflowUnsignedSub(LHS, RHS, OrigI))
2206 return SetResult(Builder->CreateNUWSub(LHS, RHS), Builder->getFalse(),
2212 case OCF_UNSIGNED_MUL: {
2213 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
2214 if (OR == OverflowResult::NeverOverflows)
2215 return SetResult(Builder->CreateNUWMul(LHS, RHS), Builder->getFalse(),
2217 if (OR == OverflowResult::AlwaysOverflows)
2218 return SetResult(Builder->CreateMul(LHS, RHS), Builder->getTrue(), true);
2220 case OCF_SIGNED_MUL:
2221 // X * undef -> undef
2222 if (isa<UndefValue>(RHS))
2223 return SetResult(RHS, UndefValue::get(Builder->getInt1Ty()), false);
2225 // X * 0 -> {0, false}
2226 if (match(RHS, m_Zero()))
2227 return SetResult(RHS, Builder->getFalse(), false);
2229 // X * 1 -> {X, false}
2230 if (match(RHS, m_One()))
2231 return SetResult(LHS, Builder->getFalse(), false);
2233 if (OCF == OCF_SIGNED_MUL)
2234 if (WillNotOverflowSignedMul(LHS, RHS, OrigI))
2235 return SetResult(Builder->CreateNSWMul(LHS, RHS), Builder->getFalse(),
2243 /// \brief Recognize and process idiom involving test for multiplication
2246 /// The caller has matched a pattern of the form:
2247 /// I = cmp u (mul(zext A, zext B), V
2248 /// The function checks if this is a test for overflow and if so replaces
2249 /// multiplication with call to 'mul.with.overflow' intrinsic.
2251 /// \param I Compare instruction.
2252 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
2253 /// the compare instruction. Must be of integer type.
2254 /// \param OtherVal The other argument of compare instruction.
2255 /// \returns Instruction which must replace the compare instruction, NULL if no
2256 /// replacement required.
2257 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
2258 Value *OtherVal, InstCombiner &IC) {
2259 // Don't bother doing this transformation for pointers, don't do it for
2261 if (!isa<IntegerType>(MulVal->getType()))
2264 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
2265 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
2266 auto *MulInstr = dyn_cast<Instruction>(MulVal);
2269 assert(MulInstr->getOpcode() == Instruction::Mul);
2271 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
2272 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
2273 assert(LHS->getOpcode() == Instruction::ZExt);
2274 assert(RHS->getOpcode() == Instruction::ZExt);
2275 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
2277 // Calculate type and width of the result produced by mul.with.overflow.
2278 Type *TyA = A->getType(), *TyB = B->getType();
2279 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
2280 WidthB = TyB->getPrimitiveSizeInBits();
2283 if (WidthB > WidthA) {
2291 // In order to replace the original mul with a narrower mul.with.overflow,
2292 // all uses must ignore upper bits of the product. The number of used low
2293 // bits must be not greater than the width of mul.with.overflow.
2294 if (MulVal->hasNUsesOrMore(2))
2295 for (User *U : MulVal->users()) {
2298 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2299 // Check if truncation ignores bits above MulWidth.
2300 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
2301 if (TruncWidth > MulWidth)
2303 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2304 // Check if AND ignores bits above MulWidth.
2305 if (BO->getOpcode() != Instruction::And)
2307 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
2308 const APInt &CVal = CI->getValue();
2309 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
2313 // Other uses prohibit this transformation.
2318 // Recognize patterns
2319 switch (I.getPredicate()) {
2320 case ICmpInst::ICMP_EQ:
2321 case ICmpInst::ICMP_NE:
2322 // Recognize pattern:
2323 // mulval = mul(zext A, zext B)
2324 // cmp eq/neq mulval, zext trunc mulval
2325 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
2326 if (Zext->hasOneUse()) {
2327 Value *ZextArg = Zext->getOperand(0);
2328 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
2329 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
2333 // Recognize pattern:
2334 // mulval = mul(zext A, zext B)
2335 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
2338 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
2339 if (ValToMask != MulVal)
2341 const APInt &CVal = CI->getValue() + 1;
2342 if (CVal.isPowerOf2()) {
2343 unsigned MaskWidth = CVal.logBase2();
2344 if (MaskWidth == MulWidth)
2345 break; // Recognized
2350 case ICmpInst::ICMP_UGT:
2351 // Recognize pattern:
2352 // mulval = mul(zext A, zext B)
2353 // cmp ugt mulval, max
2354 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2355 APInt MaxVal = APInt::getMaxValue(MulWidth);
2356 MaxVal = MaxVal.zext(CI->getBitWidth());
2357 if (MaxVal.eq(CI->getValue()))
2358 break; // Recognized
2362 case ICmpInst::ICMP_UGE:
2363 // Recognize pattern:
2364 // mulval = mul(zext A, zext B)
2365 // cmp uge mulval, max+1
2366 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2367 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2368 if (MaxVal.eq(CI->getValue()))
2369 break; // Recognized
2373 case ICmpInst::ICMP_ULE:
2374 // Recognize pattern:
2375 // mulval = mul(zext A, zext B)
2376 // cmp ule mulval, max
2377 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2378 APInt MaxVal = APInt::getMaxValue(MulWidth);
2379 MaxVal = MaxVal.zext(CI->getBitWidth());
2380 if (MaxVal.eq(CI->getValue()))
2381 break; // Recognized
2385 case ICmpInst::ICMP_ULT:
2386 // Recognize pattern:
2387 // mulval = mul(zext A, zext B)
2388 // cmp ule mulval, max + 1
2389 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2390 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2391 if (MaxVal.eq(CI->getValue()))
2392 break; // Recognized
2400 InstCombiner::BuilderTy *Builder = IC.Builder;
2401 Builder->SetInsertPoint(MulInstr);
2402 Module *M = I.getParent()->getParent()->getParent();
2404 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
2405 Value *MulA = A, *MulB = B;
2406 if (WidthA < MulWidth)
2407 MulA = Builder->CreateZExt(A, MulType);
2408 if (WidthB < MulWidth)
2409 MulB = Builder->CreateZExt(B, MulType);
2411 Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
2412 CallInst *Call = Builder->CreateCall(F, {MulA, MulB}, "umul");
2413 IC.Worklist.Add(MulInstr);
2415 // If there are uses of mul result other than the comparison, we know that
2416 // they are truncation or binary AND. Change them to use result of
2417 // mul.with.overflow and adjust properly mask/size.
2418 if (MulVal->hasNUsesOrMore(2)) {
2419 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
2420 for (User *U : MulVal->users()) {
2421 if (U == &I || U == OtherVal)
2423 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2424 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
2425 IC.ReplaceInstUsesWith(*TI, Mul);
2427 TI->setOperand(0, Mul);
2428 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2429 assert(BO->getOpcode() == Instruction::And);
2430 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
2431 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
2432 APInt ShortMask = CI->getValue().trunc(MulWidth);
2433 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
2435 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
2436 IC.Worklist.Add(Zext);
2437 IC.ReplaceInstUsesWith(*BO, Zext);
2439 llvm_unreachable("Unexpected Binary operation");
2441 IC.Worklist.Add(cast<Instruction>(U));
2444 if (isa<Instruction>(OtherVal))
2445 IC.Worklist.Add(cast<Instruction>(OtherVal));
2447 // The original icmp gets replaced with the overflow value, maybe inverted
2448 // depending on predicate.
