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()));
1078 if (IsAShr && AP1.isNegative())
1079 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1081 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1084 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1085 // There are multiple solutions if we are comparing against -1 and the LHS
1086 // of the ashr is not a power of two.
1087 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1088 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1089 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1090 } else if (AP1 == AP2.lshr(Shift)) {
1091 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1094 // Shifting const2 will never be equal to const1.
1095 return getConstant(false);
1098 /// FoldICmpCstShlCst - Handle "(icmp eq/ne (shl const2, A), const1)" ->
1099 /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)).
1100 Instruction *InstCombiner::FoldICmpCstShlCst(ICmpInst &I, Value *Op, Value *A,
1103 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1105 auto getConstant = [&I, this](bool IsTrue) {
1106 if (I.getPredicate() == I.ICMP_NE)
1108 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1111 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1112 if (I.getPredicate() == I.ICMP_NE)
1113 Pred = CmpInst::getInversePredicate(Pred);
1114 return new ICmpInst(Pred, LHS, RHS);
1117 APInt AP1 = CI1->getValue();
1118 APInt AP2 = CI2->getValue();
1120 // Don't bother doing any work for cases which InstSimplify handles.
1124 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1126 if (!AP1 && AP2TrailingZeros != 0)
1127 return getICmp(I.ICMP_UGE, A,
1128 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1131 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1133 // Get the distance between the lowest bits that are set.
1134 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1136 if (Shift > 0 && AP2.shl(Shift) == AP1)
1137 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1139 // Shifting const2 will never be equal to const1.
1140 return getConstant(false);
1143 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1145 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1148 const APInt &RHSV = RHS->getValue();
1150 switch (LHSI->getOpcode()) {
1151 case Instruction::Trunc:
1152 if (RHS->isOne() && RHSV.getBitWidth() > 1) {
1153 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1155 if (ICI.getPredicate() == ICmpInst::ICMP_SLT &&
1156 match(LHSI->getOperand(0), m_Signum(m_Value(V))))
1157 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1158 ConstantInt::get(V->getType(), 1));
1160 if (ICI.isEquality() && LHSI->hasOneUse()) {
1161 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1162 // of the high bits truncated out of x are known.
1163 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1164 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1165 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1166 computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne, 0, &ICI);
1168 // If all the high bits are known, we can do this xform.
1169 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1170 // Pull in the high bits from known-ones set.
1171 APInt NewRHS = RHS->getValue().zext(SrcBits);
1172 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1173 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1174 Builder->getInt(NewRHS));
1179 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1180 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1181 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1183 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1184 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1185 Value *CompareVal = LHSI->getOperand(0);
1187 // If the sign bit of the XorCst is not set, there is no change to
1188 // the operation, just stop using the Xor.
1189 if (!XorCst->isNegative()) {
1190 ICI.setOperand(0, CompareVal);
1195 // Was the old condition true if the operand is positive?
1196 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1198 // If so, the new one isn't.
1199 isTrueIfPositive ^= true;
1201 if (isTrueIfPositive)
1202 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1205 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1209 if (LHSI->hasOneUse()) {
1210 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1211 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1212 const APInt &SignBit = XorCst->getValue();
1213 ICmpInst::Predicate Pred = ICI.isSigned()
1214 ? ICI.getUnsignedPredicate()
1215 : ICI.getSignedPredicate();
1216 return new ICmpInst(Pred, LHSI->getOperand(0),
1217 Builder->getInt(RHSV ^ SignBit));
1220 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1221 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1222 const APInt &NotSignBit = XorCst->getValue();
1223 ICmpInst::Predicate Pred = ICI.isSigned()
1224 ? ICI.getUnsignedPredicate()
1225 : ICI.getSignedPredicate();
1226 Pred = ICI.getSwappedPredicate(Pred);
1227 return new ICmpInst(Pred, LHSI->getOperand(0),
1228 Builder->getInt(RHSV ^ NotSignBit));
1232 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1233 // iff -C is a power of 2
1234 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1235 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1236 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1238 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1239 // iff -C is a power of 2
1240 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1241 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1242 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1245 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1246 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1247 LHSI->getOperand(0)->hasOneUse()) {
1248 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1250 // If the LHS is an AND of a truncating cast, we can widen the
1251 // and/compare to be the input width without changing the value
1252 // produced, eliminating a cast.
1253 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1254 // We can do this transformation if either the AND constant does not
1255 // have its sign bit set or if it is an equality comparison.
1256 // Extending a relational comparison when we're checking the sign
1257 // bit would not work.
1258 if (ICI.isEquality() ||
1259 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1261 Builder->CreateAnd(Cast->getOperand(0),
1262 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1263 NewAnd->takeName(LHSI);
1264 return new ICmpInst(ICI.getPredicate(), NewAnd,
1265 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1269 // If the LHS is an AND of a zext, and we have an equality compare, we can
1270 // shrink the and/compare to the smaller type, eliminating the cast.
1271 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1272 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1273 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1274 // should fold the icmp to true/false in that case.
1275 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1277 Builder->CreateAnd(Cast->getOperand(0),
1278 ConstantExpr::getTrunc(AndCst, Ty));
1279 NewAnd->takeName(LHSI);
1280 return new ICmpInst(ICI.getPredicate(), NewAnd,
1281 ConstantExpr::getTrunc(RHS, Ty));
1285 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1286 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1287 // happens a LOT in code produced by the C front-end, for bitfield
1289 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1290 if (Shift && !Shift->isShift())
1294 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
1296 // This seemingly simple opportunity to fold away a shift turns out to
1297 // be rather complicated. See PR17827
1298 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1300 bool CanFold = false;
1301 unsigned ShiftOpcode = Shift->getOpcode();
1302 if (ShiftOpcode == Instruction::AShr) {
1303 // There may be some constraints that make this possible,
1304 // but nothing simple has been discovered yet.
1306 } else if (ShiftOpcode == Instruction::Shl) {
1307 // For a left shift, we can fold if the comparison is not signed.
1308 // We can also fold a signed comparison if the mask value and
1309 // comparison value are not negative. These constraints may not be
1310 // obvious, but we can prove that they are correct using an SMT
1312 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1314 } else if (ShiftOpcode == Instruction::LShr) {
1315 // For a logical right shift, we can fold if the comparison is not
1316 // signed. We can also fold a signed comparison if the shifted mask
1317 // value and the shifted comparison value are not negative.
1318 // These constraints may not be obvious, but we can prove that they
1319 // are correct using an SMT solver.
1320 if (!ICI.isSigned())
1323 ConstantInt *ShiftedAndCst =
1324 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1325 ConstantInt *ShiftedRHSCst =
1326 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1328 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1335 if (ShiftOpcode == Instruction::Shl)
1336 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1338 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1340 // Check to see if we are shifting out any of the bits being
1342 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1343 // If we shifted bits out, the fold is not going to work out.
1344 // As a special case, check to see if this means that the
1345 // result is always true or false now.
1346 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1347 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1348 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1349 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1351 ICI.setOperand(1, NewCst);
1352 Constant *NewAndCst;
1353 if (ShiftOpcode == Instruction::Shl)
1354 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1356 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1357 LHSI->setOperand(1, NewAndCst);
1358 LHSI->setOperand(0, Shift->getOperand(0));
1359 Worklist.Add(Shift); // Shift is dead.
1365 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1366 // preferable because it allows the C<<Y expression to be hoisted out
1367 // of a loop if Y is invariant and X is not.
1368 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1369 ICI.isEquality() && !Shift->isArithmeticShift() &&
1370 !isa<Constant>(Shift->getOperand(0))) {
1373 if (Shift->getOpcode() == Instruction::LShr) {
1374 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1376 // Insert a logical shift.
1377 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1380 // Compute X & (C << Y).
1382 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1384 ICI.setOperand(0, NewAnd);
1388 // (icmp pred (and (or (lshr X, Y), X), 1), 0) -->
1389 // (icmp pred (and X, (or (shl 1, Y), 1), 0))
1391 // iff pred isn't signed
1393 Value *X, *Y, *LShr;
1394 if (!ICI.isSigned() && RHSV == 0) {
1395 if (match(LHSI->getOperand(1), m_One())) {
1396 Constant *One = cast<Constant>(LHSI->getOperand(1));
1397 Value *Or = LHSI->getOperand(0);
1398 if (match(Or, m_Or(m_Value(LShr), m_Value(X))) &&
1399 match(LShr, m_LShr(m_Specific(X), m_Value(Y)))) {
1400 unsigned UsesRemoved = 0;
1401 if (LHSI->hasOneUse())
1403 if (Or->hasOneUse())
1405 if (LShr->hasOneUse())
1407 Value *NewOr = nullptr;
1408 // Compute X & ((1 << Y) | 1)
1409 if (auto *C = dyn_cast<Constant>(Y)) {
1410 if (UsesRemoved >= 1)
1412 ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1414 if (UsesRemoved >= 3)
1415 NewOr = Builder->CreateOr(Builder->CreateShl(One, Y,
1418 One, Or->getName());
1421 Value *NewAnd = Builder->CreateAnd(X, NewOr, LHSI->getName());
1422 ICI.setOperand(0, NewAnd);
1430 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1431 // bit set in (X & AndCst) will produce a result greater than RHSV.
1432 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1433 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1434 if ((NTZ < AndCst->getBitWidth()) &&
1435 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1436 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1437 Constant::getNullValue(RHS->getType()));
1441 // Try to optimize things like "A[i]&42 == 0" to index computations.
1442 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1443 if (GetElementPtrInst *GEP =
1444 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1445 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1446 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1447 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1448 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1449 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1454 // X & -C == -C -> X > u ~C
1455 // X & -C != -C -> X <= u ~C
1456 // iff C is a power of 2
1457 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1458 return new ICmpInst(
1459 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1460 : ICmpInst::ICMP_ULE,
1461 LHSI->getOperand(0), SubOne(RHS));
1463 // (icmp eq (and %A, C), 0) -> (icmp sgt (trunc %A), -1)
1464 // iff C is a power of 2
1465 if (ICI.isEquality() && LHSI->hasOneUse() && match(RHS, m_Zero())) {
1466 if (auto *CI = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1467 const APInt &AI = CI->getValue();
1468 int32_t ExactLogBase2 = AI.exactLogBase2();
1469 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1470 Type *NTy = IntegerType::get(ICI.getContext(), ExactLogBase2 + 1);
1471 Value *Trunc = Builder->CreateTrunc(LHSI->getOperand(0), NTy);
1472 return new ICmpInst(ICI.getPredicate() == ICmpInst::ICMP_EQ
1473 ? ICmpInst::ICMP_SGE
1474 : ICmpInst::ICMP_SLT,
1475 Trunc, Constant::getNullValue(NTy));
1481 case Instruction::Or: {
1483 // icmp slt signum(V) 1 --> icmp slt V, 1
1485 if (ICI.getPredicate() == ICmpInst::ICMP_SLT &&
1486 match(LHSI, m_Signum(m_Value(V))))
1487 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1488 ConstantInt::get(V->getType(), 1));
1491 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1494 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1495 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1496 // -> and (icmp eq P, null), (icmp eq Q, null).
