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 // We need TD information to know the pointer size unless this is inbounds.
233 if (!GEP->isInBounds() && !DL)
236 Constant *Init = GV->getInitializer();
237 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
240 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
241 if (ArrayElementCount > 1024) return nullptr; // Don't blow up on huge arrays.
243 // There are many forms of this optimization we can handle, for now, just do
244 // the simple index into a single-dimensional array.
246 // Require: GEP GV, 0, i {{, constant indices}}
247 if (GEP->getNumOperands() < 3 ||
248 !isa<ConstantInt>(GEP->getOperand(1)) ||
249 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
250 isa<Constant>(GEP->getOperand(2)))
253 // Check that indices after the variable are constants and in-range for the
254 // type they index. Collect the indices. This is typically for arrays of
256 SmallVector<unsigned, 4> LaterIndices;
258 Type *EltTy = Init->getType()->getArrayElementType();
259 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
260 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
261 if (!Idx) return nullptr; // Variable index.
263 uint64_t IdxVal = Idx->getZExtValue();
264 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
266 if (StructType *STy = dyn_cast<StructType>(EltTy))
267 EltTy = STy->getElementType(IdxVal);
268 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
269 if (IdxVal >= ATy->getNumElements()) return nullptr;
270 EltTy = ATy->getElementType();
272 return nullptr; // Unknown type.
275 LaterIndices.push_back(IdxVal);
278 enum { Overdefined = -3, Undefined = -2 };
280 // Variables for our state machines.
282 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
283 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
284 // and 87 is the second (and last) index. FirstTrueElement is -2 when
285 // undefined, otherwise set to the first true element. SecondTrueElement is
286 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
287 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
289 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
290 // form "i != 47 & i != 87". Same state transitions as for true elements.
291 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
293 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
294 /// define a state machine that triggers for ranges of values that the index
295 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
296 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
297 /// index in the range (inclusive). We use -2 for undefined here because we
298 /// use relative comparisons and don't want 0-1 to match -1.
299 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
301 // MagicBitvector - This is a magic bitvector where we set a bit if the
302 // comparison is true for element 'i'. If there are 64 elements or less in
303 // the array, this will fully represent all the comparison results.
304 uint64_t MagicBitvector = 0;
307 // Scan the array and see if one of our patterns matches.
308 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
309 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
310 Constant *Elt = Init->getAggregateElement(i);
311 if (!Elt) return nullptr;
313 // If this is indexing an array of structures, get the structure element.
314 if (!LaterIndices.empty())
315 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
317 // If the element is masked, handle it.
318 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
320 // Find out if the comparison would be true or false for the i'th element.
321 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
322 CompareRHS, DL, TLI);
323 // If the result is undef for this element, ignore it.
324 if (isa<UndefValue>(C)) {
325 // Extend range state machines to cover this element in case there is an
326 // undef in the middle of the range.
327 if (TrueRangeEnd == (int)i-1)
329 if (FalseRangeEnd == (int)i-1)
334 // If we can't compute the result for any of the elements, we have to give
335 // up evaluating the entire conditional.
336 if (!isa<ConstantInt>(C)) return nullptr;
338 // Otherwise, we know if the comparison is true or false for this element,
339 // update our state machines.
340 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
342 // State machine for single/double/range index comparison.
344 // Update the TrueElement state machine.
345 if (FirstTrueElement == Undefined)
346 FirstTrueElement = TrueRangeEnd = i; // First true element.
348 // Update double-compare state machine.
349 if (SecondTrueElement == Undefined)
350 SecondTrueElement = i;
352 SecondTrueElement = Overdefined;
354 // Update range state machine.
355 if (TrueRangeEnd == (int)i-1)
358 TrueRangeEnd = Overdefined;
361 // Update the FalseElement state machine.
362 if (FirstFalseElement == Undefined)
363 FirstFalseElement = FalseRangeEnd = i; // First false element.
365 // Update double-compare state machine.
366 if (SecondFalseElement == Undefined)
367 SecondFalseElement = i;
369 SecondFalseElement = Overdefined;
371 // Update range state machine.
372 if (FalseRangeEnd == (int)i-1)
375 FalseRangeEnd = Overdefined;
380 // If this element is in range, update our magic bitvector.
381 if (i < 64 && IsTrueForElt)
382 MagicBitvector |= 1ULL << i;
384 // If all of our states become overdefined, bail out early. Since the
385 // predicate is expensive, only check it every 8 elements. This is only
386 // really useful for really huge arrays.
387 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
388 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
389 FalseRangeEnd == Overdefined)
393 // Now that we've scanned the entire array, emit our new comparison(s). We
394 // order the state machines in complexity of the generated code.
395 Value *Idx = GEP->getOperand(2);
397 // If the index is larger than the pointer size of the target, truncate the
398 // index down like the GEP would do implicitly. We don't have to do this for
399 // an inbounds GEP because the index can't be out of range.
400 if (!GEP->isInBounds()) {
401 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
402 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
403 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
404 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
407 // If the comparison is only true for one or two elements, emit direct
409 if (SecondTrueElement != Overdefined) {
410 // None true -> false.
411 if (FirstTrueElement == Undefined)
412 return ReplaceInstUsesWith(ICI, Builder->getFalse());
414 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
416 // True for one element -> 'i == 47'.
417 if (SecondTrueElement == Undefined)
418 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
420 // True for two elements -> 'i == 47 | i == 72'.
421 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
422 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
423 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
424 return BinaryOperator::CreateOr(C1, C2);
427 // If the comparison is only false for one or two elements, emit direct
429 if (SecondFalseElement != Overdefined) {
430 // None false -> true.
431 if (FirstFalseElement == Undefined)
432 return ReplaceInstUsesWith(ICI, Builder->getTrue());
434 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
436 // False for one element -> 'i != 47'.
437 if (SecondFalseElement == Undefined)
438 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
440 // False for two elements -> 'i != 47 & i != 72'.
441 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
442 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
443 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
444 return BinaryOperator::CreateAnd(C1, C2);
447 // If the comparison can be replaced with a range comparison for the elements
448 // where it is true, emit the range check.
449 if (TrueRangeEnd != Overdefined) {
450 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
452 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
453 if (FirstTrueElement) {
454 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
455 Idx = Builder->CreateAdd(Idx, Offs);
458 Value *End = ConstantInt::get(Idx->getType(),
459 TrueRangeEnd-FirstTrueElement+1);
460 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
463 // False range check.
464 if (FalseRangeEnd != Overdefined) {
465 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
466 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
467 if (FirstFalseElement) {
468 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
469 Idx = Builder->CreateAdd(Idx, Offs);
472 Value *End = ConstantInt::get(Idx->getType(),
473 FalseRangeEnd-FirstFalseElement);
474 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
478 // If a magic bitvector captures the entire comparison state
479 // of this load, replace it with computation that does:
480 // ((magic_cst >> i) & 1) != 0
484 // Look for an appropriate type:
485 // - The type of Idx if the magic fits
486 // - The smallest fitting legal type if we have a DataLayout
488 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
491 Ty = DL->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
492 else if (ArrayElementCount <= 32)
493 Ty = Type::getInt32Ty(Init->getContext());
496 Value *V = Builder->CreateIntCast(Idx, Ty, false);
497 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
498 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
499 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
507 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
508 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
509 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
510 /// be complex, and scales are involved. The above expression would also be
511 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
512 /// This later form is less amenable to optimization though, and we are allowed
513 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
515 /// If we can't emit an optimized form for this expression, this returns null.
517 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
518 const DataLayout &DL = *IC.getDataLayout();
519 gep_type_iterator GTI = gep_type_begin(GEP);
521 // Check to see if this gep only has a single variable index. If so, and if
522 // any constant indices are a multiple of its scale, then we can compute this
523 // in terms of the scale of the variable index. For example, if the GEP
524 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
525 // because the expression will cross zero at the same point.
526 unsigned i, e = GEP->getNumOperands();
528 for (i = 1; i != e; ++i, ++GTI) {
529 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
530 // Compute the aggregate offset of constant indices.
531 if (CI->isZero()) continue;
533 // Handle a struct index, which adds its field offset to the pointer.
534 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
535 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
537 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
538 Offset += Size*CI->getSExtValue();
541 // Found our variable index.
546 // If there are no variable indices, we must have a constant offset, just
547 // evaluate it the general way.