2449 bool Inverse = false;
2450 switch (I.getPredicate()) {
2451 case ICmpInst::ICMP_NE:
2453 case ICmpInst::ICMP_EQ:
2456 case ICmpInst::ICMP_UGT:
2457 case ICmpInst::ICMP_UGE:
2458 if (I.getOperand(0) == MulVal)
2462 case ICmpInst::ICMP_ULT:
2463 case ICmpInst::ICMP_ULE:
2464 if (I.getOperand(1) == MulVal)
2469 llvm_unreachable("Unexpected predicate");
2472 Value *Res = Builder->CreateExtractValue(Call, 1);
2473 return BinaryOperator::CreateNot(Res);
2476 return ExtractValueInst::Create(Call, 1);
2479 // DemandedBitsLHSMask - When performing a comparison against a constant,
2480 // it is possible that not all the bits in the LHS are demanded. This helper
2481 // method computes the mask that IS demanded.
2482 static APInt DemandedBitsLHSMask(ICmpInst &I,
2483 unsigned BitWidth, bool isSignCheck) {
2485 return APInt::getSignBit(BitWidth);
2487 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2488 if (!CI) return APInt::getAllOnesValue(BitWidth);
2489 const APInt &RHS = CI->getValue();
2491 switch (I.getPredicate()) {
2492 // For a UGT comparison, we don't care about any bits that
2493 // correspond to the trailing ones of the comparand. The value of these
2494 // bits doesn't impact the outcome of the comparison, because any value
2495 // greater than the RHS must differ in a bit higher than these due to carry.
2496 case ICmpInst::ICMP_UGT: {
2497 unsigned trailingOnes = RHS.countTrailingOnes();
2498 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2502 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2503 // Any value less than the RHS must differ in a higher bit because of carries.
2504 case ICmpInst::ICMP_ULT: {
2505 unsigned trailingZeros = RHS.countTrailingZeros();
2506 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2511 return APInt::getAllOnesValue(BitWidth);
2516 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2517 /// should be swapped.
2518 /// The decision is based on how many times these two operands are reused
2519 /// as subtract operands and their positions in those instructions.
2520 /// The rational is that several architectures use the same instruction for
2521 /// both subtract and cmp, thus it is better if the order of those operands
2523 /// \return true if Op0 and Op1 should be swapped.
2524 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2525 const Value * Op1) {
2526 // Filter out pointer value as those cannot appears directly in subtract.
2527 // FIXME: we may want to go through inttoptrs or bitcasts.
2528 if (Op0->getType()->isPointerTy())
2530 // Count every uses of both Op0 and Op1 in a subtract.
2531 // Each time Op0 is the first operand, count -1: swapping is bad, the
2532 // subtract has already the same layout as the compare.
2533 // Each time Op0 is the second operand, count +1: swapping is good, the
2534 // subtract has a different layout as the compare.
2535 // At the end, if the benefit is greater than 0, Op0 should come second to
2536 // expose more CSE opportunities.
2537 int GlobalSwapBenefits = 0;
2538 for (const User *U : Op0->users()) {
2539 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
2540 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2542 // If Op0 is the first argument, this is not beneficial to swap the
2544 int LocalSwapBenefits = -1;
2545 unsigned Op1Idx = 1;
2546 if (BinOp->getOperand(Op1Idx) == Op0) {
2548 LocalSwapBenefits = 1;
2550 if (BinOp->getOperand(Op1Idx) != Op1)
2552 GlobalSwapBenefits += LocalSwapBenefits;
2554 return GlobalSwapBenefits > 0;
2557 /// \brief Check that one use is in the same block as the definition and all
2558 /// other uses are in blocks dominated by a given block
2560 /// \param DI Definition
2562 /// \param DB Block that must dominate all uses of \p DI outside
2563 /// the parent block
2564 /// \return true when \p UI is the only use of \p DI in the parent block
2565 /// and all other uses of \p DI are in blocks dominated by \p DB.
2567 bool InstCombiner::dominatesAllUses(const Instruction *DI,
2568 const Instruction *UI,
2569 const BasicBlock *DB) const {
2570 assert(DI && UI && "Instruction not defined\n");
2571 // ignore incomplete definitions
2572 if (!DI->getParent())
2574 // DI and UI must be in the same block
2575 if (DI->getParent() != UI->getParent())
2577 // Protect from self-referencing blocks
2578 if (DI->getParent() == DB)
2580 // DominatorTree available?
2583 for (const User *U : DI->users()) {
2584 auto *Usr = cast<Instruction>(U);
2585 if (Usr != UI && !DT->dominates(DB, Usr->getParent()))
2592 /// true when the instruction sequence within a block is select-cmp-br.
2594 static bool isChainSelectCmpBranch(const SelectInst *SI) {
2595 const BasicBlock *BB = SI->getParent();
2598 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
2599 if (!BI || BI->getNumSuccessors() != 2)
2601 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
2602 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
2608 /// \brief True when a select result is replaced by one of its operands
2609 /// in select-icmp sequence. This will eventually result in the elimination
2612 /// \param SI Select instruction
2613 /// \param Icmp Compare instruction
2614 /// \param SIOpd Operand that replaces the select
2617 /// - The replacement is global and requires dominator information
2618 /// - The caller is responsible for the actual replacement
2623 /// %4 = select i1 %3, %C* %0, %C* null
2624 /// %5 = icmp eq %C* %4, null
2625 /// br i1 %5, label %9, label %7
2627 /// ; <label>:7 ; preds = %entry
2628 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
2631 /// can be transformed to
2633 /// %5 = icmp eq %C* %0, null
2634 /// %6 = select i1 %3, i1 %5, i1 true
2635 /// br i1 %6, label %9, label %7
2637 /// ; <label>:7 ; preds = %entry
2638 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
2640 /// Similar when the first operand of the select is a constant or/and
2641 /// the compare is for not equal rather than equal.
2643 /// NOTE: The function is only called when the select and compare constants
2644 /// are equal, the optimization can work only for EQ predicates. This is not a
2645 /// major restriction since a NE compare should be 'normalized' to an equal
2646 /// compare, which usually happens in the combiner and test case
2647 /// select-cmp-br.ll
2649 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
2650 const ICmpInst *Icmp,
2651 const unsigned SIOpd) {
2652 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
2653 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
2654 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
2655 // The check for the unique predecessor is not the best that can be
2656 // done. But it protects efficiently against cases like when SI's
2657 // home block has two successors, Succ and Succ1, and Succ1 predecessor
2658 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
2659 // replaced can be reached on either path. So the uniqueness check
2660 // guarantees that the path all uses of SI (outside SI's parent) are on
2661 // is disjoint from all other paths out of SI. But that information
2662 // is more expensive to compute, and the trade-off here is in favor
2664 if (Succ->getUniquePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
2666 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
2673 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2674 bool Changed = false;
2675 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2676 unsigned Op0Cplxity = getComplexity(Op0);
2677 unsigned Op1Cplxity = getComplexity(Op1);
2679 /// Orders the operands of the compare so that they are listed from most
2680 /// complex to least complex. This puts constants before unary operators,
2681 /// before binary operators.