1497 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1498 Constant::getNullValue(P->getType()));
1499 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1500 Constant::getNullValue(Q->getType()));
1502 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1503 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1505 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1511 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1512 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1515 // If this is a signed comparison to 0 and the mul is sign preserving,
1516 // use the mul LHS operand instead.
1517 ICmpInst::Predicate pred = ICI.getPredicate();
1518 if (isSignTest(pred, RHS) && !Val->isZero() &&
1519 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1520 return new ICmpInst(Val->isNegative() ?
1521 ICmpInst::getSwappedPredicate(pred) : pred,
1522 LHSI->getOperand(0),
1523 Constant::getNullValue(RHS->getType()));
1528 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1529 uint32_t TypeBits = RHSV.getBitWidth();
1530 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1533 // (1 << X) pred P2 -> X pred Log2(P2)
1534 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1535 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1536 ICmpInst::Predicate Pred = ICI.getPredicate();
1537 if (ICI.isUnsigned()) {
1538 if (!RHSVIsPowerOf2) {
1539 // (1 << X) < 30 -> X <= 4
1540 // (1 << X) <= 30 -> X <= 4
1541 // (1 << X) >= 30 -> X > 4
1542 // (1 << X) > 30 -> X > 4
1543 if (Pred == ICmpInst::ICMP_ULT)
1544 Pred = ICmpInst::ICMP_ULE;
1545 else if (Pred == ICmpInst::ICMP_UGE)
1546 Pred = ICmpInst::ICMP_UGT;
1548 unsigned RHSLog2 = RHSV.logBase2();
1550 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1551 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1552 if (RHSLog2 == TypeBits-1) {
1553 if (Pred == ICmpInst::ICMP_UGE)
1554 Pred = ICmpInst::ICMP_EQ;
1555 else if (Pred == ICmpInst::ICMP_ULT)
1556 Pred = ICmpInst::ICMP_NE;
1559 return new ICmpInst(Pred, X,
1560 ConstantInt::get(RHS->getType(), RHSLog2));
1561 } else if (ICI.isSigned()) {
1562 if (RHSV.isAllOnesValue()) {
1563 // (1 << X) <= -1 -> X == 31
1564 if (Pred == ICmpInst::ICMP_SLE)
1565 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1566 ConstantInt::get(RHS->getType(), TypeBits-1));
1568 // (1 << X) > -1 -> X != 31
1569 if (Pred == ICmpInst::ICMP_SGT)
1570 return new ICmpInst(ICmpInst::ICMP_NE, X,
1571 ConstantInt::get(RHS->getType(), TypeBits-1));
1573 // (1 << X) < 0 -> X == 31
1574 // (1 << X) <= 0 -> X == 31
1575 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1576 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1577 ConstantInt::get(RHS->getType(), TypeBits-1));
1579 // (1 << X) >= 0 -> X != 31
1580 // (1 << X) > 0 -> X != 31
1581 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1582 return new ICmpInst(ICmpInst::ICMP_NE, X,
1583 ConstantInt::get(RHS->getType(), TypeBits-1));
1585 } else if (ICI.isEquality()) {
1587 return new ICmpInst(
1588 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1594 // Check that the shift amount is in range. If not, don't perform
1595 // undefined shifts. When the shift is visited it will be
1597 if (ShAmt->uge(TypeBits))
1600 if (ICI.isEquality()) {
1601 // If we are comparing against bits always shifted out, the
1602 // comparison cannot succeed.
1604 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1606 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1607 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1608 Constant *Cst = Builder->getInt1(IsICMP_NE);
1609 return ReplaceInstUsesWith(ICI, Cst);
1612 // If the shift is NUW, then it is just shifting out zeros, no need for an
1614 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1615 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1616 ConstantExpr::getLShr(RHS, ShAmt));
1618 // If the shift is NSW and we compare to 0, then it is just shifting out
1619 // sign bits, no need for an AND either.
1620 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1621 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1622 ConstantExpr::getLShr(RHS, ShAmt));
1624 if (LHSI->hasOneUse()) {
1625 // Otherwise strength reduce the shift into an and.
1626 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1627 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1628 TypeBits - ShAmtVal));
1631 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1632 return new ICmpInst(ICI.getPredicate(), And,
1633 ConstantExpr::getLShr(RHS, ShAmt));
1637 // If this is a signed comparison to 0 and the shift is sign preserving,
1638 // use the shift LHS operand instead.
1639 ICmpInst::Predicate pred = ICI.getPredicate();
1640 if (isSignTest(pred, RHS) &&
1641 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1642 return new ICmpInst(pred,
1643 LHSI->getOperand(0),
1644 Constant::getNullValue(RHS->getType()));
1646 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1647 bool TrueIfSigned = false;
1648 if (LHSI->hasOneUse() &&
1649 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1650 // (X << 31) <s 0 --> (X&1) != 0
1651 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1652 APInt::getOneBitSet(TypeBits,
1653 TypeBits-ShAmt->getZExtValue()-1));
1655 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1656 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1657 And, Constant::getNullValue(And->getType()));
1660 // Transform (icmp pred iM (shl iM %v, N), CI)
1661 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1662 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1663 // This enables to get rid of the shift in favor of a trunc which can be
1664 // free on the target. It has the additional benefit of comparing to a
1665 // smaller constant, which will be target friendly.
1666 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1667 if (LHSI->hasOneUse() &&
1668 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1669 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1670 Constant *NCI = ConstantExpr::getTrunc(
1671 ConstantExpr::getAShr(RHS,
1672 ConstantInt::get(RHS->getType(), Amt)),
1674 return new ICmpInst(ICI.getPredicate(),
1675 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1682 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1683 case Instruction::AShr: {
1684 // Handle equality comparisons of shift-by-constant.
1685 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1686 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1687 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1691 // Handle exact shr's.
1692 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1693 if (RHSV.isMinValue())
1694 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1699 case Instruction::SDiv:
1700 case Instruction::UDiv:
1701 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1702 // Fold this div into the comparison, producing a range check.
1703 // Determine, based on the divide type, what the range is being
1704 // checked. If there is an overflow on the low or high side, remember
1705 // it, otherwise compute the range [low, hi) bounding the new value.
1706 // See: InsertRangeTest above for the kinds of replacements possible.
1707 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1708 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1713 case Instruction::Sub: {
1714 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1716 const APInt &LHSV = LHSC->getValue();
1718 // C1-X <u C2 -> (X|(C2-1)) == C1
1719 // iff C1 & (C2-1) == C2-1
1720 // C2 is a power of 2
1721 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1722 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1723 return new ICmpInst(ICmpInst::ICMP_EQ,
1724 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1727 // C1-X >u C2 -> (X|C2) != C1
1728 // iff C1 & C2 == C2
1729 // C2+1 is a power of 2
1730 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1731 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1732 return new ICmpInst(ICmpInst::ICMP_NE,
1733 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1737 case Instruction::Add:
1738 // Fold: icmp pred (add X, C1), C2
1739 if (!ICI.isEquality()) {
1740 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1742 const APInt &LHSV = LHSC->getValue();
1744 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1747 if (ICI.isSigned()) {
1748 if (CR.getLower().isSignBit()) {
1749 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1750 Builder->getInt(CR.getUpper()));
1751 } else if (CR.getUpper().isSignBit()) {
1752 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1753 Builder->getInt(CR.getLower()));
1756 if (CR.getLower().isMinValue()) {
1757 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1758 Builder->getInt(CR.getUpper()));
1759 } else if (CR.getUpper().isMinValue()) {
1760 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1761 Builder->getInt(CR.getLower()));
1765 // X-C1 <u C2 -> (X & -C2) == C1
1766 // iff C1 & (C2-1) == 0
1767 // C2 is a power of 2
1768 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1769 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1770 return new ICmpInst(ICmpInst::ICMP_EQ,
1771 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1772 ConstantExpr::getNeg(LHSC));
1774 // X-C1 >u C2 -> (X & ~C2) != C1
1776 // C2+1 is a power of 2
1777 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1778 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1779 return new ICmpInst(ICmpInst::ICMP_NE,
1780 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1781 ConstantExpr::getNeg(LHSC));
1786 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1787 if (ICI.isEquality()) {
1788 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1790 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1791 // the second operand is a constant, simplify a bit.
1792 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1793 switch (BO->getOpcode()) {
1794 case Instruction::SRem:
1795 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1796 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1797 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1798 if (V.sgt(1) && V.isPowerOf2()) {
1800 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1802 return new ICmpInst(ICI.getPredicate(), NewRem,
1803 Constant::getNullValue(BO->getType()));
1807 case Instruction::Add:
1808 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1809 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1810 if (BO->hasOneUse())
1811 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1812 ConstantExpr::getSub(RHS, BOp1C));
1813 } else if (RHSV == 0) {
1814 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1815 // efficiently invertible, or if the add has just this one use.
1816 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1818 if (Value *NegVal = dyn_castNegVal(BOp1))
1819 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1820 if (Value *NegVal = dyn_castNegVal(BOp0))
1821 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1822 if (BO->hasOneUse()) {
1823 Value *Neg = Builder->CreateNeg(BOp1);
1825 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1829 case Instruction::Xor:
1830 // For the xor case, we can xor two constants together, eliminating
1831 // the explicit xor.
1832 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1833 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1834 ConstantExpr::getXor(RHS, BOC));
1835 } else if (RHSV == 0) {
1836 // Replace ((xor A, B) != 0) with (A != B)
1837 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1841 case Instruction::Sub:
1842 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1843 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1844 if (BO->hasOneUse())
1845 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1846 ConstantExpr::getSub(BOp0C, RHS));
1847 } else if (RHSV == 0) {
1848 // Replace ((sub A, B) != 0) with (A != B)
1849 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1853 case Instruction::Or:
1854 // If bits are being or'd in that are not present in the constant we
1855 // are comparing against, then the comparison could never succeed!
1856 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1857 Constant *NotCI = ConstantExpr::getNot(RHS);
1858 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1859 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1863 case Instruction::And:
1864 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1865 // If bits are being compared against that are and'd out, then the
1866 // comparison can never succeed!