548 if (i == e) return nullptr;
550 Value *VariableIdx = GEP->getOperand(i);
551 // Determine the scale factor of the variable element. For example, this is
552 // 4 if the variable index is into an array of i32.
553 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
555 // Verify that there are no other variable indices. If so, emit the hard way.
556 for (++i, ++GTI; i != e; ++i, ++GTI) {
557 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
558 if (!CI) return nullptr;
560 // Compute the aggregate offset of constant indices.
561 if (CI->isZero()) continue;
563 // Handle a struct index, which adds its field offset to the pointer.
564 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
565 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
567 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
568 Offset += Size*CI->getSExtValue();
574 // Okay, we know we have a single variable index, which must be a
575 // pointer/array/vector index. If there is no offset, life is simple, return
577 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
578 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
580 // Cast to intptrty in case a truncation occurs. If an extension is needed,
581 // we don't need to bother extending: the extension won't affect where the
582 // computation crosses zero.
583 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
584 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
589 // Otherwise, there is an index. The computation we will do will be modulo
590 // the pointer size, so get it.
591 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
593 Offset &= PtrSizeMask;
594 VariableScale &= PtrSizeMask;
596 // To do this transformation, any constant index must be a multiple of the
597 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
598 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
599 // multiple of the variable scale.
600 int64_t NewOffs = Offset / (int64_t)VariableScale;
601 if (Offset != NewOffs*(int64_t)VariableScale)
604 // Okay, we can do this evaluation. Start by converting the index to intptr.
605 if (VariableIdx->getType() != IntPtrTy)
606 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
608 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
609 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
612 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
613 /// else. At this point we know that the GEP is on the LHS of the comparison.
614 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
615 ICmpInst::Predicate Cond,
617 // Don't transform signed compares of GEPs into index compares. Even if the
618 // GEP is inbounds, the final add of the base pointer can have signed overflow
619 // and would change the result of the icmp.
620 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
621 // the maximum signed value for the pointer type.
622 if (ICmpInst::isSigned(Cond))
625 // Look through bitcasts and addrspacecasts. We do not however want to remove
627 if (!isa<GetElementPtrInst>(RHS))
628 RHS = RHS->stripPointerCasts();
630 Value *PtrBase = GEPLHS->getOperand(0);
631 if (DL && PtrBase == RHS && GEPLHS->isInBounds()) {
632 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
633 // This transformation (ignoring the base and scales) is valid because we
634 // know pointers can't overflow since the gep is inbounds. See if we can
635 // output an optimized form.
636 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
638 // If not, synthesize the offset the hard way.
640 Offset = EmitGEPOffset(GEPLHS);
641 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
642 Constant::getNullValue(Offset->getType()));
643 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
644 // If the base pointers are different, but the indices are the same, just
645 // compare the base pointer.
646 if (PtrBase != GEPRHS->getOperand(0)) {
647 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
648 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
649 GEPRHS->getOperand(0)->getType();
651 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
652 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
653 IndicesTheSame = false;
657 // If all indices are the same, just compare the base pointers.
659 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
661 // If we're comparing GEPs with two base pointers that only differ in type
662 // and both GEPs have only constant indices or just one use, then fold
663 // the compare with the adjusted indices.
664 if (DL && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
665 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
666 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
667 PtrBase->stripPointerCasts() ==
668 GEPRHS->getOperand(0)->stripPointerCasts()) {
669 Value *LOffset = EmitGEPOffset(GEPLHS);
670 Value *ROffset = EmitGEPOffset(GEPRHS);
672 // If we looked through an addrspacecast between different sized address
673 // spaces, the LHS and RHS pointers are different sized
674 // integers. Truncate to the smaller one.
675 Type *LHSIndexTy = LOffset->getType();
676 Type *RHSIndexTy = ROffset->getType();
677 if (LHSIndexTy != RHSIndexTy) {
678 if (LHSIndexTy->getPrimitiveSizeInBits() <
679 RHSIndexTy->getPrimitiveSizeInBits()) {
680 ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy);
682 LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy);
685 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
687 return ReplaceInstUsesWith(I, Cmp);
690 // Otherwise, the base pointers are different and the indices are
691 // different, bail out.
695 // If one of the GEPs has all zero indices, recurse.
696 if (GEPLHS->hasAllZeroIndices())
697 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
698 ICmpInst::getSwappedPredicate(Cond), I);
700 // If the other GEP has all zero indices, recurse.
701 if (GEPRHS->hasAllZeroIndices())
702 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
704 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
705 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
706 // If the GEPs only differ by one index, compare it.
707 unsigned NumDifferences = 0; // Keep track of # differences.
708 unsigned DiffOperand = 0; // The operand that differs.
709 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
710 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
711 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
712 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
713 // Irreconcilable differences.
717 if (NumDifferences++) break;
722 if (NumDifferences == 0) // SAME GEP?
723 return ReplaceInstUsesWith(I, // No comparison is needed here.
724 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
726 else if (NumDifferences == 1 && GEPsInBounds) {
727 Value *LHSV = GEPLHS->getOperand(DiffOperand);
728 Value *RHSV = GEPRHS->getOperand(DiffOperand);
729 // Make sure we do a signed comparison here.
730 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
734 // Only lower this if the icmp is the only user of the GEP or if we expect
735 // the result to fold to a constant!
738 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
739 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
740 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
741 Value *L = EmitGEPOffset(GEPLHS);
742 Value *R = EmitGEPOffset(GEPRHS);
743 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
749 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
750 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
751 Value *X, ConstantInt *CI,
752 ICmpInst::Predicate Pred) {
753 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
754 // so the values can never be equal. Similarly for all other "or equals"
757 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
758 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
759 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
760 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
762 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
763 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
766 // (X+1) >u X --> X <u (0-1) --> X != 255
767 // (X+2) >u X --> X <u (0-2) --> X <u 254
768 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
769 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
770 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
772 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
773 ConstantInt *SMax = ConstantInt::get(X->getContext(),
774 APInt::getSignedMaxValue(BitWidth));
776 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
777 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
778 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
779 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
780 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
781 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
782 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
783 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
785 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
786 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
787 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
788 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
789 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
790 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
792 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
793 Constant *C = Builder->getInt(CI->getValue()-1);
794 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
797 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
798 /// and CmpRHS are both known to be integer constants.
799 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
800 ConstantInt *DivRHS) {
801 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
802 const APInt &CmpRHSV = CmpRHS->getValue();
804 // FIXME: If the operand types don't match the type of the divide
805 // then don't attempt this transform. The code below doesn't have the
806 // logic to deal with a signed divide and an unsigned compare (and
807 // vice versa). This is because (x /s C1) <s C2 produces different
808 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
809 // (x /u C1) <u C2. Simply casting the operands and result won't
810 // work. :( The if statement below tests that condition and bails
812 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
813 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
815 if (DivRHS->isZero())
816 return nullptr; // The ProdOV computation fails on divide by zero.
817 if (DivIsSigned && DivRHS->isAllOnesValue())
818 return nullptr; // The overflow computation also screws up here
819 if (DivRHS->isOne()) {
820 // This eliminates some funny cases with INT_MIN.
821 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
825 // Compute Prod = CI * DivRHS. We are essentially solving an equation
826 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
827 // C2 (CI). By solving for X we can turn this into a range check
828 // instead of computing a divide.
829 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
831 // Determine if the product overflows by seeing if the product is
832 // not equal to the divide. Make sure we do the same kind of divide
833 // as in the LHS instruction that we're folding.
834 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
835 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
837 // Get the ICmp opcode
838 ICmpInst::Predicate Pred = ICI.getPredicate();
840 /// If the division is known to be exact, then there is no remainder from the
841 /// divide, so the covered range size is unit, otherwise it is the divisor.
842 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
844 // Figure out the interval that is being checked. For example, a comparison
845 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
846 // Compute this interval based on the constants involved and the signedness of
847 // the compare/divide. This computes a half-open interval, keeping track of
848 // whether either value in the interval overflows. After analysis each
849 // overflow variable is set to 0 if it's corresponding bound variable is valid
850 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
851 int LoOverflow = 0, HiOverflow = 0;
852 Constant *LoBound = nullptr, *HiBound = nullptr;
854 if (!DivIsSigned) { // udiv
855 // e.g. X/5 op 3 --> [15, 20)
857 HiOverflow = LoOverflow = ProdOV;
859 // If this is not an exact divide, then many values in the range collapse
860 // to the same result value.