2682 if (Op0Cplxity < Op1Cplxity ||
2683 (Op0Cplxity == Op1Cplxity &&
2684 swapMayExposeCSEOpportunities(Op0, Op1))) {
2686 std::swap(Op0, Op1);
2691 SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AC, &I))
2692 return ReplaceInstUsesWith(I, V);
2694 // comparing -val or val with non-zero is the same as just comparing val
2695 // ie, abs(val) != 0 -> val != 0
2696 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2698 Value *Cond, *SelectTrue, *SelectFalse;
2699 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2700 m_Value(SelectFalse)))) {
2701 if (Value *V = dyn_castNegVal(SelectTrue)) {
2702 if (V == SelectFalse)
2703 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2705 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2706 if (V == SelectTrue)
2707 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2712 Type *Ty = Op0->getType();
2714 // icmp's with boolean values can always be turned into bitwise operations
2715 if (Ty->isIntegerTy(1)) {
2716 switch (I.getPredicate()) {
2717 default: llvm_unreachable("Invalid icmp instruction!");
2718 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2719 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2720 return BinaryOperator::CreateNot(Xor);
2722 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2723 return BinaryOperator::CreateXor(Op0, Op1);
2725 case ICmpInst::ICMP_UGT:
2726 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2728 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2729 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2730 return BinaryOperator::CreateAnd(Not, Op1);
2732 case ICmpInst::ICMP_SGT:
2733 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2735 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2736 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2737 return BinaryOperator::CreateAnd(Not, Op0);
2739 case ICmpInst::ICMP_UGE:
2740 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2742 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2743 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2744 return BinaryOperator::CreateOr(Not, Op1);
2746 case ICmpInst::ICMP_SGE:
2747 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2749 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2750 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2751 return BinaryOperator::CreateOr(Not, Op0);
2756 unsigned BitWidth = 0;
2757 if (Ty->isIntOrIntVectorTy())
2758 BitWidth = Ty->getScalarSizeInBits();
2759 else // Get pointer size.
2760 BitWidth = DL.getTypeSizeInBits(Ty->getScalarType());
2762 bool isSignBit = false;
2764 // See if we are doing a comparison with a constant.
2765 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2766 Value *A = nullptr, *B = nullptr;
2768 // Match the following pattern, which is a common idiom when writing
2769 // overflow-safe integer arithmetic function. The source performs an
2770 // addition in wider type, and explicitly checks for overflow using
2771 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2772 // sadd_with_overflow intrinsic.
2774 // TODO: This could probably be generalized to handle other overflow-safe
2775 // operations if we worked out the formulas to compute the appropriate
2779 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2781 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2782 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2783 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2784 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2788 // The following transforms are only 'worth it' if the only user of the
2789 // subtraction is the icmp.
2790 if (Op0->hasOneUse()) {
2791 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2792 if (I.isEquality() && CI->isZero() &&
2793 match(Op0, m_Sub(m_Value(A), m_Value(B))))
2794 return new ICmpInst(I.getPredicate(), A, B);
2796 // (icmp sgt (sub nsw A B), -1) -> (icmp sge A, B)
2797 if (I.getPredicate() == ICmpInst::ICMP_SGT && CI->isAllOnesValue() &&
2798 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2799 return new ICmpInst(ICmpInst::ICMP_SGE, A, B);
2801 // (icmp sgt (sub nsw A B), 0) -> (icmp sgt A, B)
2802 if (I.getPredicate() == ICmpInst::ICMP_SGT && CI->isZero() &&
2803 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2804 return new ICmpInst(ICmpInst::ICMP_SGT, A, B);
2806 // (icmp slt (sub nsw A B), 0) -> (icmp slt A, B)
2807 if (I.getPredicate() == ICmpInst::ICMP_SLT && CI->isZero() &&
2808 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2809 return new ICmpInst(ICmpInst::ICMP_SLT, A, B);
2811 // (icmp slt (sub nsw A B), 1) -> (icmp sle A, B)
2812 if (I.getPredicate() == ICmpInst::ICMP_SLT && CI->isOne() &&
2813 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2814 return new ICmpInst(ICmpInst::ICMP_SLE, A, B);
2817 // If we have an icmp le or icmp ge instruction, turn it into the
2818 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2819 // them being folded in the code below. The SimplifyICmpInst code has
2820 // already handled the edge cases for us, so we just assert on them.
2821 switch (I.getPredicate()) {
2823 case ICmpInst::ICMP_ULE:
2824 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2825 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2826 Builder->getInt(CI->getValue()+1));
2827 case ICmpInst::ICMP_SLE:
2828 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2829 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2830 Builder->getInt(CI->getValue()+1));
2831 case ICmpInst::ICMP_UGE:
2832 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2833 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2834 Builder->getInt(CI->getValue()-1));
2835 case ICmpInst::ICMP_SGE:
2836 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2837 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2838 Builder->getInt(CI->getValue()-1));
2841 if (I.isEquality()) {
2843 if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
2844 match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
2845 // (icmp eq/ne (ashr/lshr const2, A), const1)
2846 if (Instruction *Inst = FoldICmpCstShrCst(I, Op0, A, CI, CI2))
2849 if (match(Op0, m_Shl(m_ConstantInt(CI2), m_Value(A)))) {
2850 // (icmp eq/ne (shl const2, A), const1)
2851 if (Instruction *Inst = FoldICmpCstShlCst(I, Op0, A, CI, CI2))
2856 // If this comparison is a normal comparison, it demands all
2857 // bits, if it is a sign bit comparison, it only demands the sign bit.
2859 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2862 // See if we can fold the comparison based on range information we can get
2863 // by checking whether bits are known to be zero or one in the input.
2864 if (BitWidth != 0) {
2865 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2866 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2868 if (SimplifyDemandedBits(I.getOperandUse(0),
2869 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2870 Op0KnownZero, Op0KnownOne, 0))
2872 if (SimplifyDemandedBits(I.getOperandUse(1),
2873 APInt::getAllOnesValue(BitWidth), Op1KnownZero,
2877 // Given the known and unknown bits, compute a range that the LHS could be
2878 // in. Compute the Min, Max and RHS values based on the known bits. For the
2879 // EQ and NE we use unsigned values.
2880 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2881 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2883 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2885 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2888 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2890 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2894 // If Min and Max are known to be the same, then SimplifyDemandedBits
2895 // figured out that the LHS is a constant. Just constant fold this now so
2896 // that code below can assume that Min != Max.