1867 if ((RHSV & ~BOC->getValue()) != 0)
1868 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1870 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1871 if (RHS == BOC && RHSV.isPowerOf2())
1872 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1873 ICmpInst::ICMP_NE, LHSI,
1874 Constant::getNullValue(RHS->getType()));
1876 // Don't perform the following transforms if the AND has multiple uses
1877 if (!BO->hasOneUse())
1880 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1881 if (BOC->getValue().isSignBit()) {
1882 Value *X = BO->getOperand(0);
1883 Constant *Zero = Constant::getNullValue(X->getType());
1884 ICmpInst::Predicate pred = isICMP_NE ?
1885 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1886 return new ICmpInst(pred, X, Zero);
1889 // ((X & ~7) == 0) --> X < 8
1890 if (RHSV == 0 && isHighOnes(BOC)) {
1891 Value *X = BO->getOperand(0);
1892 Constant *NegX = ConstantExpr::getNeg(BOC);
1893 ICmpInst::Predicate pred = isICMP_NE ?
1894 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1895 return new ICmpInst(pred, X, NegX);
1899 case Instruction::Mul:
1900 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1901 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1902 // The trivial case (mul X, 0) is handled by InstSimplify
1903 // General case : (mul X, C) != 0 iff X != 0
1904 // (mul X, C) == 0 iff X == 0
1906 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1907 Constant::getNullValue(RHS->getType()));
1913 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1914 // Handle icmp {eq|ne} <intrinsic>, intcst.
1915 switch (II->getIntrinsicID()) {
1916 case Intrinsic::bswap:
1918 ICI.setOperand(0, II->getArgOperand(0));
1919 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1921 case Intrinsic::ctlz:
1922 case Intrinsic::cttz:
1923 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1924 if (RHSV == RHS->getType()->getBitWidth()) {
1926 ICI.setOperand(0, II->getArgOperand(0));
1927 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1931 case Intrinsic::ctpop:
1932 // popcount(A) == 0 -> A == 0 and likewise for !=
1933 if (RHS->isZero()) {
1935 ICI.setOperand(0, II->getArgOperand(0));
1936 ICI.setOperand(1, RHS);
1948 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1949 /// We only handle extending casts so far.
1951 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1952 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1953 Value *LHSCIOp = LHSCI->getOperand(0);
1954 Type *SrcTy = LHSCIOp->getType();
1955 Type *DestTy = LHSCI->getType();
1958 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1959 // integer type is the same size as the pointer type.
1960 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
1961 DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1962 Value *RHSOp = nullptr;
1963 if (PtrToIntOperator *RHSC = dyn_cast<PtrToIntOperator>(ICI.getOperand(1))) {
1964 Value *RHSCIOp = RHSC->getOperand(0);
1965 if (RHSCIOp->getType()->getPointerAddressSpace() ==
1966 LHSCIOp->getType()->getPointerAddressSpace()) {
1967 RHSOp = RHSC->getOperand(0);
1968 // If the pointer types don't match, insert a bitcast.
1969 if (LHSCIOp->getType() != RHSOp->getType())
1970 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1972 } else if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1)))
1973 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1976 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1979 // The code below only handles extension cast instructions, so far.
1981 if (LHSCI->getOpcode() != Instruction::ZExt &&
1982 LHSCI->getOpcode() != Instruction::SExt)
1985 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1986 bool isSignedCmp = ICI.isSigned();
1988 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1989 // Not an extension from the same type?
1990 RHSCIOp = CI->getOperand(0);
1991 if (RHSCIOp->getType() != LHSCIOp->getType())
1994 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1995 // and the other is a zext), then we can't handle this.
1996 if (CI->getOpcode() != LHSCI->getOpcode())
1999 // Deal with equality cases early.
2000 if (ICI.isEquality())
2001 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
2003 // A signed comparison of sign extended values simplifies into a
2004 // signed comparison.
2005 if (isSignedCmp && isSignedExt)
2006 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
2008 // The other three cases all fold into an unsigned comparison.
2009 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
2012 // If we aren't dealing with a constant on the RHS, exit early
2013 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
2017 // Compute the constant that would happen if we truncated to SrcTy then
2018 // reextended to DestTy.
2019 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
2020 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
2023 // If the re-extended constant didn't change...
2025 // Deal with equality cases early.
2026 if (ICI.isEquality())
2027 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
2029 // A signed comparison of sign extended values simplifies into a
2030 // signed comparison.
2031 if (isSignedExt && isSignedCmp)
2032 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
2034 // The other three cases all fold into an unsigned comparison.
2035 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
2038 // The re-extended constant changed so the constant cannot be represented
2039 // in the shorter type. Consequently, we cannot emit a simple comparison.
2040 // All the cases that fold to true or false will have already been handled
2041 // by SimplifyICmpInst, so only deal with the tricky case.
2043 if (isSignedCmp || !isSignedExt)
2046 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
2047 // should have been folded away previously and not enter in here.
2049 // We're performing an unsigned comp with a sign extended value.
2050 // This is true if the input is >= 0. [aka >s -1]
2051 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
2052 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
2054 // Finally, return the value computed.
2055 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
2056 return ReplaceInstUsesWith(ICI, Result);
2058 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
2059 return BinaryOperator::CreateNot(Result);
2062 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
2063 /// I = icmp ugt (add (add A, B), CI2), CI1
2064 /// If this is of the form:
2066 /// if (sum+128 >u 255)
2067 /// Then replace it with llvm.sadd.with.overflow.i8.
2069 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
2070 ConstantInt *CI2, ConstantInt *CI1,
2072 // The transformation we're trying to do here is to transform this into an
2073 // llvm.sadd.with.overflow. To do this, we have to replace the original add
2074 // with a narrower add, and discard the add-with-constant that is part of the
2075 // range check (if we can't eliminate it, this isn't profitable).
2077 // In order to eliminate the add-with-constant, the compare can be its only
2079 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
2080 if (!AddWithCst->hasOneUse()) return nullptr;
2082 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
2083 if (!CI2->getValue().isPowerOf2()) return nullptr;
2084 unsigned NewWidth = CI2->getValue().countTrailingZeros();
2085 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
2087 // The width of the new add formed is 1 more than the bias.
2090 // Check to see that CI1 is an all-ones value with NewWidth bits.
2091 if (CI1->getBitWidth() == NewWidth ||
2092 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
2095 // This is only really a signed overflow check if the inputs have been
2096 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
2097 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
2098 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
2099 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
2100 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
2103 // In order to replace the original add with a narrower
2104 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
2105 // and truncates that discard the high bits of the add. Verify that this is
2107 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
2108 for (User *U : OrigAdd->users()) {
2109 if (U == AddWithCst) continue;
2111 // Only accept truncates for now. We would really like a nice recursive
2112 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
2113 // chain to see which bits of a value are actually demanded. If the
2114 // original add had another add which was then immediately truncated, we
2115 // could still do the transformation.
2116 TruncInst *TI = dyn_cast<TruncInst>(U);
2117 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
2121 // If the pattern matches, truncate the inputs to the narrower type and
2122 // use the sadd_with_overflow intrinsic to efficiently compute both the
2123 // result and the overflow bit.
2124 Module *M = I.getParent()->getParent()->getParent();
2126 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
2127 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
2130 InstCombiner::BuilderTy *Builder = IC.Builder;
2132 // Put the new code above the original add, in case there are any uses of the
2133 // add between the add and the compare.
2134 Builder->SetInsertPoint(OrigAdd);
2136 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
2137 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
2138 CallInst *Call = Builder->CreateCall(F, {TruncA, TruncB}, "sadd");
2139 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
2140 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
2142 // The inner add was the result of the narrow add, zero extended to the
2143 // wider type. Replace it with the result computed by the intrinsic.
2144 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
2146 // The original icmp gets replaced with the overflow value.
2147 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
2150 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
2151 Value *RHS, Instruction &OrigI,
2152 Value *&Result, Constant *&Overflow) {
2153 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
2154 std::swap(LHS, RHS);
2156 auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
2158 Overflow = OverflowVal;
2160 Result->takeName(&OrigI);
2164 // If the overflow check was an add followed by a compare, the insertion point
2165 // may be pointing to the compare. We want to insert the new instructions
2166 // before the add in case there are uses of the add between the add and the
2168 Builder->SetInsertPoint(&OrigI);
2172 llvm_unreachable("bad overflow check kind!");
2174 case OCF_UNSIGNED_ADD: {
2175 OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
2176 if (OR == OverflowResult::NeverOverflows)
2177 return SetResult(Builder->CreateNUWAdd(LHS, RHS), Builder->getFalse(),
2180 if (OR == OverflowResult::AlwaysOverflows)
2181 return SetResult(Builder->CreateAdd(LHS, RHS), Builder->getTrue(), true);
2183 // FALL THROUGH uadd into sadd
2184 case OCF_SIGNED_ADD: {
2185 // X + 0 -> {X, false}
2186 if (match(RHS, m_Zero()))
2187 return SetResult(LHS, Builder->getFalse(), false);
2189 // We can strength reduce this signed add into a regular add if we can prove
2190 // that it will never overflow.
2191 if (OCF == OCF_SIGNED_ADD)
2192 if (WillNotOverflowSignedAdd(LHS, RHS, OrigI))
2193 return SetResult(Builder->CreateNSWAdd(LHS, RHS), Builder->getFalse(),
2198 case OCF_UNSIGNED_SUB:
2199 case OCF_SIGNED_SUB: {
2200 // X - 0 -> {X, false}
2201 if (match(RHS, m_Zero()))
2202 return SetResult(LHS, Builder->getFalse(), false);
2204 if (OCF == OCF_SIGNED_SUB) {
2205 if (WillNotOverflowSignedSub(LHS, RHS, OrigI))
2206 return SetResult(Builder->CreateNSWSub(LHS, RHS), Builder->getFalse(),
2209 if (WillNotOverflowUnsignedSub(LHS, RHS, OrigI))
2210 return SetResult(Builder->CreateNUWSub(LHS, RHS), Builder->getFalse(),
2216 case OCF_UNSIGNED_MUL: {
2217 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
2218 if (OR == OverflowResult::NeverOverflows)
2219 return SetResult(Builder->CreateNUWMul(LHS, RHS), Builder->getFalse(),
2221 if (OR == OverflowResult::AlwaysOverflows)
2222 return SetResult(Builder->CreateMul(LHS, RHS), Builder->getTrue(), true);
2224 case OCF_SIGNED_MUL:
2225 // X * undef -> undef
2226 if (isa<UndefValue>(RHS))
2227 return SetResult(RHS, UndefValue::get(Builder->getInt1Ty()), false);
2229 // X * 0 -> {0, false}
2230 if (match(RHS, m_Zero()))
2231 return SetResult(RHS, Builder->getFalse(), false);
2233 // X * 1 -> {X, false}
2234 if (match(RHS, m_One()))
2235 return SetResult(LHS, Builder->getFalse(), false);
2237 if (OCF == OCF_SIGNED_MUL)
2238 if (WillNotOverflowSignedMul(LHS, RHS, OrigI))
2239 return SetResult(Builder->CreateNSWMul(LHS, RHS), Builder->getFalse(),
2247 /// \brief Recognize and process idiom involving test for multiplication
2250 /// The caller has matched a pattern of the form:
2251 /// I = cmp u (mul(zext A, zext B), V
2252 /// The function checks if this is a test for overflow and if so replaces
2253 /// multiplication with call to 'mul.with.overflow' intrinsic.