861 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
864 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
865 if (CmpRHSV == 0) { // (X / pos) op 0
866 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
867 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
869 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
870 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
871 HiOverflow = LoOverflow = ProdOV;
873 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
874 } else { // (X / pos) op neg
875 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
876 HiBound = AddOne(Prod);
877 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
879 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
880 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
883 } else if (DivRHS->isNegative()) { // Divisor is < 0.
885 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
886 if (CmpRHSV == 0) { // (X / neg) op 0
887 // e.g. X/-5 op 0 --> [-4, 5)
888 LoBound = AddOne(RangeSize);
889 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
890 if (HiBound == DivRHS) { // -INTMIN = INTMIN
891 HiOverflow = 1; // [INTMIN+1, overflow)
892 HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
894 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
895 // e.g. X/-5 op 3 --> [-19, -14)
896 HiBound = AddOne(Prod);
897 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
899 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
900 } else { // (X / neg) op neg
901 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
902 LoOverflow = HiOverflow = ProdOV;
904 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
907 // Dividing by a negative swaps the condition. LT <-> GT
908 Pred = ICmpInst::getSwappedPredicate(Pred);
911 Value *X = DivI->getOperand(0);
913 default: llvm_unreachable("Unhandled icmp opcode!");
914 case ICmpInst::ICMP_EQ:
915 if (LoOverflow && HiOverflow)
916 return ReplaceInstUsesWith(ICI, Builder->getFalse());
918 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
919 ICmpInst::ICMP_UGE, X, LoBound);
921 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
922 ICmpInst::ICMP_ULT, X, HiBound);
923 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
925 case ICmpInst::ICMP_NE:
926 if (LoOverflow && HiOverflow)
927 return ReplaceInstUsesWith(ICI, Builder->getTrue());
929 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
930 ICmpInst::ICMP_ULT, X, LoBound);
932 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
933 ICmpInst::ICMP_UGE, X, HiBound);
934 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
935 DivIsSigned, false));
936 case ICmpInst::ICMP_ULT:
937 case ICmpInst::ICMP_SLT:
938 if (LoOverflow == +1) // Low bound is greater than input range.
939 return ReplaceInstUsesWith(ICI, Builder->getTrue());
940 if (LoOverflow == -1) // Low bound is less than input range.
941 return ReplaceInstUsesWith(ICI, Builder->getFalse());
942 return new ICmpInst(Pred, X, LoBound);
943 case ICmpInst::ICMP_UGT:
944 case ICmpInst::ICMP_SGT:
945 if (HiOverflow == +1) // High bound greater than input range.
946 return ReplaceInstUsesWith(ICI, Builder->getFalse());
947 if (HiOverflow == -1) // High bound less than input range.
948 return ReplaceInstUsesWith(ICI, Builder->getTrue());
949 if (Pred == ICmpInst::ICMP_UGT)
950 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
951 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
955 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
956 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
957 ConstantInt *ShAmt) {
958 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
960 // Check that the shift amount is in range. If not, don't perform
961 // undefined shifts. When the shift is visited it will be
963 uint32_t TypeBits = CmpRHSV.getBitWidth();
964 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
965 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
968 if (!ICI.isEquality()) {
969 // If we have an unsigned comparison and an ashr, we can't simplify this.
970 // Similarly for signed comparisons with lshr.
971 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
974 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
975 // by a power of 2. Since we already have logic to simplify these,
976 // transform to div and then simplify the resultant comparison.
977 if (Shr->getOpcode() == Instruction::AShr &&
978 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
981 // Revisit the shift (to delete it).
985 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
988 Shr->getOpcode() == Instruction::AShr ?
989 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
990 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
992 ICI.setOperand(0, Tmp);
994 // If the builder folded the binop, just return it.
995 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
999 // Otherwise, fold this div/compare.
1000 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1001 TheDiv->getOpcode() == Instruction::UDiv);
1003 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1004 assert(Res && "This div/cst should have folded!");
1009 // If we are comparing against bits always shifted out, the
1010 // comparison cannot succeed.
1011 APInt Comp = CmpRHSV << ShAmtVal;
1012 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1013 if (Shr->getOpcode() == Instruction::LShr)
1014 Comp = Comp.lshr(ShAmtVal);
1016 Comp = Comp.ashr(ShAmtVal);
1018 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1019 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1020 Constant *Cst = Builder->getInt1(IsICMP_NE);
1021 return ReplaceInstUsesWith(ICI, Cst);
1024 // Otherwise, check to see if the bits shifted out are known to be zero.
1025 // If so, we can compare against the unshifted value:
1026 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1027 if (Shr->hasOneUse() && Shr->isExact())
1028 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1030 if (Shr->hasOneUse()) {
1031 // Otherwise strength reduce the shift into an and.
1032 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1033 Constant *Mask = Builder->getInt(Val);
1035 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1036 Mask, Shr->getName()+".mask");
1037 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1042 /// FoldICmpCstShrCst - Handle "(icmp eq/ne (ashr/lshr const2, A), const1)" ->
1043 /// (icmp eq/ne A, Log2(const2/const1)) ->
1044 /// (icmp eq/ne A, Log2(const2) - Log2(const1)).
1045 Instruction *InstCombiner::FoldICmpCstShrCst(ICmpInst &I, Value *Op, Value *A,
1048 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1050 auto getConstant = [&I, this](bool IsTrue) {
1051 if (I.getPredicate() == I.ICMP_NE)
1053 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1056 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1057 if (I.getPredicate() == I.ICMP_NE)
1058 Pred = CmpInst::getInversePredicate(Pred);
1059 return new ICmpInst(Pred, LHS, RHS);
1062 APInt AP1 = CI1->getValue();
1063 APInt AP2 = CI2->getValue();
1065 // Don't bother doing any work for cases which InstSimplify handles.
1068 bool IsAShr = isa<AShrOperator>(Op);
1070 if (AP2.isAllOnesValue())
1072 if (AP2.isNegative() != AP1.isNegative())
1079 // 'A' must be large enough to shift out the highest set bit.
1080 return getICmp(I.ICMP_UGT, A,
1081 ConstantInt::get(A->getType(), AP2.logBase2()));
1084 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1086 // Get the distance between the highest bit that's set.
1088 // Both the constants are negative, take their positive to calculate log.
1089 if (IsAShr && AP1.isNegative())
1090 // Get the ones' complement of AP2 and AP1 when computing the distance.
1091 Shift = (~AP2).logBase2() - (~AP1).logBase2();
1093 Shift = AP2.logBase2() - AP1.logBase2();
1096 if (IsAShr ? AP1 == AP2.ashr(Shift) : AP1 == AP2.lshr(Shift))
1097 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1099 // Shifting const2 will never be equal to const1.
1100 return getConstant(false);
1103 /// FoldICmpCstShlCst - Handle "(icmp eq/ne (shl const2, A), const1)" ->
1104 /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)).
1105 Instruction *InstCombiner::FoldICmpCstShlCst(ICmpInst &I, Value *Op, Value *A,
1108 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1110 auto getConstant = [&I, this](bool IsTrue) {
1111 if (I.getPredicate() == I.ICMP_NE)
1113 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1116 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1117 if (I.getPredicate() == I.ICMP_NE)
1118 Pred = CmpInst::getInversePredicate(Pred);
1119 return new ICmpInst(Pred, LHS, RHS);
1122 APInt AP1 = CI1->getValue();
1123 APInt AP2 = CI2->getValue();
1125 // Don't bother doing any work for cases which InstSimplify handles.
1129 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1131 if (!AP1 && AP2TrailingZeros != 0)
1132 return getICmp(I.ICMP_UGE, A,
1133 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1136 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1138 // Get the distance between the lowest bits that are set.
1139 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1141 if (Shift > 0 && AP2.shl(Shift) == AP1)
1142 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1144 // Shifting const2 will never be equal to const1.
1145 return getConstant(false);
1148 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1150 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1153 const APInt &RHSV = RHS->getValue();
1155 switch (LHSI->getOpcode()) {
1156 case Instruction::Trunc:
1157 if (ICI.isEquality() && LHSI->hasOneUse()) {
1158 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1159 // of the high bits truncated out of x are known.
1160 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1161 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1162 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1163 computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne, 0, &ICI);
1165 // If all the high bits are known, we can do this xform.
1166 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1167 // Pull in the high bits from known-ones set.