2897 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2898 return new ICmpInst(I.getPredicate(),
2899 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2900 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2901 return new ICmpInst(I.getPredicate(), Op0,
2902 ConstantInt::get(Op1->getType(), Op1Min));
2904 // Based on the range information we know about the LHS, see if we can
2905 // simplify this comparison. For example, (x&4) < 8 is always true.
2906 switch (I.getPredicate()) {
2907 default: llvm_unreachable("Unknown icmp opcode!");
2908 case ICmpInst::ICMP_EQ: {
2909 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2910 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2912 // If all bits are known zero except for one, then we know at most one
2913 // bit is set. If the comparison is against zero, then this is a check
2914 // to see if *that* bit is set.
2915 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2916 if (~Op1KnownZero == 0) {
2917 // If the LHS is an AND with the same constant, look through it.
2918 Value *LHS = nullptr;
2919 ConstantInt *LHSC = nullptr;
2920 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2921 LHSC->getValue() != Op0KnownZeroInverted)
2924 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2925 // then turn "((1 << x)&8) == 0" into "x != 3".
2926 // or turn "((1 << x)&7) == 0" into "x > 2".
2928 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2929 APInt ValToCheck = Op0KnownZeroInverted;
2930 if (ValToCheck.isPowerOf2()) {
2931 unsigned CmpVal = ValToCheck.countTrailingZeros();
2932 return new ICmpInst(ICmpInst::ICMP_NE, X,
2933 ConstantInt::get(X->getType(), CmpVal));
2934 } else if ((++ValToCheck).isPowerOf2()) {
2935 unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
2936 return new ICmpInst(ICmpInst::ICMP_UGT, X,
2937 ConstantInt::get(X->getType(), CmpVal));
2941 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2942 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2944 if (Op0KnownZeroInverted == 1 &&
2945 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2946 return new ICmpInst(ICmpInst::ICMP_NE, X,
2947 ConstantInt::get(X->getType(),
2948 CI->countTrailingZeros()));
2953 case ICmpInst::ICMP_NE: {
2954 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2955 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2957 // If all bits are known zero except for one, then we know at most one
2958 // bit is set. If the comparison is against zero, then this is a check
2959 // to see if *that* bit is set.
2960 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2961 if (~Op1KnownZero == 0) {
2962 // If the LHS is an AND with the same constant, look through it.
2963 Value *LHS = nullptr;
2964 ConstantInt *LHSC = nullptr;
2965 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2966 LHSC->getValue() != Op0KnownZeroInverted)
2969 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2970 // then turn "((1 << x)&8) != 0" into "x == 3".
2971 // or turn "((1 << x)&7) != 0" into "x < 3".
2973 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2974 APInt ValToCheck = Op0KnownZeroInverted;
2975 if (ValToCheck.isPowerOf2()) {
2976 unsigned CmpVal = ValToCheck.countTrailingZeros();
2977 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2978 ConstantInt::get(X->getType(), CmpVal));
2979 } else if ((++ValToCheck).isPowerOf2()) {
2980 unsigned CmpVal = ValToCheck.countTrailingZeros();
2981 return new ICmpInst(ICmpInst::ICMP_ULT, X,
2982 ConstantInt::get(X->getType(), CmpVal));
2986 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2987 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2989 if (Op0KnownZeroInverted == 1 &&
2990 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2991 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2992 ConstantInt::get(X->getType(),
2993 CI->countTrailingZeros()));
2998 case ICmpInst::ICMP_ULT:
2999 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
3000 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3001 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
3002 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3003 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
3004 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
3005 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3006 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
3007 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
3008 Builder->getInt(CI->getValue()-1));
3010 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
3011 if (CI->isMinValue(true))
3012 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
3013 Constant::getAllOnesValue(Op0->getType()));
3016 case ICmpInst::ICMP_UGT:
3017 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
3018 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3019 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
3020 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3022 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
3023 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
3024 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3025 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
3026 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
3027 Builder->getInt(CI->getValue()+1));
3029 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
3030 if (CI->isMaxValue(true))
3031 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
3032 Constant::getNullValue(Op0->getType()));
3035 case ICmpInst::ICMP_SLT:
3036 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
3037 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3038 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
3039 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3040 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
3041 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
3042 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3043 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
3044 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
3045 Builder->getInt(CI->getValue()-1));
3048 case ICmpInst::ICMP_SGT:
3049 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
3050 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3051 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
3052 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3054 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
3055 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
3056 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3057 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
3058 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
3059 Builder->getInt(CI->getValue()+1));
3062 case ICmpInst::ICMP_SGE:
3063 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
3064 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
3065 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3066 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
3067 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3069 case ICmpInst::ICMP_SLE:
3070 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
3071 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
3072 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3073 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
3074 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3076 case ICmpInst::ICMP_UGE:
3077 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
3078 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
3079 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3080 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
3081 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3083 case ICmpInst::ICMP_ULE:
3084 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
3085 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
3086 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3087 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
3088 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3092 // Turn a signed comparison into an unsigned one if both operands
3093 // are known to have the same sign.
3095 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
3096 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
3097 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
3100 // Test if the ICmpInst instruction is used exclusively by a select as
3101 // part of a minimum or maximum operation. If so, refrain from doing
3102 // any other folding. This helps out other analyses which understand
3103 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
3104 // and CodeGen. And in this case, at least one of the comparison
3105 // operands has at least one user besides the compare (the select),
3106 // which would often largely negate the benefit of folding anyway.
3108 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
3109 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
3110 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
3113 // See if we are doing a comparison between a constant and an instruction that
3114 // can be folded into the comparison.
3115 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3116 // Since the RHS is a ConstantInt (CI), if the left hand side is an
3117 // instruction, see if that instruction also has constants so that the
3118 // instruction can be folded into the icmp
3119 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3120 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
3124 // Handle icmp with constant (but not simple integer constant) RHS
3125 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3126 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3127 switch (LHSI->getOpcode()) {
3128 case Instruction::GetElementPtr:
3129 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3130 if (RHSC->isNullValue() &&
3131 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3132 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
3133 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3135 case Instruction::PHI:
3136 // Only fold icmp into the PHI if the phi and icmp are in the same
3137 // block. If in the same block, we're encouraging jump threading. If
3138 // not, we are just pessimizing the code by making an i1 phi.
3139 if (LHSI->getParent() == I.getParent())
3140 if (Instruction *NV = FoldOpIntoPhi(I))
3143 case Instruction::Select: {
3144 // If either operand of the select is a constant, we can fold the
3145 // comparison into the select arms, which will cause one to be
3146 // constant folded and the select turned into a bitwise or.
3147 Value *Op1 = nullptr, *Op2 = nullptr;
3148 ConstantInt *CI = 0;
3149 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3150 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3151 CI = dyn_cast<ConstantInt>(Op1);
3153 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3154 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3155 CI = dyn_cast<ConstantInt>(Op2);
3158 // We only want to perform this transformation if it will not lead to
3159 // additional code. This is true if either both sides of the select
3160 // fold to a constant (in which case the icmp is replaced with a select
3161 // which will usually simplify) or this is the only user of the
3162 // select (in which case we are trading a select+icmp for a simpler
3163 // select+icmp) or all uses of the select can be replaced based on
3164 // dominance information ("Global cases").