2255 /// \param I Compare instruction.
2256 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
2257 /// the compare instruction. Must be of integer type.
2258 /// \param OtherVal The other argument of compare instruction.
2259 /// \returns Instruction which must replace the compare instruction, NULL if no
2260 /// replacement required.
2261 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
2262 Value *OtherVal, InstCombiner &IC) {
2263 // Don't bother doing this transformation for pointers, don't do it for
2265 if (!isa<IntegerType>(MulVal->getType()))
2268 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
2269 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
2270 auto *MulInstr = dyn_cast<Instruction>(MulVal);
2273 assert(MulInstr->getOpcode() == Instruction::Mul);
2275 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
2276 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
2277 assert(LHS->getOpcode() == Instruction::ZExt);
2278 assert(RHS->getOpcode() == Instruction::ZExt);
2279 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
2281 // Calculate type and width of the result produced by mul.with.overflow.
2282 Type *TyA = A->getType(), *TyB = B->getType();
2283 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
2284 WidthB = TyB->getPrimitiveSizeInBits();
2287 if (WidthB > WidthA) {
2295 // In order to replace the original mul with a narrower mul.with.overflow,
2296 // all uses must ignore upper bits of the product. The number of used low
2297 // bits must be not greater than the width of mul.with.overflow.
2298 if (MulVal->hasNUsesOrMore(2))
2299 for (User *U : MulVal->users()) {
2302 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2303 // Check if truncation ignores bits above MulWidth.
2304 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
2305 if (TruncWidth > MulWidth)
2307 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2308 // Check if AND ignores bits above MulWidth.
2309 if (BO->getOpcode() != Instruction::And)
2311 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
2312 const APInt &CVal = CI->getValue();
2313 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
2317 // Other uses prohibit this transformation.
2322 // Recognize patterns
2323 switch (I.getPredicate()) {
2324 case ICmpInst::ICMP_EQ:
2325 case ICmpInst::ICMP_NE:
2326 // Recognize pattern:
2327 // mulval = mul(zext A, zext B)
2328 // cmp eq/neq mulval, zext trunc mulval
2329 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
2330 if (Zext->hasOneUse()) {
2331 Value *ZextArg = Zext->getOperand(0);
2332 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
2333 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
2337 // Recognize pattern:
2338 // mulval = mul(zext A, zext B)
2339 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
2342 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
2343 if (ValToMask != MulVal)
2345 const APInt &CVal = CI->getValue() + 1;
2346 if (CVal.isPowerOf2()) {
2347 unsigned MaskWidth = CVal.logBase2();
2348 if (MaskWidth == MulWidth)
2349 break; // Recognized
2354 case ICmpInst::ICMP_UGT:
2355 // Recognize pattern:
2356 // mulval = mul(zext A, zext B)
2357 // cmp ugt mulval, max
2358 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2359 APInt MaxVal = APInt::getMaxValue(MulWidth);
2360 MaxVal = MaxVal.zext(CI->getBitWidth());
2361 if (MaxVal.eq(CI->getValue()))
2362 break; // Recognized
2366 case ICmpInst::ICMP_UGE:
2367 // Recognize pattern:
2368 // mulval = mul(zext A, zext B)
2369 // cmp uge mulval, max+1
2370 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2371 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2372 if (MaxVal.eq(CI->getValue()))
2373 break; // Recognized
2377 case ICmpInst::ICMP_ULE:
2378 // Recognize pattern:
2379 // mulval = mul(zext A, zext B)
2380 // cmp ule mulval, max
2381 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2382 APInt MaxVal = APInt::getMaxValue(MulWidth);
2383 MaxVal = MaxVal.zext(CI->getBitWidth());
2384 if (MaxVal.eq(CI->getValue()))
2385 break; // Recognized
2389 case ICmpInst::ICMP_ULT:
2390 // Recognize pattern:
2391 // mulval = mul(zext A, zext B)
2392 // cmp ule mulval, max + 1
2393 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2394 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2395 if (MaxVal.eq(CI->getValue()))
2396 break; // Recognized
2404 InstCombiner::BuilderTy *Builder = IC.Builder;
2405 Builder->SetInsertPoint(MulInstr);
2406 Module *M = I.getParent()->getParent()->getParent();
2408 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
2409 Value *MulA = A, *MulB = B;
2410 if (WidthA < MulWidth)
2411 MulA = Builder->CreateZExt(A, MulType);
2412 if (WidthB < MulWidth)
2413 MulB = Builder->CreateZExt(B, MulType);
2415 Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
2416 CallInst *Call = Builder->CreateCall(F, {MulA, MulB}, "umul");
2417 IC.Worklist.Add(MulInstr);
2419 // If there are uses of mul result other than the comparison, we know that
2420 // they are truncation or binary AND. Change them to use result of
2421 // mul.with.overflow and adjust properly mask/size.
2422 if (MulVal->hasNUsesOrMore(2)) {
2423 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
2424 for (User *U : MulVal->users()) {
2425 if (U == &I || U == OtherVal)
2427 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2428 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
2429 IC.ReplaceInstUsesWith(*TI, Mul);
2431 TI->setOperand(0, Mul);
2432 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2433 assert(BO->getOpcode() == Instruction::And);
2434 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
2435 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
2436 APInt ShortMask = CI->getValue().trunc(MulWidth);
2437 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
2439 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
2440 IC.Worklist.Add(Zext);
2441 IC.ReplaceInstUsesWith(*BO, Zext);
2443 llvm_unreachable("Unexpected Binary operation");
2445 IC.Worklist.Add(cast<Instruction>(U));
2448 if (isa<Instruction>(OtherVal))
2449 IC.Worklist.Add(cast<Instruction>(OtherVal));
2451 // The original icmp gets replaced with the overflow value, maybe inverted
2452 // depending on predicate.
2453 bool Inverse = false;
2454 switch (I.getPredicate()) {
2455 case ICmpInst::ICMP_NE:
2457 case ICmpInst::ICMP_EQ:
2460 case ICmpInst::ICMP_UGT:
2461 case ICmpInst::ICMP_UGE:
2462 if (I.getOperand(0) == MulVal)
2466 case ICmpInst::ICMP_ULT:
2467 case ICmpInst::ICMP_ULE:
2468 if (I.getOperand(1) == MulVal)
2473 llvm_unreachable("Unexpected predicate");
2476 Value *Res = Builder->CreateExtractValue(Call, 1);
2477 return BinaryOperator::CreateNot(Res);
2480 return ExtractValueInst::Create(Call, 1);
2483 // DemandedBitsLHSMask - When performing a comparison against a constant,
2484 // it is possible that not all the bits in the LHS are demanded. This helper
2485 // method computes the mask that IS demanded.
2486 static APInt DemandedBitsLHSMask(ICmpInst &I,
2487 unsigned BitWidth, bool isSignCheck) {
2489 return APInt::getSignBit(BitWidth);
2491 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2492 if (!CI) return APInt::getAllOnesValue(BitWidth);
2493 const APInt &RHS = CI->getValue();
2495 switch (I.getPredicate()) {
2496 // For a UGT comparison, we don't care about any bits that
2497 // correspond to the trailing ones of the comparand. The value of these
2498 // bits doesn't impact the outcome of the comparison, because any value
2499 // greater than the RHS must differ in a bit higher than these due to carry.
2500 case ICmpInst::ICMP_UGT: {
2501 unsigned trailingOnes = RHS.countTrailingOnes();
2502 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2506 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2507 // Any value less than the RHS must differ in a higher bit because of carries.
2508 case ICmpInst::ICMP_ULT: {
2509 unsigned trailingZeros = RHS.countTrailingZeros();
2510 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2515 return APInt::getAllOnesValue(BitWidth);
2520 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2521 /// should be swapped.
2522 /// The decision is based on how many times these two operands are reused
2523 /// as subtract operands and their positions in those instructions.
2524 /// The rational is that several architectures use the same instruction for
2525 /// both subtract and cmp, thus it is better if the order of those operands
2527 /// \return true if Op0 and Op1 should be swapped.
2528 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2529 const Value * Op1) {
2530 // Filter out pointer value as those cannot appears directly in subtract.
2531 // FIXME: we may want to go through inttoptrs or bitcasts.
2532 if (Op0->getType()->isPointerTy())
2534 // Count every uses of both Op0 and Op1 in a subtract.
2535 // Each time Op0 is the first operand, count -1: swapping is bad, the
2536 // subtract has already the same layout as the compare.
2537 // Each time Op0 is the second operand, count +1: swapping is good, the
2538 // subtract has a different layout as the compare.
2539 // At the end, if the benefit is greater than 0, Op0 should come second to
2540 // expose more CSE opportunities.
2541 int GlobalSwapBenefits = 0;
2542 for (const User *U : Op0->users()) {
2543 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
2544 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2546 // If Op0 is the first argument, this is not beneficial to swap the
2548 int LocalSwapBenefits = -1;
2549 unsigned Op1Idx = 1;
2550 if (BinOp->getOperand(Op1Idx) == Op0) {
2552 LocalSwapBenefits = 1;
2554 if (BinOp->getOperand(Op1Idx) != Op1)
2556 GlobalSwapBenefits += LocalSwapBenefits;
2558 return GlobalSwapBenefits > 0;
2561 /// \brief Check that one use is in the same block as the definition and all
2562 /// other uses are in blocks dominated by a given block
2564 /// \param DI Definition
2566 /// \param DB Block that must dominate all uses of \p DI outside
2567 /// the parent block
2568 /// \return true when \p UI is the only use of \p DI in the parent block
2569 /// and all other uses of \p DI are in blocks dominated by \p DB.
2571 bool InstCombiner::dominatesAllUses(const Instruction *DI,
2572 const Instruction *UI,
2573 const BasicBlock *DB) const {
2574 assert(DI && UI && "Instruction not defined\n");
2575 // ignore incomplete definitions
2576 if (!DI->getParent())
2578 // DI and UI must be in the same block
2579 if (DI->getParent() != UI->getParent())
2581 // Protect from self-referencing blocks
2582 if (DI->getParent() == DB)
2584 // DominatorTree available?
2587 for (const User *U : DI->users()) {
2588 auto *Usr = cast<Instruction>(U);
2589 if (Usr != UI && !DT->dominates(DB, Usr->getParent()))
2596 /// true when the instruction sequence within a block is select-cmp-br.