1168 APInt NewRHS = RHS->getValue().zext(SrcBits);
1169 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1170 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1171 Builder->getInt(NewRHS));
1176 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1177 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1178 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1180 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1181 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1182 Value *CompareVal = LHSI->getOperand(0);
1184 // If the sign bit of the XorCst is not set, there is no change to
1185 // the operation, just stop using the Xor.
1186 if (!XorCst->isNegative()) {
1187 ICI.setOperand(0, CompareVal);
1192 // Was the old condition true if the operand is positive?
1193 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1195 // If so, the new one isn't.
1196 isTrueIfPositive ^= true;
1198 if (isTrueIfPositive)
1199 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1202 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1206 if (LHSI->hasOneUse()) {
1207 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1208 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1209 const APInt &SignBit = XorCst->getValue();
1210 ICmpInst::Predicate Pred = ICI.isSigned()
1211 ? ICI.getUnsignedPredicate()
1212 : ICI.getSignedPredicate();
1213 return new ICmpInst(Pred, LHSI->getOperand(0),
1214 Builder->getInt(RHSV ^ SignBit));
1217 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1218 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1219 const APInt &NotSignBit = XorCst->getValue();
1220 ICmpInst::Predicate Pred = ICI.isSigned()
1221 ? ICI.getUnsignedPredicate()
1222 : ICI.getSignedPredicate();
1223 Pred = ICI.getSwappedPredicate(Pred);
1224 return new ICmpInst(Pred, LHSI->getOperand(0),
1225 Builder->getInt(RHSV ^ NotSignBit));
1229 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1230 // iff -C is a power of 2
1231 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1232 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1233 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1235 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1236 // iff -C is a power of 2
1237 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1238 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1239 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1242 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1243 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1244 LHSI->getOperand(0)->hasOneUse()) {
1245 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1247 // If the LHS is an AND of a truncating cast, we can widen the
1248 // and/compare to be the input width without changing the value
1249 // produced, eliminating a cast.
1250 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1251 // We can do this transformation if either the AND constant does not
1252 // have its sign bit set or if it is an equality comparison.
1253 // Extending a relational comparison when we're checking the sign
1254 // bit would not work.
1255 if (ICI.isEquality() ||
1256 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1258 Builder->CreateAnd(Cast->getOperand(0),
1259 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1260 NewAnd->takeName(LHSI);
1261 return new ICmpInst(ICI.getPredicate(), NewAnd,
1262 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1266 // If the LHS is an AND of a zext, and we have an equality compare, we can
1267 // shrink the and/compare to the smaller type, eliminating the cast.
1268 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1269 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1270 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1271 // should fold the icmp to true/false in that case.
1272 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1274 Builder->CreateAnd(Cast->getOperand(0),
1275 ConstantExpr::getTrunc(AndCst, Ty));
1276 NewAnd->takeName(LHSI);
1277 return new ICmpInst(ICI.getPredicate(), NewAnd,
1278 ConstantExpr::getTrunc(RHS, Ty));
1282 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1283 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1284 // happens a LOT in code produced by the C front-end, for bitfield
1286 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1287 if (Shift && !Shift->isShift())
1291 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
1293 // This seemingly simple opportunity to fold away a shift turns out to
1294 // be rather complicated. See PR17827
1295 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1297 bool CanFold = false;
1298 unsigned ShiftOpcode = Shift->getOpcode();
1299 if (ShiftOpcode == Instruction::AShr) {
1300 // There may be some constraints that make this possible,
1301 // but nothing simple has been discovered yet.
1303 } else if (ShiftOpcode == Instruction::Shl) {
1304 // For a left shift, we can fold if the comparison is not signed.
1305 // We can also fold a signed comparison if the mask value and
1306 // comparison value are not negative. These constraints may not be
1307 // obvious, but we can prove that they are correct using an SMT
1309 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1311 } else if (ShiftOpcode == Instruction::LShr) {
1312 // For a logical right shift, we can fold if the comparison is not
1313 // signed. We can also fold a signed comparison if the shifted mask
1314 // value and the shifted comparison value are not negative.
1315 // These constraints may not be obvious, but we can prove that they
1316 // are correct using an SMT solver.
1317 if (!ICI.isSigned())
1320 ConstantInt *ShiftedAndCst =
1321 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1322 ConstantInt *ShiftedRHSCst =
1323 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1325 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1332 if (ShiftOpcode == Instruction::Shl)
1333 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1335 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1337 // Check to see if we are shifting out any of the bits being
1339 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1340 // If we shifted bits out, the fold is not going to work out.
1341 // As a special case, check to see if this means that the
1342 // result is always true or false now.
1343 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1344 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1345 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1346 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1348 ICI.setOperand(1, NewCst);
1349 Constant *NewAndCst;
1350 if (ShiftOpcode == Instruction::Shl)
1351 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1353 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1354 LHSI->setOperand(1, NewAndCst);
1355 LHSI->setOperand(0, Shift->getOperand(0));
1356 Worklist.Add(Shift); // Shift is dead.
1362 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1363 // preferable because it allows the C<<Y expression to be hoisted out
1364 // of a loop if Y is invariant and X is not.
1365 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1366 ICI.isEquality() && !Shift->isArithmeticShift() &&
1367 !isa<Constant>(Shift->getOperand(0))) {
1370 if (Shift->getOpcode() == Instruction::LShr) {
1371 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1373 // Insert a logical shift.
1374 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1377 // Compute X & (C << Y).
1379 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1381 ICI.setOperand(0, NewAnd);
1385 // (icmp pred (and (or (lshr X, Y), X), 1), 0) -->
1386 // (icmp pred (and X, (or (shl 1, Y), 1), 0))
1388 // iff pred isn't signed
1390 Value *X, *Y, *LShr;
1391 if (!ICI.isSigned() && RHSV == 0) {
1392 if (match(LHSI->getOperand(1), m_One())) {
1393 Constant *One = cast<Constant>(LHSI->getOperand(1));
1394 Value *Or = LHSI->getOperand(0);
1395 if (match(Or, m_Or(m_Value(LShr), m_Value(X))) &&
1396 match(LShr, m_LShr(m_Specific(X), m_Value(Y)))) {
1397 unsigned UsesRemoved = 0;
1398 if (LHSI->hasOneUse())
1400 if (Or->hasOneUse())
1402 if (LShr->hasOneUse())
1404 Value *NewOr = nullptr;
1405 // Compute X & ((1 << Y) | 1)
1406 if (auto *C = dyn_cast<Constant>(Y)) {
1407 if (UsesRemoved >= 1)
1409 ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1411 if (UsesRemoved >= 3)
1412 NewOr = Builder->CreateOr(Builder->CreateShl(One, Y,
1415 One, Or->getName());
1418 Value *NewAnd = Builder->CreateAnd(X, NewOr, LHSI->getName());
1419 ICI.setOperand(0, NewAnd);
1427 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1428 // bit set in (X & AndCst) will produce a result greater than RHSV.
1429 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1430 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1431 if ((NTZ < AndCst->getBitWidth()) &&
1432 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1433 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1434 Constant::getNullValue(RHS->getType()));
1438 // Try to optimize things like "A[i]&42 == 0" to index computations.
1439 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1440 if (GetElementPtrInst *GEP =
1441 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1442 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1443 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1444 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1445 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1446 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1451 // X & -C == -C -> X > u ~C
1452 // X & -C != -C -> X <= u ~C
1453 // iff C is a power of 2
1454 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1455 return new ICmpInst(
1456 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1457 : ICmpInst::ICMP_ULE,
1458 LHSI->getOperand(0), SubOne(RHS));
1461 case Instruction::Or: {
1462 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1465 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1466 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1467 // -> and (icmp eq P, null), (icmp eq Q, null).
1468 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1469 Constant::getNullValue(P->getType()));
1470 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1471 Constant::getNullValue(Q->getType()));
1473 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1474 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1476 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1482 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1483 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1486 // If this is a signed comparison to 0 and the mul is sign preserving,
1487 // use the mul LHS operand instead.
1488 ICmpInst::Predicate pred = ICI.getPredicate();
1489 if (isSignTest(pred, RHS) && !Val->isZero() &&
1490 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1491 return new ICmpInst(Val->isNegative() ?