3165 bool Transform = false;
3168 else if (Op1 || Op2) {
3170 if (LHSI->hasOneUse())
3173 else if (CI && !CI->isZero())
3174 // When Op1 is constant try replacing select with second operand.
3175 // Otherwise Op2 is constant and try replacing select with first
3177 Transform = replacedSelectWithOperand(cast<SelectInst>(LHSI), &I,
3182 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
3185 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
3187 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3191 case Instruction::IntToPtr:
3192 // icmp pred inttoptr(X), null -> icmp pred X, 0
3193 if (RHSC->isNullValue() &&
3194 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3195 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
3196 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3199 case Instruction::Load:
3200 // Try to optimize things like "A[i] > 4" to index computations.
3201 if (GetElementPtrInst *GEP =
3202 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3203 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3204 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3205 !cast<LoadInst>(LHSI)->isVolatile())
3206 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3213 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
3214 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
3215 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
3217 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
3218 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
3219 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
3222 // Test to see if the operands of the icmp are casted versions of other
3223 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
3225 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
3226 if (Op0->getType()->isPointerTy() &&
3227 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
3228 // We keep moving the cast from the left operand over to the right
3229 // operand, where it can often be eliminated completely.
3230 Op0 = CI->getOperand(0);
3232 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
3233 // so eliminate it as well.
3234 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
3235 Op1 = CI2->getOperand(0);
3237 // If Op1 is a constant, we can fold the cast into the constant.
3238 if (Op0->getType() != Op1->getType()) {
3239 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3240 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
3242 // Otherwise, cast the RHS right before the icmp
3243 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
3246 return new ICmpInst(I.getPredicate(), Op0, Op1);
3250 if (isa<CastInst>(Op0)) {
3251 // Handle the special case of: icmp (cast bool to X), <cst>
3252 // This comes up when you have code like
3255 // For generality, we handle any zero-extension of any operand comparison
3256 // with a constant or another cast from the same type.
3257 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
3258 if (Instruction *R = visitICmpInstWithCastAndCast(I))
3262 // Special logic for binary operators.
3263 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3264 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3266 CmpInst::Predicate Pred = I.getPredicate();
3267 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3268 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3269 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
3270 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3271 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3272 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3273 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
3274 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3275 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3277 // Analyze the case when either Op0 or Op1 is an add instruction.
3278 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3279 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3280 if (BO0 && BO0->getOpcode() == Instruction::Add)
3281 A = BO0->getOperand(0), B = BO0->getOperand(1);
3282 if (BO1 && BO1->getOpcode() == Instruction::Add)
3283 C = BO1->getOperand(0), D = BO1->getOperand(1);
3285 // icmp (X+cst) < 0 --> X < -cst
3286 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred) && match(Op1, m_Zero()))
3287 if (ConstantInt *RHSC = dyn_cast_or_null<ConstantInt>(B))
3288 if (!RHSC->isMinValue(/*isSigned=*/true))
3289 return new ICmpInst(Pred, A, ConstantExpr::getNeg(RHSC));
3291 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3292 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3293 return new ICmpInst(Pred, A == Op1 ? B : A,
3294 Constant::getNullValue(Op1->getType()));
3296 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3297 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3298 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3301 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3302 if (A && C && (A == C || A == D || B == C || B == D) &&
3303 NoOp0WrapProblem && NoOp1WrapProblem &&
3304 // Try not to increase register pressure.
3305 BO0->hasOneUse() && BO1->hasOneUse()) {
3306 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3309 // C + B == C + D -> B == D
3312 } else if (A == D) {
3313 // D + B == C + D -> B == C
3316 } else if (B == C) {
3317 // A + C == C + D -> A == D
3322 // A + D == C + D -> A == C
3326 return new ICmpInst(Pred, Y, Z);
3329 // icmp slt (X + -1), Y -> icmp sle X, Y
3330 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3331 match(B, m_AllOnes()))
3332 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3334 // icmp sge (X + -1), Y -> icmp sgt X, Y
3335 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3336 match(B, m_AllOnes()))
3337 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3339 // icmp sle (X + 1), Y -> icmp slt X, Y
3340 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
3342 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3344 // icmp sgt (X + 1), Y -> icmp sge X, Y
3345 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
3347 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3349 // if C1 has greater magnitude than C2:
3350 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3351 // s.t. C3 = C1 - C2
3353 // if C2 has greater magnitude than C1:
3354 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3355 // s.t. C3 = C2 - C1
3356 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3357 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3358 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3359 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3360 const APInt &AP1 = C1->getValue();
3361 const APInt &AP2 = C2->getValue();
3362 if (AP1.isNegative() == AP2.isNegative()) {
3363 APInt AP1Abs = C1->getValue().abs();
3364 APInt AP2Abs = C2->getValue().abs();
3365 if (AP1Abs.uge(AP2Abs)) {
3366 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
3367 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
3368 return new ICmpInst(Pred, NewAdd, C);
3370 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
3371 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
3372 return new ICmpInst(Pred, A, NewAdd);
3378 // Analyze the case when either Op0 or Op1 is a sub instruction.
3379 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3380 A = nullptr; B = nullptr; C = nullptr; D = nullptr;
3381 if (BO0 && BO0->getOpcode() == Instruction::Sub)
3382 A = BO0->getOperand(0), B = BO0->getOperand(1);
3383 if (BO1 && BO1->getOpcode() == Instruction::Sub)
3384 C = BO1->getOperand(0), D = BO1->getOperand(1);
3386 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3387 if (A == Op1 && NoOp0WrapProblem)
3388 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3390 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3391 if (C == Op0 && NoOp1WrapProblem)
3392 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3394 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3395 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3396 // Try not to increase register pressure.
3397 BO0->hasOneUse() && BO1->hasOneUse())
3398 return new ICmpInst(Pred, A, C);
3400 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3401 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3402 // Try not to increase register pressure.
3403 BO0->hasOneUse() && BO1->hasOneUse())
3404 return new ICmpInst(Pred, D, B);
3406 // icmp (0-X) < cst --> x > -cst
3407 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3409 if (match(BO0, m_Neg(m_Value(X))))
3410 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3411 if (!RHSC->isMinValue(/*isSigned=*/true))
3412 return new ICmpInst(I.getSwappedPredicate(), X,
3413 ConstantExpr::getNeg(RHSC));
3416 BinaryOperator *SRem = nullptr;
3417 // icmp (srem X, Y), Y
3418 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
3419 Op1 == BO0->getOperand(1))
3421 // icmp Y, (srem X, Y)
3422 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3423 Op0 == BO1->getOperand(1))
3426 // We don't check hasOneUse to avoid increasing register pressure because
3427 // the value we use is the same value this instruction was already using.