2598 static bool isChainSelectCmpBranch(const SelectInst *SI) {
2599 const BasicBlock *BB = SI->getParent();
2602 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
2603 if (!BI || BI->getNumSuccessors() != 2)
2605 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
2606 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
2612 /// \brief True when a select result is replaced by one of its operands
2613 /// in select-icmp sequence. This will eventually result in the elimination
2616 /// \param SI Select instruction
2617 /// \param Icmp Compare instruction
2618 /// \param SIOpd Operand that replaces the select
2621 /// - The replacement is global and requires dominator information
2622 /// - The caller is responsible for the actual replacement
2627 /// %4 = select i1 %3, %C* %0, %C* null
2628 /// %5 = icmp eq %C* %4, null
2629 /// br i1 %5, label %9, label %7
2631 /// ; <label>:7 ; preds = %entry
2632 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
2635 /// can be transformed to
2637 /// %5 = icmp eq %C* %0, null
2638 /// %6 = select i1 %3, i1 %5, i1 true
2639 /// br i1 %6, label %9, label %7
2641 /// ; <label>:7 ; preds = %entry
2642 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
2644 /// Similar when the first operand of the select is a constant or/and
2645 /// the compare is for not equal rather than equal.
2647 /// NOTE: The function is only called when the select and compare constants
2648 /// are equal, the optimization can work only for EQ predicates. This is not a
2649 /// major restriction since a NE compare should be 'normalized' to an equal
2650 /// compare, which usually happens in the combiner and test case
2651 /// select-cmp-br.ll
2653 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
2654 const ICmpInst *Icmp,
2655 const unsigned SIOpd) {
2656 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
2657 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
2658 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
2659 // The check for the unique predecessor is not the best that can be
2660 // done. But it protects efficiently against cases like when SI's
2661 // home block has two successors, Succ and Succ1, and Succ1 predecessor
2662 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
2663 // replaced can be reached on either path. So the uniqueness check
2664 // guarantees that the path all uses of SI (outside SI's parent) are on
2665 // is disjoint from all other paths out of SI. But that information
2666 // is more expensive to compute, and the trade-off here is in favor
2668 if (Succ->getUniquePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
2670 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
2677 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2678 bool Changed = false;
2679 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2680 unsigned Op0Cplxity = getComplexity(Op0);
2681 unsigned Op1Cplxity = getComplexity(Op1);
2683 /// Orders the operands of the compare so that they are listed from most
2684 /// complex to least complex. This puts constants before unary operators,
2685 /// before binary operators.
2686 if (Op0Cplxity < Op1Cplxity ||
2687 (Op0Cplxity == Op1Cplxity &&
2688 swapMayExposeCSEOpportunities(Op0, Op1))) {
2690 std::swap(Op0, Op1);
2695 SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AC, &I))
2696 return ReplaceInstUsesWith(I, V);
2698 // comparing -val or val with non-zero is the same as just comparing val
2699 // ie, abs(val) != 0 -> val != 0
2700 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2702 Value *Cond, *SelectTrue, *SelectFalse;
2703 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2704 m_Value(SelectFalse)))) {
2705 if (Value *V = dyn_castNegVal(SelectTrue)) {
2706 if (V == SelectFalse)
2707 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2709 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2710 if (V == SelectTrue)
2711 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2716 Type *Ty = Op0->getType();
2718 // icmp's with boolean values can always be turned into bitwise operations
2719 if (Ty->isIntegerTy(1)) {
2720 switch (I.getPredicate()) {
2721 default: llvm_unreachable("Invalid icmp instruction!");
2722 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2723 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2724 return BinaryOperator::CreateNot(Xor);
2726 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2727 return BinaryOperator::CreateXor(Op0, Op1);
2729 case ICmpInst::ICMP_UGT:
2730 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2732 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2733 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2734 return BinaryOperator::CreateAnd(Not, Op1);
2736 case ICmpInst::ICMP_SGT:
2737 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2739 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2740 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2741 return BinaryOperator::CreateAnd(Not, Op0);
2743 case ICmpInst::ICMP_UGE:
2744 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2746 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2747 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2748 return BinaryOperator::CreateOr(Not, Op1);
2750 case ICmpInst::ICMP_SGE:
2751 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2753 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2754 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2755 return BinaryOperator::CreateOr(Not, Op0);
2760 unsigned BitWidth = 0;
2761 if (Ty->isIntOrIntVectorTy())
2762 BitWidth = Ty->getScalarSizeInBits();
2763 else // Get pointer size.
2764 BitWidth = DL.getTypeSizeInBits(Ty->getScalarType());
2766 bool isSignBit = false;
2768 // See if we are doing a comparison with a constant.
2769 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2770 Value *A = nullptr, *B = nullptr;
2772 // Match the following pattern, which is a common idiom when writing
2773 // overflow-safe integer arithmetic function. The source performs an
2774 // addition in wider type, and explicitly checks for overflow using
2775 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2776 // sadd_with_overflow intrinsic.
2778 // TODO: This could probably be generalized to handle other overflow-safe
2779 // operations if we worked out the formulas to compute the appropriate
2783 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2785 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2786 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2787 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2788 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2792 // The following transforms are only 'worth it' if the only user of the
2793 // subtraction is the icmp.
2794 if (Op0->hasOneUse()) {
2795 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2796 if (I.isEquality() && CI->isZero() &&
2797 match(Op0, m_Sub(m_Value(A), m_Value(B))))
2798 return new ICmpInst(I.getPredicate(), A, B);
2800 // (icmp sgt (sub nsw A B), -1) -> (icmp sge A, B)
2801 if (I.getPredicate() == ICmpInst::ICMP_SGT && CI->isAllOnesValue() &&
2802 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2803 return new ICmpInst(ICmpInst::ICMP_SGE, A, B);
2805 // (icmp sgt (sub nsw A B), 0) -> (icmp sgt A, B)
2806 if (I.getPredicate() == ICmpInst::ICMP_SGT && CI->isZero() &&
2807 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2808 return new ICmpInst(ICmpInst::ICMP_SGT, A, B);
2810 // (icmp slt (sub nsw A B), 0) -> (icmp slt A, B)
2811 if (I.getPredicate() == ICmpInst::ICMP_SLT && CI->isZero() &&
2812 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2813 return new ICmpInst(ICmpInst::ICMP_SLT, A, B);
2815 // (icmp slt (sub nsw A B), 1) -> (icmp sle A, B)
2816 if (I.getPredicate() == ICmpInst::ICMP_SLT && CI->isOne() &&
2817 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2818 return new ICmpInst(ICmpInst::ICMP_SLE, A, B);
2821 // If we have an icmp le or icmp ge instruction, turn it into the
2822 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2823 // them being folded in the code below. The SimplifyICmpInst code has
2824 // already handled the edge cases for us, so we just assert on them.
2825 switch (I.getPredicate()) {
2827 case ICmpInst::ICMP_ULE:
2828 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2829 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2830 Builder->getInt(CI->getValue()+1));
2831 case ICmpInst::ICMP_SLE:
2832 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2833 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2834 Builder->getInt(CI->getValue()+1));
2835 case ICmpInst::ICMP_UGE:
2836 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2837 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2838 Builder->getInt(CI->getValue()-1));
2839 case ICmpInst::ICMP_SGE:
2840 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2841 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2842 Builder->getInt(CI->getValue()-1));
2845 if (I.isEquality()) {
2847 if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
2848 match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
2849 // (icmp eq/ne (ashr/lshr const2, A), const1)
2850 if (Instruction *Inst = FoldICmpCstShrCst(I, Op0, A, CI, CI2))
2853 if (match(Op0, m_Shl(m_ConstantInt(CI2), m_Value(A)))) {
2854 // (icmp eq/ne (shl const2, A), const1)
2855 if (Instruction *Inst = FoldICmpCstShlCst(I, Op0, A, CI, CI2))
2860 // If this comparison is a normal comparison, it demands all
2861 // bits, if it is a sign bit comparison, it only demands the sign bit.
2863 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2866 // See if we can fold the comparison based on range information we can get
2867 // by checking whether bits are known to be zero or one in the input.
2868 if (BitWidth != 0) {
2869 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2870 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2872 if (SimplifyDemandedBits(I.getOperandUse(0),
2873 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2874 Op0KnownZero, Op0KnownOne, 0))
2876 if (SimplifyDemandedBits(I.getOperandUse(1),
2877 APInt::getAllOnesValue(BitWidth), Op1KnownZero,
2881 // Given the known and unknown bits, compute a range that the LHS could be
2882 // in. Compute the Min, Max and RHS values based on the known bits. For the
2883 // EQ and NE we use unsigned values.
2884 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2885 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2887 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2889 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2892 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2894 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2898 // If Min and Max are known to be the same, then SimplifyDemandedBits
2899 // figured out that the LHS is a constant. Just constant fold this now so
2900 // that code below can assume that Min != Max.
2901 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2902 return new ICmpInst(I.getPredicate(),
2903 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2904 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2905 return new ICmpInst(I.getPredicate(), Op0,
2906 ConstantInt::get(Op1->getType(), Op1Min));
2908 // Based on the range information we know about the LHS, see if we can
2909 // simplify this comparison. For example, (x&4) < 8 is always true.
2910 switch (I.getPredicate()) {
2911 default: llvm_unreachable("Unknown icmp opcode!");
2912 case ICmpInst::ICMP_EQ: {
2913 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2914 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2916 // If all bits are known zero except for one, then we know at most one
2917 // bit is set. If the comparison is against zero, then this is a check
2918 // to see if *that* bit is set.
2919 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2920 if (~Op1KnownZero == 0) {
2921 // If the LHS is an AND with the same constant, look through it.
2922 Value *LHS = nullptr;
2923 ConstantInt *LHSC = nullptr;
2924 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2925 LHSC->getValue() != Op0KnownZeroInverted)
2928 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2929 // then turn "((1 << x)&8) == 0" into "x != 3".
2930 // or turn "((1 << x)&7) == 0" into "x > 2".
2932 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2933 APInt ValToCheck = Op0KnownZeroInverted;
2934 if (ValToCheck.isPowerOf2()) {
2935 unsigned CmpVal = ValToCheck.countTrailingZeros();
2936 return new ICmpInst(ICmpInst::ICMP_NE, X,
2937 ConstantInt::get(X->getType(), CmpVal));
2938 } else if ((++ValToCheck).isPowerOf2()) {
2939 unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
2940 return new ICmpInst(ICmpInst::ICMP_UGT, X,
2941 ConstantInt::get(X->getType(), CmpVal));
2945 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2946 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2948 if (Op0KnownZeroInverted == 1 &&
2949 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2950 return new ICmpInst(ICmpInst::ICMP_NE, X,
2951 ConstantInt::get(X->getType(),
2952 CI->countTrailingZeros()));
2957 case ICmpInst::ICMP_NE: {
2958 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2959 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2961 // If all bits are known zero except for one, then we know at most one
2962 // bit is set. If the comparison is against zero, then this is a check
2963 // to see if *that* bit is set.
2964 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2965 if (~Op1KnownZero == 0) {
2966 // If the LHS is an AND with the same constant, look through it.
2967 Value *LHS = nullptr;
2968 ConstantInt *LHSC = nullptr;
2969 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2970 LHSC->getValue() != Op0KnownZeroInverted)
2973 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2974 // then turn "((1 << x)&8) != 0" into "x == 3".