1492 ICmpInst::getSwappedPredicate(pred) : pred,
1493 LHSI->getOperand(0),
1494 Constant::getNullValue(RHS->getType()));
1499 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1500 uint32_t TypeBits = RHSV.getBitWidth();
1501 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1504 // (1 << X) pred P2 -> X pred Log2(P2)
1505 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1506 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1507 ICmpInst::Predicate Pred = ICI.getPredicate();
1508 if (ICI.isUnsigned()) {
1509 if (!RHSVIsPowerOf2) {
1510 // (1 << X) < 30 -> X <= 4
1511 // (1 << X) <= 30 -> X <= 4
1512 // (1 << X) >= 30 -> X > 4
1513 // (1 << X) > 30 -> X > 4
1514 if (Pred == ICmpInst::ICMP_ULT)
1515 Pred = ICmpInst::ICMP_ULE;
1516 else if (Pred == ICmpInst::ICMP_UGE)
1517 Pred = ICmpInst::ICMP_UGT;
1519 unsigned RHSLog2 = RHSV.logBase2();
1521 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1522 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1523 if (RHSLog2 == TypeBits-1) {
1524 if (Pred == ICmpInst::ICMP_UGE)
1525 Pred = ICmpInst::ICMP_EQ;
1526 else if (Pred == ICmpInst::ICMP_ULT)
1527 Pred = ICmpInst::ICMP_NE;
1530 return new ICmpInst(Pred, X,
1531 ConstantInt::get(RHS->getType(), RHSLog2));
1532 } else if (ICI.isSigned()) {
1533 if (RHSV.isAllOnesValue()) {
1534 // (1 << X) <= -1 -> X == 31
1535 if (Pred == ICmpInst::ICMP_SLE)
1536 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1537 ConstantInt::get(RHS->getType(), TypeBits-1));
1539 // (1 << X) > -1 -> X != 31
1540 if (Pred == ICmpInst::ICMP_SGT)
1541 return new ICmpInst(ICmpInst::ICMP_NE, X,
1542 ConstantInt::get(RHS->getType(), TypeBits-1));
1544 // (1 << X) < 0 -> X == 31
1545 // (1 << X) <= 0 -> X == 31
1546 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1547 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1548 ConstantInt::get(RHS->getType(), TypeBits-1));
1550 // (1 << X) >= 0 -> X != 31
1551 // (1 << X) > 0 -> X != 31
1552 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1553 return new ICmpInst(ICmpInst::ICMP_NE, X,
1554 ConstantInt::get(RHS->getType(), TypeBits-1));
1556 } else if (ICI.isEquality()) {
1558 return new ICmpInst(
1559 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1565 // Check that the shift amount is in range. If not, don't perform
1566 // undefined shifts. When the shift is visited it will be
1568 if (ShAmt->uge(TypeBits))
1571 if (ICI.isEquality()) {
1572 // If we are comparing against bits always shifted out, the
1573 // comparison cannot succeed.
1575 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1577 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1578 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1579 Constant *Cst = Builder->getInt1(IsICMP_NE);
1580 return ReplaceInstUsesWith(ICI, Cst);
1583 // If the shift is NUW, then it is just shifting out zeros, no need for an
1585 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1586 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1587 ConstantExpr::getLShr(RHS, ShAmt));
1589 // If the shift is NSW and we compare to 0, then it is just shifting out
1590 // sign bits, no need for an AND either.
1591 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1592 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1593 ConstantExpr::getLShr(RHS, ShAmt));
1595 if (LHSI->hasOneUse()) {
1596 // Otherwise strength reduce the shift into an and.
1597 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1598 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1599 TypeBits - ShAmtVal));
1602 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1603 return new ICmpInst(ICI.getPredicate(), And,
1604 ConstantExpr::getLShr(RHS, ShAmt));
1608 // If this is a signed comparison to 0 and the shift is sign preserving,
1609 // use the shift LHS operand instead.
1610 ICmpInst::Predicate pred = ICI.getPredicate();
1611 if (isSignTest(pred, RHS) &&
1612 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1613 return new ICmpInst(pred,
1614 LHSI->getOperand(0),
1615 Constant::getNullValue(RHS->getType()));
1617 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1618 bool TrueIfSigned = false;
1619 if (LHSI->hasOneUse() &&
1620 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1621 // (X << 31) <s 0 --> (X&1) != 0
1622 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1623 APInt::getOneBitSet(TypeBits,
1624 TypeBits-ShAmt->getZExtValue()-1));
1626 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1627 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1628 And, Constant::getNullValue(And->getType()));
1631 // Transform (icmp pred iM (shl iM %v, N), CI)
1632 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1633 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1634 // This enables to get rid of the shift in favor of a trunc which can be
1635 // free on the target. It has the additional benefit of comparing to a
1636 // smaller constant, which will be target friendly.
1637 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1638 if (LHSI->hasOneUse() &&
1639 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1640 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1641 Constant *NCI = ConstantExpr::getTrunc(
1642 ConstantExpr::getAShr(RHS,
1643 ConstantInt::get(RHS->getType(), Amt)),
1645 return new ICmpInst(ICI.getPredicate(),
1646 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1653 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1654 case Instruction::AShr: {
1655 // Handle equality comparisons of shift-by-constant.
1656 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1657 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1658 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1662 // Handle exact shr's.
1663 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1664 if (RHSV.isMinValue())
1665 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1670 case Instruction::SDiv:
1671 case Instruction::UDiv:
1672 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1673 // Fold this div into the comparison, producing a range check.
1674 // Determine, based on the divide type, what the range is being
1675 // checked. If there is an overflow on the low or high side, remember
1676 // it, otherwise compute the range [low, hi) bounding the new value.
1677 // See: InsertRangeTest above for the kinds of replacements possible.
1678 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1679 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1684 case Instruction::Sub: {
1685 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1687 const APInt &LHSV = LHSC->getValue();
1689 // C1-X <u C2 -> (X|(C2-1)) == C1
1690 // iff C1 & (C2-1) == C2-1
1691 // C2 is a power of 2
1692 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1693 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1694 return new ICmpInst(ICmpInst::ICMP_EQ,
1695 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1698 // C1-X >u C2 -> (X|C2) != C1
1699 // iff C1 & C2 == C2
1700 // C2+1 is a power of 2
1701 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1702 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1703 return new ICmpInst(ICmpInst::ICMP_NE,
1704 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1708 case Instruction::Add:
1709 // Fold: icmp pred (add X, C1), C2
1710 if (!ICI.isEquality()) {
1711 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1713 const APInt &LHSV = LHSC->getValue();
1715 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1718 if (ICI.isSigned()) {
1719 if (CR.getLower().isSignBit()) {
1720 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1721 Builder->getInt(CR.getUpper()));
1722 } else if (CR.getUpper().isSignBit()) {
1723 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1724 Builder->getInt(CR.getLower()));
1727 if (CR.getLower().isMinValue()) {
1728 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1729 Builder->getInt(CR.getUpper()));
1730 } else if (CR.getUpper().isMinValue()) {
1731 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1732 Builder->getInt(CR.getLower()));
1736 // X-C1 <u C2 -> (X & -C2) == C1
1737 // iff C1 & (C2-1) == 0
1738 // C2 is a power of 2
1739 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1740 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1741 return new ICmpInst(ICmpInst::ICMP_EQ,
1742 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1743 ConstantExpr::getNeg(LHSC));
1745 // X-C1 >u C2 -> (X & ~C2) != C1
1747 // C2+1 is a power of 2
1748 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1749 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1750 return new ICmpInst(ICmpInst::ICMP_NE,
1751 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1752 ConstantExpr::getNeg(LHSC));
1757 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1758 if (ICI.isEquality()) {
1759 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1761 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1762 // the second operand is a constant, simplify a bit.
1763 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1764 switch (BO->getOpcode()) {
1765 case Instruction::SRem:
1766 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1767 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1768 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1769 if (V.sgt(1) && V.isPowerOf2()) {
1771 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1773 return new ICmpInst(ICI.getPredicate(), NewRem,
1774 Constant::getNullValue(BO->getType()));
1778 case Instruction::Add:
1779 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1780 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1781 if (BO->hasOneUse())
1782 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1783 ConstantExpr::getSub(RHS, BOp1C));
1784 } else if (RHSV == 0) {
1785 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1786 // efficiently invertible, or if the add has just this one use.