3428 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3430 case ICmpInst::ICMP_EQ:
3431 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3432 case ICmpInst::ICMP_NE:
3433 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3434 case ICmpInst::ICMP_SGT:
3435 case ICmpInst::ICMP_SGE:
3436 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3437 Constant::getAllOnesValue(SRem->getType()));
3438 case ICmpInst::ICMP_SLT:
3439 case ICmpInst::ICMP_SLE:
3440 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3441 Constant::getNullValue(SRem->getType()));
3445 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
3446 BO0->hasOneUse() && BO1->hasOneUse() &&
3447 BO0->getOperand(1) == BO1->getOperand(1)) {
3448 switch (BO0->getOpcode()) {
3450 case Instruction::Add:
3451 case Instruction::Sub:
3452 case Instruction::Xor:
3453 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3454 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3455 BO1->getOperand(0));
3456 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
3457 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3458 if (CI->getValue().isSignBit()) {
3459 ICmpInst::Predicate Pred = I.isSigned()
3460 ? I.getUnsignedPredicate()
3461 : I.getSignedPredicate();
3462 return new ICmpInst(Pred, BO0->getOperand(0),
3463 BO1->getOperand(0));
3466 if (CI->isMaxValue(true)) {
3467 ICmpInst::Predicate Pred = I.isSigned()
3468 ? I.getUnsignedPredicate()
3469 : I.getSignedPredicate();
3470 Pred = I.getSwappedPredicate(Pred);
3471 return new ICmpInst(Pred, BO0->getOperand(0),
3472 BO1->getOperand(0));
3476 case Instruction::Mul:
3477 if (!I.isEquality())
3480 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3481 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
3482 // Mask = -1 >> count-trailing-zeros(Cst).
3483 if (!CI->isZero() && !CI->isOne()) {
3484 const APInt &AP = CI->getValue();
3485 ConstantInt *Mask = ConstantInt::get(I.getContext(),
3486 APInt::getLowBitsSet(AP.getBitWidth(),
3488 AP.countTrailingZeros()));
3489 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
3490 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
3491 return new ICmpInst(I.getPredicate(), And1, And2);
3495 case Instruction::UDiv:
3496 case Instruction::LShr:
3500 case Instruction::SDiv:
3501 case Instruction::AShr:
3502 if (!BO0->isExact() || !BO1->isExact())
3504 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3505 BO1->getOperand(0));
3506 case Instruction::Shl: {
3507 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3508 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3511 if (!NSW && I.isSigned())
3513 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3514 BO1->getOperand(0));
3520 // Transform A & (L - 1) `ult` L --> L != 0
3521 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3523 m_CombineOr(m_And(m_Value(), LSubOne), m_And(LSubOne, m_Value()));
3525 if (match(BO0, BitwiseAnd) && I.getPredicate() == ICmpInst::ICMP_ULT) {
3526 auto *Zero = Constant::getNullValue(BO0->getType());
3527 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3533 // Transform (A & ~B) == 0 --> (A & B) != 0
3534 // and (A & ~B) != 0 --> (A & B) == 0
3535 // if A is a power of 2.
3536 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
3537 match(Op1, m_Zero()) &&
3538 isKnownToBeAPowerOfTwo(A, DL, false, 0, AC, &I, DT) && I.isEquality())
3539 return new ICmpInst(I.getInversePredicate(),
3540 Builder->CreateAnd(A, B),
3543 // ~x < ~y --> y < x
3544 // ~x < cst --> ~cst < x
3545 if (match(Op0, m_Not(m_Value(A)))) {
3546 if (match(Op1, m_Not(m_Value(B))))
3547 return new ICmpInst(I.getPredicate(), B, A);
3548 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3549 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
3552 Instruction *AddI = nullptr;
3553 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
3554 m_Instruction(AddI))) &&
3555 isa<IntegerType>(A->getType())) {
3558 if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result,
3560 ReplaceInstUsesWith(*AddI, Result);
3561 return ReplaceInstUsesWith(I, Overflow);
3565 // (zext a) * (zext b) --> llvm.umul.with.overflow.
3566 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3567 if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
3570 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3571 if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
3576 if (I.isEquality()) {
3577 Value *A, *B, *C, *D;
3579 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3580 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3581 Value *OtherVal = A == Op1 ? B : A;
3582 return new ICmpInst(I.getPredicate(), OtherVal,
3583 Constant::getNullValue(A->getType()));
3586 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3587 // A^c1 == C^c2 --> A == C^(c1^c2)
3588 ConstantInt *C1, *C2;
3589 if (match(B, m_ConstantInt(C1)) &&
3590 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
3591 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3592 Value *Xor = Builder->CreateXor(C, NC);
3593 return new ICmpInst(I.getPredicate(), A, Xor);
3596 // A^B == A^D -> B == D
3597 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
3598 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
3599 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
3600 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
3604 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
3605 (A == Op0 || B == Op0)) {
3606 // A == (A^B) -> B == 0
3607 Value *OtherVal = A == Op0 ? B : A;
3608 return new ICmpInst(I.getPredicate(), OtherVal,
3609 Constant::getNullValue(A->getType()));
3612 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3613 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3614 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3615 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3618 X = B; Y = D; Z = A;
3619 } else if (A == D) {
3620 X = B; Y = C; Z = A;
3621 } else if (B == C) {
3622 X = A; Y = D; Z = B;
3623 } else if (B == D) {
3624 X = A; Y = C; Z = B;
3627 if (X) { // Build (X^Y) & Z
3628 Op1 = Builder->CreateXor(X, Y);
3629 Op1 = Builder->CreateAnd(Op1, Z);
3630 I.setOperand(0, Op1);
3631 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3636 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3637 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3639 if ((Op0->hasOneUse() &&
3640 match(Op0, m_ZExt(m_Value(A))) &&
3641 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3642 (Op1->hasOneUse() &&
3643 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3644 match(Op1, m_ZExt(m_Value(A))))) {
3645 APInt Pow2 = Cst1->getValue() + 1;
3646 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3647 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3648 return new ICmpInst(I.getPredicate(), A,
3649 Builder->CreateTrunc(B, A->getType()));
3652 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3653 // For lshr and ashr pairs.
3654 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3655 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3656 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3657 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3658 unsigned TypeBits = Cst1->getBitWidth();
3659 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3660 if (ShAmt < TypeBits && ShAmt != 0) {
3661 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
3662 ? ICmpInst::ICMP_UGE
3663 : ICmpInst::ICMP_ULT;
3664 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3665 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3666 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
3670 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3671 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3672 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3673 unsigned TypeBits = Cst1->getBitWidth();
3674 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3675 if (ShAmt < TypeBits && ShAmt != 0) {
3676 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3677 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3678 Value *And = Builder->CreateAnd(Xor, Builder->getInt(AndVal),
3679 I.getName() + ".mask");
3680 return new ICmpInst(I.getPredicate(), And,
3681 Constant::getNullValue(Cst1->getType()));
3685 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3686 // "icmp (and X, mask), cst"
3688 if (Op0->hasOneUse() &&
3689 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
3690 m_ConstantInt(ShAmt))))) &&
3691 match(Op1, m_ConstantInt(Cst1)) &&
3692 // Only do this when A has multiple uses. This is most important to do
3693 // when it exposes other optimizations.