2975 // or turn "((1 << x)&7) != 0" into "x < 3".
2977 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2978 APInt ValToCheck = Op0KnownZeroInverted;
2979 if (ValToCheck.isPowerOf2()) {
2980 unsigned CmpVal = ValToCheck.countTrailingZeros();
2981 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2982 ConstantInt::get(X->getType(), CmpVal));
2983 } else if ((++ValToCheck).isPowerOf2()) {
2984 unsigned CmpVal = ValToCheck.countTrailingZeros();
2985 return new ICmpInst(ICmpInst::ICMP_ULT, X,
2986 ConstantInt::get(X->getType(), CmpVal));
2990 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2991 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2993 if (Op0KnownZeroInverted == 1 &&
2994 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2995 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2996 ConstantInt::get(X->getType(),
2997 CI->countTrailingZeros()));
3002 case ICmpInst::ICMP_ULT:
3003 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
3004 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3005 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
3006 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3007 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
3008 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
3009 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3010 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
3011 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
3012 Builder->getInt(CI->getValue()-1));
3014 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
3015 if (CI->isMinValue(true))
3016 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
3017 Constant::getAllOnesValue(Op0->getType()));
3020 case ICmpInst::ICMP_UGT:
3021 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
3022 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3023 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
3024 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3026 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
3027 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
3028 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3029 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
3030 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
3031 Builder->getInt(CI->getValue()+1));
3033 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
3034 if (CI->isMaxValue(true))
3035 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
3036 Constant::getNullValue(Op0->getType()));
3039 case ICmpInst::ICMP_SLT:
3040 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
3041 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3042 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
3043 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3044 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
3045 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
3046 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3047 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
3048 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
3049 Builder->getInt(CI->getValue()-1));
3052 case ICmpInst::ICMP_SGT:
3053 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
3054 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3055 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
3056 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3058 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
3059 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
3060 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3061 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
3062 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
3063 Builder->getInt(CI->getValue()+1));
3066 case ICmpInst::ICMP_SGE:
3067 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
3068 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
3069 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3070 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
3071 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3073 case ICmpInst::ICMP_SLE:
3074 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
3075 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
3076 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3077 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
3078 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3080 case ICmpInst::ICMP_UGE:
3081 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
3082 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
3083 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3084 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
3085 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3087 case ICmpInst::ICMP_ULE:
3088 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
3089 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
3090 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3091 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
3092 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3096 // Turn a signed comparison into an unsigned one if both operands
3097 // are known to have the same sign.
3099 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
3100 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
3101 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
3104 // Test if the ICmpInst instruction is used exclusively by a select as
3105 // part of a minimum or maximum operation. If so, refrain from doing
3106 // any other folding. This helps out other analyses which understand
3107 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
3108 // and CodeGen. And in this case, at least one of the comparison
3109 // operands has at least one user besides the compare (the select),
3110 // which would often largely negate the benefit of folding anyway.
3112 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
3113 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
3114 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
3117 // See if we are doing a comparison between a constant and an instruction that
3118 // can be folded into the comparison.
3119 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3120 // Since the RHS is a ConstantInt (CI), if the left hand side is an
3121 // instruction, see if that instruction also has constants so that the
3122 // instruction can be folded into the icmp
3123 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3124 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
3128 // Handle icmp with constant (but not simple integer constant) RHS
3129 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3130 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3131 switch (LHSI->getOpcode()) {
3132 case Instruction::GetElementPtr:
3133 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3134 if (RHSC->isNullValue() &&
3135 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3136 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
3137 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3139 case Instruction::PHI:
3140 // Only fold icmp into the PHI if the phi and icmp are in the same
3141 // block. If in the same block, we're encouraging jump threading. If
3142 // not, we are just pessimizing the code by making an i1 phi.
3143 if (LHSI->getParent() == I.getParent())
3144 if (Instruction *NV = FoldOpIntoPhi(I))
3147 case Instruction::Select: {
3148 // If either operand of the select is a constant, we can fold the
3149 // comparison into the select arms, which will cause one to be
3150 // constant folded and the select turned into a bitwise or.
3151 Value *Op1 = nullptr, *Op2 = nullptr;
3152 ConstantInt *CI = 0;
3153 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3154 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3155 CI = dyn_cast<ConstantInt>(Op1);
3157 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3158 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3159 CI = dyn_cast<ConstantInt>(Op2);
3162 // We only want to perform this transformation if it will not lead to
3163 // additional code. This is true if either both sides of the select
3164 // fold to a constant (in which case the icmp is replaced with a select
3165 // which will usually simplify) or this is the only user of the
3166 // select (in which case we are trading a select+icmp for a simpler
3167 // select+icmp) or all uses of the select can be replaced based on
3168 // dominance information ("Global cases").
3169 bool Transform = false;
3172 else if (Op1 || Op2) {
3174 if (LHSI->hasOneUse())
3177 else if (CI && !CI->isZero())
3178 // When Op1 is constant try replacing select with second operand.
3179 // Otherwise Op2 is constant and try replacing select with first
3181 Transform = replacedSelectWithOperand(cast<SelectInst>(LHSI), &I,
3186 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
3189 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
3191 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3195 case Instruction::IntToPtr:
3196 // icmp pred inttoptr(X), null -> icmp pred X, 0
3197 if (RHSC->isNullValue() &&
3198 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3199 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
3200 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3203 case Instruction::Load:
3204 // Try to optimize things like "A[i] > 4" to index computations.
3205 if (GetElementPtrInst *GEP =
3206 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3207 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3208 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3209 !cast<LoadInst>(LHSI)->isVolatile())
3210 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3217 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
3218 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
3219 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
3221 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
3222 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
3223 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
3226 // Test to see if the operands of the icmp are casted versions of other
3227 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
3229 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
3230 if (Op0->getType()->isPointerTy() &&
3231 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
3232 // We keep moving the cast from the left operand over to the right
3233 // operand, where it can often be eliminated completely.
3234 Op0 = CI->getOperand(0);
3236 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
3237 // so eliminate it as well.
3238 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
3239 Op1 = CI2->getOperand(0);
3241 // If Op1 is a constant, we can fold the cast into the constant.
3242 if (Op0->getType() != Op1->getType()) {
3243 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3244 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
3246 // Otherwise, cast the RHS right before the icmp
3247 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
3250 return new ICmpInst(I.getPredicate(), Op0, Op1);
3254 if (isa<CastInst>(Op0)) {
3255 // Handle the special case of: icmp (cast bool to X), <cst>
3256 // This comes up when you have code like
3259 // For generality, we handle any zero-extension of any operand comparison
3260 // with a constant or another cast from the same type.
3261 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
3262 if (Instruction *R = visitICmpInstWithCastAndCast(I))
3266 // Special logic for binary operators.
3267 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3268 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3270 CmpInst::Predicate Pred = I.getPredicate();
3271 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3272 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3273 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
3274 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3275 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3276 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3277 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
3278 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3279 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3281 // Analyze the case when either Op0 or Op1 is an add instruction.
3282 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3283 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3284 if (BO0 && BO0->getOpcode() == Instruction::Add)
3285 A = BO0->getOperand(0), B = BO0->getOperand(1);
3286 if (BO1 && BO1->getOpcode() == Instruction::Add)
3287 C = BO1->getOperand(0), D = BO1->getOperand(1);
3289 // icmp (X+cst) < 0 --> X < -cst
3290 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred) && match(Op1, m_Zero()))
3291 if (ConstantInt *RHSC = dyn_cast_or_null<ConstantInt>(B))
3292 if (!RHSC->isMinValue(/*isSigned=*/true))
3293 return new ICmpInst(Pred, A, ConstantExpr::getNeg(RHSC));
3295 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3296 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3297 return new ICmpInst(Pred, A == Op1 ? B : A,
3298 Constant::getNullValue(Op1->getType()));
3300 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3301 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3302 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3305 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3306 if (A && C && (A == C || A == D || B == C || B == D) &&
3307 NoOp0WrapProblem && NoOp1WrapProblem &&
3308 // Try not to increase register pressure.
3309 BO0->hasOneUse() && BO1->hasOneUse()) {
3310 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3313 // C + B == C + D -> B == D
3316 } else if (A == D) {
3317 // D + B == C + D -> B == C
3320 } else if (B == C) {
3321 // A + C == C + D -> A == D
3326 // A + D == C + D -> A == C
3330 return new ICmpInst(Pred, Y, Z);
3333 // icmp slt (X + -1), Y -> icmp sle X, Y
3334 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3335 match(B, m_AllOnes()))
3336 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3338 // icmp sge (X + -1), Y -> icmp sgt X, Y
3339 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3340 match(B, m_AllOnes()))
3341 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3343 // icmp sle (X + 1), Y -> icmp slt X, Y
3344 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
3346 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3348 // icmp sgt (X + 1), Y -> icmp sge X, Y
3349 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
3351 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3353 // if C1 has greater magnitude than C2:
3354 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3355 // s.t. C3 = C1 - C2
3357 // if C2 has greater magnitude than C1:
3358 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3359 // s.t. C3 = C2 - C1
3360 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3361 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3362 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3363 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3364 const APInt &AP1 = C1->getValue();
3365 const APInt &AP2 = C2->getValue();
3366 if (AP1.isNegative() == AP2.isNegative()) {
3367 APInt AP1Abs = C1->getValue().abs();
3368 APInt AP2Abs = C2->getValue().abs();
3369 if (AP1Abs.uge(AP2Abs)) {
3370 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
3371 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
3372 return new ICmpInst(Pred, NewAdd, C);
3374 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
3375 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
3376 return new ICmpInst(Pred, A, NewAdd);
3382 // Analyze the case when either Op0 or Op1 is a sub instruction.
3383 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3384 A = nullptr; B = nullptr; C = nullptr; D = nullptr;
3385 if (BO0 && BO0->getOpcode() == Instruction::Sub)
3386 A = BO0->getOperand(0), B = BO0->getOperand(1);
3387 if (BO1 && BO1->getOpcode() == Instruction::Sub)
3388 C = BO1->getOperand(0), D = BO1->getOperand(1);
3390 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3391 if (A == Op1 && NoOp0WrapProblem)
3392 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3394 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3395 if (C == Op0 && NoOp1WrapProblem)
3396 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3398 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3399 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3400 // Try not to increase register pressure.
3401 BO0->hasOneUse() && BO1->hasOneUse())
3402 return new ICmpInst(Pred, A, C);
3404 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3405 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3406 // Try not to increase register pressure.