1787 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1789 if (Value *NegVal = dyn_castNegVal(BOp1))
1790 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1791 if (Value *NegVal = dyn_castNegVal(BOp0))
1792 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1793 if (BO->hasOneUse()) {
1794 Value *Neg = Builder->CreateNeg(BOp1);
1796 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1800 case Instruction::Xor:
1801 // For the xor case, we can xor two constants together, eliminating
1802 // the explicit xor.
1803 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1804 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1805 ConstantExpr::getXor(RHS, BOC));
1806 } else if (RHSV == 0) {
1807 // Replace ((xor A, B) != 0) with (A != B)
1808 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1812 case Instruction::Sub:
1813 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1814 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1815 if (BO->hasOneUse())
1816 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1817 ConstantExpr::getSub(BOp0C, RHS));
1818 } else if (RHSV == 0) {
1819 // Replace ((sub A, B) != 0) with (A != B)
1820 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1824 case Instruction::Or:
1825 // If bits are being or'd in that are not present in the constant we
1826 // are comparing against, then the comparison could never succeed!
1827 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1828 Constant *NotCI = ConstantExpr::getNot(RHS);
1829 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1830 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1834 case Instruction::And:
1835 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1836 // If bits are being compared against that are and'd out, then the
1837 // comparison can never succeed!
1838 if ((RHSV & ~BOC->getValue()) != 0)
1839 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1841 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1842 if (RHS == BOC && RHSV.isPowerOf2())
1843 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1844 ICmpInst::ICMP_NE, LHSI,
1845 Constant::getNullValue(RHS->getType()));
1847 // Don't perform the following transforms if the AND has multiple uses
1848 if (!BO->hasOneUse())
1851 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1852 if (BOC->getValue().isSignBit()) {
1853 Value *X = BO->getOperand(0);
1854 Constant *Zero = Constant::getNullValue(X->getType());
1855 ICmpInst::Predicate pred = isICMP_NE ?
1856 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1857 return new ICmpInst(pred, X, Zero);
1860 // ((X & ~7) == 0) --> X < 8
1861 if (RHSV == 0 && isHighOnes(BOC)) {
1862 Value *X = BO->getOperand(0);
1863 Constant *NegX = ConstantExpr::getNeg(BOC);
1864 ICmpInst::Predicate pred = isICMP_NE ?
1865 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1866 return new ICmpInst(pred, X, NegX);
1870 case Instruction::Mul:
1871 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1872 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1873 // The trivial case (mul X, 0) is handled by InstSimplify
1874 // General case : (mul X, C) != 0 iff X != 0
1875 // (mul X, C) == 0 iff X == 0
1877 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1878 Constant::getNullValue(RHS->getType()));
1884 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1885 // Handle icmp {eq|ne} <intrinsic>, intcst.
1886 switch (II->getIntrinsicID()) {
1887 case Intrinsic::bswap:
1889 ICI.setOperand(0, II->getArgOperand(0));
1890 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1892 case Intrinsic::ctlz:
1893 case Intrinsic::cttz:
1894 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1895 if (RHSV == RHS->getType()->getBitWidth()) {
1897 ICI.setOperand(0, II->getArgOperand(0));
1898 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1902 case Intrinsic::ctpop:
1903 // popcount(A) == 0 -> A == 0 and likewise for !=
1904 if (RHS->isZero()) {
1906 ICI.setOperand(0, II->getArgOperand(0));
1907 ICI.setOperand(1, RHS);
1919 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1920 /// We only handle extending casts so far.
1922 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1923 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1924 Value *LHSCIOp = LHSCI->getOperand(0);
1925 Type *SrcTy = LHSCIOp->getType();
1926 Type *DestTy = LHSCI->getType();
1929 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1930 // integer type is the same size as the pointer type.
1931 if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
1932 DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1933 Value *RHSOp = nullptr;
1934 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1935 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1936 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1937 RHSOp = RHSC->getOperand(0);
1938 // If the pointer types don't match, insert a bitcast.
1939 if (LHSCIOp->getType() != RHSOp->getType())
1940 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1944 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1947 // The code below only handles extension cast instructions, so far.
1949 if (LHSCI->getOpcode() != Instruction::ZExt &&
1950 LHSCI->getOpcode() != Instruction::SExt)
1953 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1954 bool isSignedCmp = ICI.isSigned();
1956 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1957 // Not an extension from the same type?
1958 RHSCIOp = CI->getOperand(0);
1959 if (RHSCIOp->getType() != LHSCIOp->getType())
1962 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1963 // and the other is a zext), then we can't handle this.
1964 if (CI->getOpcode() != LHSCI->getOpcode())
1967 // Deal with equality cases early.
1968 if (ICI.isEquality())
1969 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1971 // A signed comparison of sign extended values simplifies into a
1972 // signed comparison.
1973 if (isSignedCmp && isSignedExt)
1974 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1976 // The other three cases all fold into an unsigned comparison.
1977 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1980 // If we aren't dealing with a constant on the RHS, exit early
1981 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1985 // Compute the constant that would happen if we truncated to SrcTy then
1986 // reextended to DestTy.
1987 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1988 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1991 // If the re-extended constant didn't change...
1993 // Deal with equality cases early.
1994 if (ICI.isEquality())
1995 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1997 // A signed comparison of sign extended values simplifies into a
1998 // signed comparison.
1999 if (isSignedExt && isSignedCmp)
2000 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
2002 // The other three cases all fold into an unsigned comparison.
2003 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
2006 // The re-extended constant changed so the constant cannot be represented
2007 // in the shorter type. Consequently, we cannot emit a simple comparison.
2008 // All the cases that fold to true or false will have already been handled
2009 // by SimplifyICmpInst, so only deal with the tricky case.
2011 if (isSignedCmp || !isSignedExt)
2014 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
2015 // should have been folded away previously and not enter in here.
2017 // We're performing an unsigned comp with a sign extended value.
2018 // This is true if the input is >= 0. [aka >s -1]
2019 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
2020 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
2022 // Finally, return the value computed.
2023 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
2024 return ReplaceInstUsesWith(ICI, Result);
2026 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
2027 return BinaryOperator::CreateNot(Result);
2030 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
2031 /// I = icmp ugt (add (add A, B), CI2), CI1
2032 /// If this is of the form:
2034 /// if (sum+128 >u 255)
2035 /// Then replace it with llvm.sadd.with.overflow.i8.
2037 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
2038 ConstantInt *CI2, ConstantInt *CI1,
2040 // The transformation we're trying to do here is to transform this into an
2041 // llvm.sadd.with.overflow. To do this, we have to replace the original add
2042 // with a narrower add, and discard the add-with-constant that is part of the
2043 // range check (if we can't eliminate it, this isn't profitable).
2045 // In order to eliminate the add-with-constant, the compare can be its only
2047 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
2048 if (!AddWithCst->hasOneUse()) return nullptr;
2050 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
2051 if (!CI2->getValue().isPowerOf2()) return nullptr;
2052 unsigned NewWidth = CI2->getValue().countTrailingZeros();
2053 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
2055 // The width of the new add formed is 1 more than the bias.
2058 // Check to see that CI1 is an all-ones value with NewWidth bits.
2059 if (CI1->getBitWidth() == NewWidth ||
2060 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
2063 // This is only really a signed overflow check if the inputs have been
2064 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
2065 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
2066 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
2067 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
2068 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
2071 // In order to replace the original add with a narrower
2072 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
2073 // and truncates that discard the high bits of the add. Verify that this is
2075 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
2076 for (User *U : OrigAdd->users()) {
2077 if (U == AddWithCst) continue;
2079 // Only accept truncates for now. We would really like a nice recursive
2080 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
2081 // chain to see which bits of a value are actually demanded. If the
2082 // original add had another add which was then immediately truncated, we
2083 // could still do the transformation.
2084 TruncInst *TI = dyn_cast<TruncInst>(U);
2085 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
2089 // If the pattern matches, truncate the inputs to the narrower type and
2090 // use the sadd_with_overflow intrinsic to efficiently compute both the
2091 // result and the overflow bit.
2092 Module *M = I.getParent()->getParent()->getParent();
2094 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
2095 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
2098 InstCombiner::BuilderTy *Builder = IC.Builder;
2100 // Put the new code above the original add, in case there are any uses of the
2101 // add between the add and the compare.
2102 Builder->SetInsertPoint(OrigAdd);
2104 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
2105 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
2106 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
2107 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
2108 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
2110 // The inner add was the result of the narrow add, zero extended to the
2111 // wider type. Replace it with the result computed by the intrinsic.