3695 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3697 if (ShAmt < ASize) {
3699 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3702 APInt CmpV = Cst1->getValue().zext(ASize);
3705 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3706 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3711 // The 'cmpxchg' instruction returns an aggregate containing the old value and
3712 // an i1 which indicates whether or not we successfully did the swap.
3714 // Replace comparisons between the old value and the expected value with the
3715 // indicator that 'cmpxchg' returns.
3717 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
3718 // spuriously fail. In those cases, the old value may equal the expected
3719 // value but it is possible for the swap to not occur.
3720 if (I.getPredicate() == ICmpInst::ICMP_EQ)
3721 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
3722 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
3723 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
3725 return ExtractValueInst::Create(ACXI, 1);
3728 Value *X; ConstantInt *Cst;
3730 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3731 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3734 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3735 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3737 return Changed ? &I : nullptr;
3740 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3741 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3744 if (!isa<ConstantFP>(RHSC)) return nullptr;
3745 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3747 // Get the width of the mantissa. We don't want to hack on conversions that
3748 // might lose information from the integer, e.g. "i64 -> float"
3749 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3750 if (MantissaWidth == -1) return nullptr; // Unknown.
3752 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3754 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3756 if (I.isEquality()) {
3757 FCmpInst::Predicate P = I.getPredicate();
3758 bool IsExact = false;
3759 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
3760 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
3762 // If the floating point constant isn't an integer value, we know if we will
3763 // ever compare equal / not equal to it.
3765 // TODO: Can never be -0.0 and other non-representable values
3766 APFloat RHSRoundInt(RHS);
3767 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
3768 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
3769 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
3770 return ReplaceInstUsesWith(I, Builder->getFalse());
3772 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
3773 return ReplaceInstUsesWith(I, Builder->getTrue());
3777 // TODO: If the constant is exactly representable, is it always OK to do
3778 // equality compares as integer?
3781 // Check to see that the input is converted from an integer type that is small
3782 // enough that preserves all bits. TODO: check here for "known" sign bits.
3783 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3784 unsigned InputSize = IntTy->getScalarSizeInBits();
3786 // Following test does NOT adjust InputSize downwards for signed inputs,
3787 // because the most negative value still requires all the mantissa bits
3788 // to distinguish it from one less than that value.
3789 if ((int)InputSize > MantissaWidth) {
3790 // Conversion would lose accuracy. Check if loss can impact comparison.
3791 int Exp = ilogb(RHS);
3792 if (Exp == APFloat::IEK_Inf) {
3793 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
3794 if (MaxExponent < (int)InputSize - !LHSUnsigned)
3795 // Conversion could create infinity.
3798 // Note that if RHS is zero or NaN, then Exp is negative
3799 // and first condition is trivially false.
3800 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
3801 // Conversion could affect comparison.
3806 // Otherwise, we can potentially simplify the comparison. We know that it
3807 // will always come through as an integer value and we know the constant is
3808 // not a NAN (it would have been previously simplified).
3809 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3811 ICmpInst::Predicate Pred;
3812 switch (I.getPredicate()) {
3813 default: llvm_unreachable("Unexpected predicate!");
3814 case FCmpInst::FCMP_UEQ:
3815 case FCmpInst::FCMP_OEQ:
3816 Pred = ICmpInst::ICMP_EQ;
3818 case FCmpInst::FCMP_UGT:
3819 case FCmpInst::FCMP_OGT:
3820 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3822 case FCmpInst::FCMP_UGE:
3823 case FCmpInst::FCMP_OGE:
3824 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3826 case FCmpInst::FCMP_ULT:
3827 case FCmpInst::FCMP_OLT:
3828 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3830 case FCmpInst::FCMP_ULE:
3831 case FCmpInst::FCMP_OLE:
3832 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3834 case FCmpInst::FCMP_UNE:
3835 case FCmpInst::FCMP_ONE:
3836 Pred = ICmpInst::ICMP_NE;
3838 case FCmpInst::FCMP_ORD:
3839 return ReplaceInstUsesWith(I, Builder->getTrue());
3840 case FCmpInst::FCMP_UNO:
3841 return ReplaceInstUsesWith(I, Builder->getFalse());
3844 // Now we know that the APFloat is a normal number, zero or inf.
3846 // See if the FP constant is too large for the integer. For example,
3847 // comparing an i8 to 300.0.
3848 unsigned IntWidth = IntTy->getScalarSizeInBits();
3851 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3852 // and large values.
3853 APFloat SMax(RHS.getSemantics());
3854 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3855 APFloat::rmNearestTiesToEven);
3856 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3857 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3858 Pred == ICmpInst::ICMP_SLE)
3859 return ReplaceInstUsesWith(I, Builder->getTrue());
3860 return ReplaceInstUsesWith(I, Builder->getFalse());
3863 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3864 // +INF and large values.
3865 APFloat UMax(RHS.getSemantics());
3866 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3867 APFloat::rmNearestTiesToEven);
3868 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3869 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3870 Pred == ICmpInst::ICMP_ULE)
3871 return ReplaceInstUsesWith(I, Builder->getTrue());
3872 return ReplaceInstUsesWith(I, Builder->getFalse());
3877 // See if the RHS value is < SignedMin.
3878 APFloat SMin(RHS.getSemantics());
3879 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3880 APFloat::rmNearestTiesToEven);
3881 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3882 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3883 Pred == ICmpInst::ICMP_SGE)
3884 return ReplaceInstUsesWith(I, Builder->getTrue());
3885 return ReplaceInstUsesWith(I, Builder->getFalse());
3888 // See if the RHS value is < UnsignedMin.
3889 APFloat SMin(RHS.getSemantics());
3890 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3891 APFloat::rmNearestTiesToEven);
3892 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3893 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3894 Pred == ICmpInst::ICMP_UGE)
3895 return ReplaceInstUsesWith(I, Builder->getTrue());
3896 return ReplaceInstUsesWith(I, Builder->getFalse());
3900 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3901 // [0, UMAX], but it may still be fractional. See if it is fractional by
3902 // casting the FP value to the integer value and back, checking for equality.
3903 // Don't do this for zero, because -0.0 is not fractional.
3904 Constant *RHSInt = LHSUnsigned
3905 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3906 : ConstantExpr::getFPToSI(RHSC, IntTy);
3907 if (!RHS.isZero()) {
3908 bool Equal = LHSUnsigned
3909 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3910 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3912 // If we had a comparison against a fractional value, we have to adjust
3913 // the compare predicate and sometimes the value. RHSC is rounded towards
3914 // zero at this point.