3407 BO0->hasOneUse() && BO1->hasOneUse())
3408 return new ICmpInst(Pred, D, B);
3410 // icmp (0-X) < cst --> x > -cst
3411 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3413 if (match(BO0, m_Neg(m_Value(X))))
3414 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3415 if (!RHSC->isMinValue(/*isSigned=*/true))
3416 return new ICmpInst(I.getSwappedPredicate(), X,
3417 ConstantExpr::getNeg(RHSC));
3420 BinaryOperator *SRem = nullptr;
3421 // icmp (srem X, Y), Y
3422 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
3423 Op1 == BO0->getOperand(1))
3425 // icmp Y, (srem X, Y)
3426 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3427 Op0 == BO1->getOperand(1))
3430 // We don't check hasOneUse to avoid increasing register pressure because
3431 // the value we use is the same value this instruction was already using.
3432 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3434 case ICmpInst::ICMP_EQ:
3435 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3436 case ICmpInst::ICMP_NE:
3437 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3438 case ICmpInst::ICMP_SGT:
3439 case ICmpInst::ICMP_SGE:
3440 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3441 Constant::getAllOnesValue(SRem->getType()));
3442 case ICmpInst::ICMP_SLT:
3443 case ICmpInst::ICMP_SLE:
3444 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3445 Constant::getNullValue(SRem->getType()));
3449 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
3450 BO0->hasOneUse() && BO1->hasOneUse() &&
3451 BO0->getOperand(1) == BO1->getOperand(1)) {
3452 switch (BO0->getOpcode()) {
3454 case Instruction::Add:
3455 case Instruction::Sub:
3456 case Instruction::Xor:
3457 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3458 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3459 BO1->getOperand(0));
3460 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
3461 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3462 if (CI->getValue().isSignBit()) {
3463 ICmpInst::Predicate Pred = I.isSigned()
3464 ? I.getUnsignedPredicate()
3465 : I.getSignedPredicate();
3466 return new ICmpInst(Pred, BO0->getOperand(0),
3467 BO1->getOperand(0));
3470 if (CI->isMaxValue(true)) {
3471 ICmpInst::Predicate Pred = I.isSigned()
3472 ? I.getUnsignedPredicate()
3473 : I.getSignedPredicate();
3474 Pred = I.getSwappedPredicate(Pred);
3475 return new ICmpInst(Pred, BO0->getOperand(0),
3476 BO1->getOperand(0));
3480 case Instruction::Mul:
3481 if (!I.isEquality())
3484 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3485 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
3486 // Mask = -1 >> count-trailing-zeros(Cst).
3487 if (!CI->isZero() && !CI->isOne()) {
3488 const APInt &AP = CI->getValue();
3489 ConstantInt *Mask = ConstantInt::get(I.getContext(),
3490 APInt::getLowBitsSet(AP.getBitWidth(),
3492 AP.countTrailingZeros()));
3493 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
3494 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
3495 return new ICmpInst(I.getPredicate(), And1, And2);
3499 case Instruction::UDiv:
3500 case Instruction::LShr:
3504 case Instruction::SDiv:
3505 case Instruction::AShr:
3506 if (!BO0->isExact() || !BO1->isExact())
3508 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3509 BO1->getOperand(0));
3510 case Instruction::Shl: {
3511 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3512 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3515 if (!NSW && I.isSigned())
3517 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3518 BO1->getOperand(0));
3524 // Transform A & (L - 1) `ult` L --> L != 0
3525 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3527 m_CombineOr(m_And(m_Value(), LSubOne), m_And(LSubOne, m_Value()));
3529 if (match(BO0, BitwiseAnd) && I.getPredicate() == ICmpInst::ICMP_ULT) {
3530 auto *Zero = Constant::getNullValue(BO0->getType());
3531 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3537 // Transform (A & ~B) == 0 --> (A & B) != 0
3538 // and (A & ~B) != 0 --> (A & B) == 0
3539 // if A is a power of 2.
3540 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
3541 match(Op1, m_Zero()) &&
3542 isKnownToBeAPowerOfTwo(A, DL, false, 0, AC, &I, DT) && I.isEquality())
3543 return new ICmpInst(I.getInversePredicate(),
3544 Builder->CreateAnd(A, B),
3547 // ~x < ~y --> y < x
3548 // ~x < cst --> ~cst < x
3549 if (match(Op0, m_Not(m_Value(A)))) {
3550 if (match(Op1, m_Not(m_Value(B))))
3551 return new ICmpInst(I.getPredicate(), B, A);
3552 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3553 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
3556 Instruction *AddI = nullptr;
3557 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
3558 m_Instruction(AddI))) &&
3559 isa<IntegerType>(A->getType())) {
3562 if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result,
3564 ReplaceInstUsesWith(*AddI, Result);
3565 return ReplaceInstUsesWith(I, Overflow);
3569 // (zext a) * (zext b) --> llvm.umul.with.overflow.
3570 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3571 if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
3574 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3575 if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
3580 if (I.isEquality()) {
3581 Value *A, *B, *C, *D;
3583 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3584 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3585 Value *OtherVal = A == Op1 ? B : A;
3586 return new ICmpInst(I.getPredicate(), OtherVal,
3587 Constant::getNullValue(A->getType()));
3590 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3591 // A^c1 == C^c2 --> A == C^(c1^c2)
3592 ConstantInt *C1, *C2;
3593 if (match(B, m_ConstantInt(C1)) &&
3594 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
3595 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3596 Value *Xor = Builder->CreateXor(C, NC);
3597 return new ICmpInst(I.getPredicate(), A, Xor);
3600 // A^B == A^D -> B == D
3601 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
3602 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
3603 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
3604 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
3608 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
3609 (A == Op0 || B == Op0)) {
3610 // A == (A^B) -> B == 0
3611 Value *OtherVal = A == Op0 ? B : A;
3612 return new ICmpInst(I.getPredicate(), OtherVal,
3613 Constant::getNullValue(A->getType()));
3616 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3617 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3618 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3619 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3622 X = B; Y = D; Z = A;
3623 } else if (A == D) {
3624 X = B; Y = C; Z = A;
3625 } else if (B == C) {
3626 X = A; Y = D; Z = B;
3627 } else if (B == D) {
3628 X = A; Y = C; Z = B;
3631 if (X) { // Build (X^Y) & Z
3632 Op1 = Builder->CreateXor(X, Y);
3633 Op1 = Builder->CreateAnd(Op1, Z);
3634 I.setOperand(0, Op1);
3635 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3640 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3641 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3643 if ((Op0->hasOneUse() &&
3644 match(Op0, m_ZExt(m_Value(A))) &&
3645 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3646 (Op1->hasOneUse() &&
3647 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3648 match(Op1, m_ZExt(m_Value(A))))) {
3649 APInt Pow2 = Cst1->getValue() + 1;
3650 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3651 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3652 return new ICmpInst(I.getPredicate(), A,
3653 Builder->CreateTrunc(B, A->getType()));
3656 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3657 // For lshr and ashr pairs.
3658 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3659 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3660 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3661 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3662 unsigned TypeBits = Cst1->getBitWidth();
3663 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3664 if (ShAmt < TypeBits && ShAmt != 0) {
3665 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
3666 ? ICmpInst::ICMP_UGE
3667 : ICmpInst::ICMP_ULT;
3668 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3669 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3670 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
3674 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3675 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3676 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3677 unsigned TypeBits = Cst1->getBitWidth();
3678 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3679 if (ShAmt < TypeBits && ShAmt != 0) {
3680 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3681 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3682 Value *And = Builder->CreateAnd(Xor, Builder->getInt(AndVal),
3683 I.getName() + ".mask");
3684 return new ICmpInst(I.getPredicate(), And,
3685 Constant::getNullValue(Cst1->getType()));
3689 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3690 // "icmp (and X, mask), cst"
3692 if (Op0->hasOneUse() &&
3693 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
3694 m_ConstantInt(ShAmt))))) &&
3695 match(Op1, m_ConstantInt(Cst1)) &&
3696 // Only do this when A has multiple uses. This is most important to do
3697 // when it exposes other optimizations.
3699 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3701 if (ShAmt < ASize) {
3703 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3706 APInt CmpV = Cst1->getValue().zext(ASize);
3709 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3710 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3715 // The 'cmpxchg' instruction returns an aggregate containing the old value and
3716 // an i1 which indicates whether or not we successfully did the swap.
3718 // Replace comparisons between the old value and the expected value with the
3719 // indicator that 'cmpxchg' returns.
3721 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
3722 // spuriously fail. In those cases, the old value may equal the expected
3723 // value but it is possible for the swap to not occur.
3724 if (I.getPredicate() == ICmpInst::ICMP_EQ)
3725 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
3726 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
3727 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
3729 return ExtractValueInst::Create(ACXI, 1);
3732 Value *X; ConstantInt *Cst;
3734 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3735 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3738 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3739 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3741 return Changed ? &I : nullptr;
3744 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3745 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3748 if (!isa<ConstantFP>(RHSC)) return nullptr;
3749 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3751 // Get the width of the mantissa. We don't want to hack on conversions that
3752 // might lose information from the integer, e.g. "i64 -> float"
3753 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3754 if (MantissaWidth == -1) return nullptr; // Unknown.
3756 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3758 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3760 if (I.isEquality()) {
3761 FCmpInst::Predicate P = I.getPredicate();
3762 bool IsExact = false;
3763 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
3764 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
3766 // If the floating point constant isn't an integer value, we know if we will
3767 // ever compare equal / not equal to it.
3769 // TODO: Can never be -0.0 and other non-representable values
3770 APFloat RHSRoundInt(RHS);
3771 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
3772 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
3773 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
3774 return ReplaceInstUsesWith(I, Builder->getFalse());
3776 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
3777 return ReplaceInstUsesWith(I, Builder->getTrue());
3781 // TODO: If the constant is exactly representable, is it always OK to do
3782 // equality compares as integer?
3785 // Check to see that the input is converted from an integer type that is small
3786 // enough that preserves all bits. TODO: check here for "known" sign bits.
3787 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3788 unsigned InputSize = IntTy->getScalarSizeInBits();
3790 // Following test does NOT adjust InputSize downwards for signed inputs,
3791 // because the most negative value still requires all the mantissa bits
3792 // to distinguish it from one less than that value.
3793 if ((int)InputSize > MantissaWidth) {
3794 // Conversion would lose accuracy. Check if loss can impact comparison.
3795 int Exp = ilogb(RHS);
3796 if (Exp == APFloat::IEK_Inf) {
3797 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
3798 if (MaxExponent < (int)InputSize - !LHSUnsigned)
3799 // Conversion could create infinity.
3802 // Note that if RHS is zero or NaN, then Exp is negative
3803 // and first condition is trivially false.
3804 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
3805 // Conversion could affect comparison.
3810 // Otherwise, we can potentially simplify the comparison. We know that it
3811 // will always come through as an integer value and we know the constant is
3812 // not a NAN (it would have been previously simplified).