2112 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
2114 // The original icmp gets replaced with the overflow value.
2115 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
2118 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
2120 // Don't bother doing this transformation for pointers, don't do it for
2122 if (!isa<IntegerType>(OrigAddV->getType())) return nullptr;
2124 // If the add is a constant expr, then we don't bother transforming it.
2125 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
2126 if (!OrigAdd) return nullptr;
2128 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
2130 // Put the new code above the original add, in case there are any uses of the
2131 // add between the add and the compare.
2132 InstCombiner::BuilderTy *Builder = IC.Builder;
2133 Builder->SetInsertPoint(OrigAdd);
2135 Module *M = I.getParent()->getParent()->getParent();
2136 Type *Ty = LHS->getType();
2137 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
2138 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
2139 Value *Add = Builder->CreateExtractValue(Call, 0);
2141 IC.ReplaceInstUsesWith(*OrigAdd, Add);
2143 // The original icmp gets replaced with the overflow value.
2144 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2147 /// \brief Recognize and process idiom involving test for multiplication
2150 /// The caller has matched a pattern of the form:
2151 /// I = cmp u (mul(zext A, zext B), V
2152 /// The function checks if this is a test for overflow and if so replaces
2153 /// multiplication with call to 'mul.with.overflow' intrinsic.
2155 /// \param I Compare instruction.
2156 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
2157 /// the compare instruction. Must be of integer type.
2158 /// \param OtherVal The other argument of compare instruction.
2159 /// \returns Instruction which must replace the compare instruction, NULL if no
2160 /// replacement required.
2161 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
2162 Value *OtherVal, InstCombiner &IC) {
2163 // Don't bother doing this transformation for pointers, don't do it for
2165 if (!isa<IntegerType>(MulVal->getType()))
2168 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
2169 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
2170 Instruction *MulInstr = cast<Instruction>(MulVal);
2171 assert(MulInstr->getOpcode() == Instruction::Mul);
2173 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
2174 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
2175 assert(LHS->getOpcode() == Instruction::ZExt);
2176 assert(RHS->getOpcode() == Instruction::ZExt);
2177 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
2179 // Calculate type and width of the result produced by mul.with.overflow.
2180 Type *TyA = A->getType(), *TyB = B->getType();
2181 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
2182 WidthB = TyB->getPrimitiveSizeInBits();
2185 if (WidthB > WidthA) {
2193 // In order to replace the original mul with a narrower mul.with.overflow,
2194 // all uses must ignore upper bits of the product. The number of used low
2195 // bits must be not greater than the width of mul.with.overflow.
2196 if (MulVal->hasNUsesOrMore(2))
2197 for (User *U : MulVal->users()) {
2200 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2201 // Check if truncation ignores bits above MulWidth.
2202 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
2203 if (TruncWidth > MulWidth)
2205 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2206 // Check if AND ignores bits above MulWidth.
2207 if (BO->getOpcode() != Instruction::And)
2209 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
2210 const APInt &CVal = CI->getValue();
2211 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
2215 // Other uses prohibit this transformation.
2220 // Recognize patterns
2221 switch (I.getPredicate()) {
2222 case ICmpInst::ICMP_EQ:
2223 case ICmpInst::ICMP_NE:
2224 // Recognize pattern:
2225 // mulval = mul(zext A, zext B)
2226 // cmp eq/neq mulval, zext trunc mulval
2227 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
2228 if (Zext->hasOneUse()) {
2229 Value *ZextArg = Zext->getOperand(0);
2230 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
2231 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
2235 // Recognize pattern:
2236 // mulval = mul(zext A, zext B)
2237 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
2240 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
2241 if (ValToMask != MulVal)
2243 const APInt &CVal = CI->getValue() + 1;
2244 if (CVal.isPowerOf2()) {
2245 unsigned MaskWidth = CVal.logBase2();
2246 if (MaskWidth == MulWidth)
2247 break; // Recognized
2252 case ICmpInst::ICMP_UGT:
2253 // Recognize pattern:
2254 // mulval = mul(zext A, zext B)
2255 // cmp ugt mulval, max
2256 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2257 APInt MaxVal = APInt::getMaxValue(MulWidth);
2258 MaxVal = MaxVal.zext(CI->getBitWidth());
2259 if (MaxVal.eq(CI->getValue()))
2260 break; // Recognized
2264 case ICmpInst::ICMP_UGE:
2265 // Recognize pattern:
2266 // mulval = mul(zext A, zext B)
2267 // cmp uge mulval, max+1
2268 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2269 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2270 if (MaxVal.eq(CI->getValue()))
2271 break; // Recognized
2275 case ICmpInst::ICMP_ULE:
2276 // Recognize pattern:
2277 // mulval = mul(zext A, zext B)
2278 // cmp ule mulval, max
2279 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2280 APInt MaxVal = APInt::getMaxValue(MulWidth);
2281 MaxVal = MaxVal.zext(CI->getBitWidth());
2282 if (MaxVal.eq(CI->getValue()))
2283 break; // Recognized
2287 case ICmpInst::ICMP_ULT:
2288 // Recognize pattern:
2289 // mulval = mul(zext A, zext B)
2290 // cmp ule mulval, max + 1
2291 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2292 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2293 if (MaxVal.eq(CI->getValue()))
2294 break; // Recognized
2302 InstCombiner::BuilderTy *Builder = IC.Builder;
2303 Builder->SetInsertPoint(MulInstr);
2304 Module *M = I.getParent()->getParent()->getParent();
2306 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
2307 Value *MulA = A, *MulB = B;
2308 if (WidthA < MulWidth)
2309 MulA = Builder->CreateZExt(A, MulType);
2310 if (WidthB < MulWidth)
2311 MulB = Builder->CreateZExt(B, MulType);
2313 Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
2314 CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
2315 IC.Worklist.Add(MulInstr);
2317 // If there are uses of mul result other than the comparison, we know that
2318 // they are truncation or binary AND. Change them to use result of
2319 // mul.with.overflow and adjust properly mask/size.
2320 if (MulVal->hasNUsesOrMore(2)) {
2321 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
2322 for (User *U : MulVal->users()) {
2323 if (U == &I || U == OtherVal)
2325 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2326 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
2327 IC.ReplaceInstUsesWith(*TI, Mul);
2329 TI->setOperand(0, Mul);
2330 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2331 assert(BO->getOpcode() == Instruction::And);
2332 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
2333 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
2334 APInt ShortMask = CI->getValue().trunc(MulWidth);
2335 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
2337 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
2338 IC.Worklist.Add(Zext);
2339 IC.ReplaceInstUsesWith(*BO, Zext);
2341 llvm_unreachable("Unexpected Binary operation");
2343 IC.Worklist.Add(cast<Instruction>(U));
2346 if (isa<Instruction>(OtherVal))
2347 IC.Worklist.Add(cast<Instruction>(OtherVal));
2349 // The original icmp gets replaced with the overflow value, maybe inverted
2350 // depending on predicate.
2351 bool Inverse = false;
2352 switch (I.getPredicate()) {
2353 case ICmpInst::ICMP_NE:
2355 case ICmpInst::ICMP_EQ:
2358 case ICmpInst::ICMP_UGT:
2359 case ICmpInst::ICMP_UGE:
2360 if (I.getOperand(0) == MulVal)
2364 case ICmpInst::ICMP_ULT:
2365 case ICmpInst::ICMP_ULE:
2366 if (I.getOperand(1) == MulVal)
2371 llvm_unreachable("Unexpected predicate");
2374 Value *Res = Builder->CreateExtractValue(Call, 1);
2375 return BinaryOperator::CreateNot(Res);
2378 return ExtractValueInst::Create(Call, 1);
2381 // DemandedBitsLHSMask - When performing a comparison against a constant,
2382 // it is possible that not all the bits in the LHS are demanded. This helper
2383 // method computes the mask that IS demanded.
2384 static APInt DemandedBitsLHSMask(ICmpInst &I,
2385 unsigned BitWidth, bool isSignCheck) {
2387 return APInt::getSignBit(BitWidth);
2389 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2390 if (!CI) return APInt::getAllOnesValue(BitWidth);
2391 const APInt &RHS = CI->getValue();
2393 switch (I.getPredicate()) {
2394 // For a UGT comparison, we don't care about any bits that
2395 // correspond to the trailing ones of the comparand. The value of these
2396 // bits doesn't impact the outcome of the comparison, because any value
2397 // greater than the RHS must differ in a bit higher than these due to carry.