3916 default: llvm_unreachable("Unexpected integer comparison!");
3917 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3918 return ReplaceInstUsesWith(I, Builder->getTrue());
3919 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3920 return ReplaceInstUsesWith(I, Builder->getFalse());
3921 case ICmpInst::ICMP_ULE:
3922 // (float)int <= 4.4 --> int <= 4
3923 // (float)int <= -4.4 --> false
3924 if (RHS.isNegative())
3925 return ReplaceInstUsesWith(I, Builder->getFalse());
3927 case ICmpInst::ICMP_SLE:
3928 // (float)int <= 4.4 --> int <= 4
3929 // (float)int <= -4.4 --> int < -4
3930 if (RHS.isNegative())
3931 Pred = ICmpInst::ICMP_SLT;
3933 case ICmpInst::ICMP_ULT:
3934 // (float)int < -4.4 --> false
3935 // (float)int < 4.4 --> int <= 4
3936 if (RHS.isNegative())
3937 return ReplaceInstUsesWith(I, Builder->getFalse());
3938 Pred = ICmpInst::ICMP_ULE;
3940 case ICmpInst::ICMP_SLT:
3941 // (float)int < -4.4 --> int < -4
3942 // (float)int < 4.4 --> int <= 4
3943 if (!RHS.isNegative())
3944 Pred = ICmpInst::ICMP_SLE;
3946 case ICmpInst::ICMP_UGT:
3947 // (float)int > 4.4 --> int > 4
3948 // (float)int > -4.4 --> true
3949 if (RHS.isNegative())
3950 return ReplaceInstUsesWith(I, Builder->getTrue());
3952 case ICmpInst::ICMP_SGT:
3953 // (float)int > 4.4 --> int > 4
3954 // (float)int > -4.4 --> int >= -4
3955 if (RHS.isNegative())
3956 Pred = ICmpInst::ICMP_SGE;
3958 case ICmpInst::ICMP_UGE:
3959 // (float)int >= -4.4 --> true
3960 // (float)int >= 4.4 --> int > 4
3961 if (RHS.isNegative())
3962 return ReplaceInstUsesWith(I, Builder->getTrue());
3963 Pred = ICmpInst::ICMP_UGT;
3965 case ICmpInst::ICMP_SGE:
3966 // (float)int >= -4.4 --> int >= -4
3967 // (float)int >= 4.4 --> int > 4
3968 if (!RHS.isNegative())
3969 Pred = ICmpInst::ICMP_SGT;
3975 // Lower this FP comparison into an appropriate integer version of the
3977 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3980 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3981 bool Changed = false;
3983 /// Orders the operands of the compare so that they are listed from most
3984 /// complex to least complex. This puts constants before unary operators,
3985 /// before binary operators.
3986 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3991 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3993 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1,
3994 I.getFastMathFlags(), DL, TLI, DT, AC, &I))
3995 return ReplaceInstUsesWith(I, V);
3997 // Simplify 'fcmp pred X, X'
3999 switch (I.getPredicate()) {
4000 default: llvm_unreachable("Unknown predicate!");
4001 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4002 case FCmpInst::FCMP_ULT: // True if unordered or less than
4003 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4004 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4005 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4006 I.setPredicate(FCmpInst::FCMP_UNO);
4007 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4010 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4011 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4012 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4013 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4014 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4015 I.setPredicate(FCmpInst::FCMP_ORD);
4016 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4021 // Test if the FCmpInst instruction is used exclusively by a select as
4022 // part of a minimum or maximum operation. If so, refrain from doing
4023 // any other folding. This helps out other analyses which understand
4024 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4025 // and CodeGen. And in this case, at least one of the comparison
4026 // operands has at least one user besides the compare (the select),
4027 // which would often largely negate the benefit of folding anyway.
4029 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
4030 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
4031 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
4034 // Handle fcmp with constant RHS
4035 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4036 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4037 switch (LHSI->getOpcode()) {
4038 case Instruction::FPExt: {
4039 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
4040 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
4041 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
4045 const fltSemantics *Sem;
4046 // FIXME: This shouldn't be here.
4047 if (LHSExt->getSrcTy()->isHalfTy())
4048 Sem = &APFloat::IEEEhalf;
4049 else if (LHSExt->getSrcTy()->isFloatTy())
4050 Sem = &APFloat::IEEEsingle;
4051 else if (LHSExt->getSrcTy()->isDoubleTy())
4052 Sem = &APFloat::IEEEdouble;
4053 else if (LHSExt->getSrcTy()->isFP128Ty())
4054 Sem = &APFloat::IEEEquad;
4055 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
4056 Sem = &APFloat::x87DoubleExtended;
4057 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
4058 Sem = &APFloat::PPCDoubleDouble;
4063 APFloat F = RHSF->getValueAPF();
4064 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
4066 // Avoid lossy conversions and denormals. Zero is a special case
4067 // that's OK to convert.
4071 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
4072 APFloat::cmpLessThan) || Fabs.isZero()))
4074 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
4075 ConstantFP::get(RHSC->getContext(), F));
4078 case Instruction::PHI:
4079 // Only fold fcmp into the PHI if the phi and fcmp are in the same
4080 // block. If in the same block, we're encouraging jump threading. If
4081 // not, we are just pessimizing the code by making an i1 phi.
4082 if (LHSI->getParent() == I.getParent())
4083 if (Instruction *NV = FoldOpIntoPhi(I))
4086 case Instruction::SIToFP:
4087 case Instruction::UIToFP:
4088 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
4091 case Instruction::FSub: {
4092 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
4094 if (match(LHSI, m_FNeg(m_Value(Op))))
4095 return new FCmpInst(I.getSwappedPredicate(), Op,
4096 ConstantExpr::getFNeg(RHSC));
4099 case Instruction::Load:
4100 if (GetElementPtrInst *GEP =
4101 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
4102 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
4103 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
4104 !cast<LoadInst>(LHSI)->isVolatile())
4105 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
4109 case Instruction::Call: {
4110 if (!RHSC->isNullValue())
4113 CallInst *CI = cast<CallInst>(LHSI);
4114 const Function *F = CI->getCalledFunction();
4118 // Various optimization for fabs compared with zero.
4120 if (F->getIntrinsicID() == Intrinsic::fabs ||
4121 (TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
4122 (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
4123 Func == LibFunc::fabsl))) {
4124 switch (I.getPredicate()) {
4127 // fabs(x) < 0 --> false
4128 case FCmpInst::FCMP_OLT:
4129 return ReplaceInstUsesWith(I, Builder->getFalse());
4130 // fabs(x) > 0 --> x != 0
4131 case FCmpInst::FCMP_OGT:
4132 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), RHSC);
4133 // fabs(x) <= 0 --> x == 0
4134 case FCmpInst::FCMP_OLE:
4135 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), RHSC);
4136 // fabs(x) >= 0 --> !isnan(x)
4137 case FCmpInst::FCMP_OGE:
4138 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), RHSC);
4139 // fabs(x) == 0 --> x == 0
4140 // fabs(x) != 0 --> x != 0
4141 case FCmpInst::FCMP_OEQ:
4142 case FCmpInst::FCMP_UEQ:
4143 case FCmpInst::FCMP_ONE:
4144 case FCmpInst::FCMP_UNE:
4145 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), RHSC);
4152 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
4154 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
4155 return new FCmpInst(I.getSwappedPredicate(), X, Y);
4157 // fcmp (fpext x), (fpext y) -> fcmp x, y
4158 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
4159 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
4160 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
4161 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
4162 RHSExt->getOperand(0));
4164 return Changed ? &I : nullptr;