3813 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3815 ICmpInst::Predicate Pred;
3816 switch (I.getPredicate()) {
3817 default: llvm_unreachable("Unexpected predicate!");
3818 case FCmpInst::FCMP_UEQ:
3819 case FCmpInst::FCMP_OEQ:
3820 Pred = ICmpInst::ICMP_EQ;
3822 case FCmpInst::FCMP_UGT:
3823 case FCmpInst::FCMP_OGT:
3824 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3826 case FCmpInst::FCMP_UGE:
3827 case FCmpInst::FCMP_OGE:
3828 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3830 case FCmpInst::FCMP_ULT:
3831 case FCmpInst::FCMP_OLT:
3832 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3834 case FCmpInst::FCMP_ULE:
3835 case FCmpInst::FCMP_OLE:
3836 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3838 case FCmpInst::FCMP_UNE:
3839 case FCmpInst::FCMP_ONE:
3840 Pred = ICmpInst::ICMP_NE;
3842 case FCmpInst::FCMP_ORD:
3843 return ReplaceInstUsesWith(I, Builder->getTrue());
3844 case FCmpInst::FCMP_UNO:
3845 return ReplaceInstUsesWith(I, Builder->getFalse());
3848 // Now we know that the APFloat is a normal number, zero or inf.
3850 // See if the FP constant is too large for the integer. For example,
3851 // comparing an i8 to 300.0.
3852 unsigned IntWidth = IntTy->getScalarSizeInBits();
3855 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3856 // and large values.
3857 APFloat SMax(RHS.getSemantics());
3858 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3859 APFloat::rmNearestTiesToEven);
3860 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3861 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3862 Pred == ICmpInst::ICMP_SLE)
3863 return ReplaceInstUsesWith(I, Builder->getTrue());
3864 return ReplaceInstUsesWith(I, Builder->getFalse());
3867 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3868 // +INF and large values.
3869 APFloat UMax(RHS.getSemantics());
3870 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3871 APFloat::rmNearestTiesToEven);
3872 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3873 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3874 Pred == ICmpInst::ICMP_ULE)
3875 return ReplaceInstUsesWith(I, Builder->getTrue());
3876 return ReplaceInstUsesWith(I, Builder->getFalse());
3881 // See if the RHS value is < SignedMin.
3882 APFloat SMin(RHS.getSemantics());
3883 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3884 APFloat::rmNearestTiesToEven);
3885 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3886 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3887 Pred == ICmpInst::ICMP_SGE)
3888 return ReplaceInstUsesWith(I, Builder->getTrue());
3889 return ReplaceInstUsesWith(I, Builder->getFalse());
3892 // See if the RHS value is < UnsignedMin.
3893 APFloat SMin(RHS.getSemantics());
3894 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3895 APFloat::rmNearestTiesToEven);
3896 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3897 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3898 Pred == ICmpInst::ICMP_UGE)
3899 return ReplaceInstUsesWith(I, Builder->getTrue());
3900 return ReplaceInstUsesWith(I, Builder->getFalse());
3904 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3905 // [0, UMAX], but it may still be fractional. See if it is fractional by
3906 // casting the FP value to the integer value and back, checking for equality.
3907 // Don't do this for zero, because -0.0 is not fractional.
3908 Constant *RHSInt = LHSUnsigned
3909 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3910 : ConstantExpr::getFPToSI(RHSC, IntTy);
3911 if (!RHS.isZero()) {
3912 bool Equal = LHSUnsigned
3913 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3914 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3916 // If we had a comparison against a fractional value, we have to adjust
3917 // the compare predicate and sometimes the value. RHSC is rounded towards
3918 // zero at this point.
3920 default: llvm_unreachable("Unexpected integer comparison!");
3921 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3922 return ReplaceInstUsesWith(I, Builder->getTrue());
3923 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3924 return ReplaceInstUsesWith(I, Builder->getFalse());
3925 case ICmpInst::ICMP_ULE:
3926 // (float)int <= 4.4 --> int <= 4
3927 // (float)int <= -4.4 --> false
3928 if (RHS.isNegative())
3929 return ReplaceInstUsesWith(I, Builder->getFalse());
3931 case ICmpInst::ICMP_SLE:
3932 // (float)int <= 4.4 --> int <= 4
3933 // (float)int <= -4.4 --> int < -4
3934 if (RHS.isNegative())
3935 Pred = ICmpInst::ICMP_SLT;
3937 case ICmpInst::ICMP_ULT:
3938 // (float)int < -4.4 --> false
3939 // (float)int < 4.4 --> int <= 4
3940 if (RHS.isNegative())
3941 return ReplaceInstUsesWith(I, Builder->getFalse());
3942 Pred = ICmpInst::ICMP_ULE;
3944 case ICmpInst::ICMP_SLT:
3945 // (float)int < -4.4 --> int < -4
3946 // (float)int < 4.4 --> int <= 4
3947 if (!RHS.isNegative())
3948 Pred = ICmpInst::ICMP_SLE;
3950 case ICmpInst::ICMP_UGT:
3951 // (float)int > 4.4 --> int > 4
3952 // (float)int > -4.4 --> true
3953 if (RHS.isNegative())
3954 return ReplaceInstUsesWith(I, Builder->getTrue());
3956 case ICmpInst::ICMP_SGT:
3957 // (float)int > 4.4 --> int > 4
3958 // (float)int > -4.4 --> int >= -4
3959 if (RHS.isNegative())
3960 Pred = ICmpInst::ICMP_SGE;
3962 case ICmpInst::ICMP_UGE:
3963 // (float)int >= -4.4 --> true
3964 // (float)int >= 4.4 --> int > 4
3965 if (RHS.isNegative())
3966 return ReplaceInstUsesWith(I, Builder->getTrue());
3967 Pred = ICmpInst::ICMP_UGT;
3969 case ICmpInst::ICMP_SGE:
3970 // (float)int >= -4.4 --> int >= -4
3971 // (float)int >= 4.4 --> int > 4
3972 if (!RHS.isNegative())
3973 Pred = ICmpInst::ICMP_SGT;
3979 // Lower this FP comparison into an appropriate integer version of the
3981 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3984 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3985 bool Changed = false;
3987 /// Orders the operands of the compare so that they are listed from most
3988 /// complex to least complex. This puts constants before unary operators,
3989 /// before binary operators.
3990 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3995 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3997 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1,
3998 I.getFastMathFlags(), DL, TLI, DT, AC, &I))
3999 return ReplaceInstUsesWith(I, V);
4001 // Simplify 'fcmp pred X, X'
4003 switch (I.getPredicate()) {
4004 default: llvm_unreachable("Unknown predicate!");
4005 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4006 case FCmpInst::FCMP_ULT: // True if unordered or less than
4007 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4008 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4009 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4010 I.setPredicate(FCmpInst::FCMP_UNO);
4011 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4014 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4015 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4016 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4017 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4018 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4019 I.setPredicate(FCmpInst::FCMP_ORD);
4020 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4025 // Test if the FCmpInst instruction is used exclusively by a select as
4026 // part of a minimum or maximum operation. If so, refrain from doing
4027 // any other folding. This helps out other analyses which understand
4028 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4029 // and CodeGen. And in this case, at least one of the comparison
4030 // operands has at least one user besides the compare (the select),
4031 // which would often largely negate the benefit of folding anyway.
4033 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
4034 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
4035 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
4038 // Handle fcmp with constant RHS
4039 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4040 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4041 switch (LHSI->getOpcode()) {
4042 case Instruction::FPExt: {
4043 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
4044 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
4045 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
4049 const fltSemantics *Sem;
4050 // FIXME: This shouldn't be here.
4051 if (LHSExt->getSrcTy()->isHalfTy())
4052 Sem = &APFloat::IEEEhalf;
4053 else if (LHSExt->getSrcTy()->isFloatTy())
4054 Sem = &APFloat::IEEEsingle;
4055 else if (LHSExt->getSrcTy()->isDoubleTy())
4056 Sem = &APFloat::IEEEdouble;
4057 else if (LHSExt->getSrcTy()->isFP128Ty())
4058 Sem = &APFloat::IEEEquad;
4059 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
4060 Sem = &APFloat::x87DoubleExtended;
4061 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
4062 Sem = &APFloat::PPCDoubleDouble;
4067 APFloat F = RHSF->getValueAPF();
4068 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
4070 // Avoid lossy conversions and denormals. Zero is a special case
4071 // that's OK to convert.
4075 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
4076 APFloat::cmpLessThan) || Fabs.isZero()))
4078 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
4079 ConstantFP::get(RHSC->getContext(), F));
4082 case Instruction::PHI:
4083 // Only fold fcmp into the PHI if the phi and fcmp are in the same
4084 // block. If in the same block, we're encouraging jump threading. If
4085 // not, we are just pessimizing the code by making an i1 phi.
4086 if (LHSI->getParent() == I.getParent())
4087 if (Instruction *NV = FoldOpIntoPhi(I))
4090 case Instruction::SIToFP:
4091 case Instruction::UIToFP:
4092 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
4095 case Instruction::FSub: {
4096 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
4098 if (match(LHSI, m_FNeg(m_Value(Op))))
4099 return new FCmpInst(I.getSwappedPredicate(), Op,
4100 ConstantExpr::getFNeg(RHSC));
4103 case Instruction::Load:
4104 if (GetElementPtrInst *GEP =
4105 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
4106 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
4107 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
4108 !cast<LoadInst>(LHSI)->isVolatile())
4109 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
4113 case Instruction::Call: {
4114 if (!RHSC->isNullValue())
4117 CallInst *CI = cast<CallInst>(LHSI);
4118 const Function *F = CI->getCalledFunction();
4122 // Various optimization for fabs compared with zero.
4124 if (F->getIntrinsicID() == Intrinsic::fabs ||
4125 (TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
4126 (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
4127 Func == LibFunc::fabsl))) {
4128 switch (I.getPredicate()) {
4131 // fabs(x) < 0 --> false
4132 case FCmpInst::FCMP_OLT:
4133 return ReplaceInstUsesWith(I, Builder->getFalse());
4134 // fabs(x) > 0 --> x != 0
4135 case FCmpInst::FCMP_OGT:
4136 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), RHSC);
4137 // fabs(x) <= 0 --> x == 0
4138 case FCmpInst::FCMP_OLE:
4139 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), RHSC);
4140 // fabs(x) >= 0 --> !isnan(x)
4141 case FCmpInst::FCMP_OGE:
4142 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), RHSC);
4143 // fabs(x) == 0 --> x == 0
4144 // fabs(x) != 0 --> x != 0
4145 case FCmpInst::FCMP_OEQ:
4146 case FCmpInst::FCMP_UEQ:
4147 case FCmpInst::FCMP_ONE:
4148 case FCmpInst::FCMP_UNE:
4149 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), RHSC);
4156 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
4158 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
4159 return new FCmpInst(I.getSwappedPredicate(), X, Y);
4161 // fcmp (fpext x), (fpext y) -> fcmp x, y
4162 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
4163 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
4164 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
4165 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
4166 RHSExt->getOperand(0));
4168 return Changed ? &I : nullptr;