2398 case ICmpInst::ICMP_UGT: {
2399 unsigned trailingOnes = RHS.countTrailingOnes();
2400 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2404 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2405 // Any value less than the RHS must differ in a higher bit because of carries.
2406 case ICmpInst::ICMP_ULT: {
2407 unsigned trailingZeros = RHS.countTrailingZeros();
2408 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2413 return APInt::getAllOnesValue(BitWidth);
2418 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2419 /// should be swapped.
2420 /// The decision is based on how many times these two operands are reused
2421 /// as subtract operands and their positions in those instructions.
2422 /// The rational is that several architectures use the same instruction for
2423 /// both subtract and cmp, thus it is better if the order of those operands
2425 /// \return true if Op0 and Op1 should be swapped.
2426 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2427 const Value * Op1) {
2428 // Filter out pointer value as those cannot appears directly in subtract.
2429 // FIXME: we may want to go through inttoptrs or bitcasts.
2430 if (Op0->getType()->isPointerTy())
2432 // Count every uses of both Op0 and Op1 in a subtract.
2433 // Each time Op0 is the first operand, count -1: swapping is bad, the
2434 // subtract has already the same layout as the compare.
2435 // Each time Op0 is the second operand, count +1: swapping is good, the
2436 // subtract has a different layout as the compare.
2437 // At the end, if the benefit is greater than 0, Op0 should come second to
2438 // expose more CSE opportunities.
2439 int GlobalSwapBenefits = 0;
2440 for (const User *U : Op0->users()) {
2441 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
2442 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2444 // If Op0 is the first argument, this is not beneficial to swap the
2446 int LocalSwapBenefits = -1;
2447 unsigned Op1Idx = 1;
2448 if (BinOp->getOperand(Op1Idx) == Op0) {
2450 LocalSwapBenefits = 1;
2452 if (BinOp->getOperand(Op1Idx) != Op1)
2454 GlobalSwapBenefits += LocalSwapBenefits;
2456 return GlobalSwapBenefits > 0;
2459 /// \brief Check that one use is in the same block as the definition and all
2460 /// other uses are in blocks dominated by a given block
2462 /// \param DI Definition
2464 /// \param DB Block that must dominate all uses of \p DI outside
2465 /// the parent block
2466 /// \return true when \p UI is the only use of \p DI in the parent block
2467 /// and all other uses of \p DI are in blocks dominated by \p DB.
2469 bool InstCombiner::dominatesAllUses(const Instruction *DI,
2470 const Instruction *UI,
2471 const BasicBlock *DB) const {
2472 assert(DI && UI && "Instruction not defined\n");
2473 // ignore incomplete definitions
2474 if (!DI->getParent())
2476 // DI and UI must be in the same block
2477 if (DI->getParent() != UI->getParent())
2479 // Protect from self-referencing blocks
2480 if (DI->getParent() == DB)
2482 // DominatorTree available?
2485 for (const User *U : DI->users()) {
2486 auto *Usr = cast<Instruction>(U);
2487 if (Usr != UI && !DT->dominates(DB, Usr->getParent()))
2494 /// true when the instruction sequence within a block is select-cmp-br.
2496 static bool isChainSelectCmpBranch(const SelectInst *SI) {
2497 const BasicBlock *BB = SI->getParent();
2500 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
2501 if (!BI || BI->getNumSuccessors() != 2)
2503 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
2504 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
2510 /// \brief True when a select result is replaced by one of its operands
2511 /// in select-icmp sequence. This will eventually result in the elimination
2514 /// \param SI Select instruction
2515 /// \param Icmp Compare instruction
2516 /// \param SIOpd Operand that replaces the select
2519 /// - The replacement is global and requires dominator information
2520 /// - The caller is responsible for the actual replacement
2525 /// %4 = select i1 %3, %C* %0, %C* null
2526 /// %5 = icmp eq %C* %4, null
2527 /// br i1 %5, label %9, label %7
2529 /// ; <label>:7 ; preds = %entry
2530 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
2533 /// can be transformed to
2535 /// %5 = icmp eq %C* %0, null
2536 /// %6 = select i1 %3, i1 %5, i1 true
2537 /// br i1 %6, label %9, label %7
2539 /// ; <label>:7 ; preds = %entry
2540 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
2542 /// Similar when the first operand of the select is a constant or/and
2543 /// the compare is for not equal rather than equal.
2545 /// NOTE: The function is only called when the select and compare constants
2546 /// are equal, the optimization can work only for EQ predicates. This is not a
2547 /// major restriction since a NE compare should be 'normalized' to an equal
2548 /// compare, which usually happens in the combiner and test case
2549 /// select-cmp-br.ll
2551 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
2552 const ICmpInst *Icmp,
2553 const unsigned SIOpd) {
2554 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
2555 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
2556 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
2557 // The check for the unique predecessor is not the best that can be
2558 // done. But it protects efficiently against cases like when SI's
2559 // home block has two successors, Succ and Succ1, and Succ1 predecessor
2560 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
2561 // replaced can be reached on either path. So the uniqueness check
2562 // guarantees that the path all uses of SI (outside SI's parent) are on
2563 // is disjoint from all other paths out of SI. But that information
2564 // is more expensive to compute, and the trade-off here is in favor
2566 if (Succ->getUniquePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
2568 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
2575 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2576 bool Changed = false;
2577 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2578 unsigned Op0Cplxity = getComplexity(Op0);
2579 unsigned Op1Cplxity = getComplexity(Op1);
2581 /// Orders the operands of the compare so that they are listed from most
2582 /// complex to least complex. This puts constants before unary operators,
2583 /// before binary operators.
2584 if (Op0Cplxity < Op1Cplxity ||
2585 (Op0Cplxity == Op1Cplxity &&
2586 swapMayExposeCSEOpportunities(Op0, Op1))) {
2588 std::swap(Op0, Op1);
2592 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AC))
2593 return ReplaceInstUsesWith(I, V);
2595 // comparing -val or val with non-zero is the same as just comparing val
2596 // ie, abs(val) != 0 -> val != 0
2597 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2599 Value *Cond, *SelectTrue, *SelectFalse;
2600 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2601 m_Value(SelectFalse)))) {
2602 if (Value *V = dyn_castNegVal(SelectTrue)) {
2603 if (V == SelectFalse)
2604 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2606 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2607 if (V == SelectTrue)
2608 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2613 Type *Ty = Op0->getType();
2615 // icmp's with boolean values can always be turned into bitwise operations
2616 if (Ty->isIntegerTy(1)) {
2617 switch (I.getPredicate()) {
2618 default: llvm_unreachable("Invalid icmp instruction!");
2619 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2620 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2621 return BinaryOperator::CreateNot(Xor);
2623 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2624 return BinaryOperator::CreateXor(Op0, Op1);
2626 case ICmpInst::ICMP_UGT:
2627 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2629 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2630 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2631 return BinaryOperator::CreateAnd(Not, Op1);
2633 case ICmpInst::ICMP_SGT:
2634 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2636 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2637 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2638 return BinaryOperator::CreateAnd(Not, Op0);
2640 case ICmpInst::ICMP_UGE:
2641 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2643 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2644 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2645 return BinaryOperator::CreateOr(Not, Op1);
2647 case ICmpInst::ICMP_SGE:
2648 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2650 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2651 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2652 return BinaryOperator::CreateOr(Not, Op0);
2657 unsigned BitWidth = 0;
2658 if (Ty->isIntOrIntVectorTy())
2659 BitWidth = Ty->getScalarSizeInBits();
2660 else if (DL) // Pointers require DL info to get their size.
2661 BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
2663 bool isSignBit = false;
2665 // See if we are doing a comparison with a constant.
2666 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2667 Value *A = nullptr, *B = nullptr;
2669 // Match the following pattern, which is a common idiom when writing
2670 // overflow-safe integer arithmetic function. The source performs an
2671 // addition in wider type, and explicitly checks for overflow using
2672 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2673 // sadd_with_overflow intrinsic.
2675 // TODO: This could probably be generalized to handle other overflow-safe
2676 // operations if we worked out the formulas to compute the appropriate
2680 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2682 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2683 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2684 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2685 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2689 // The following transforms are only 'worth it' if the only user of the
2690 // subtraction is the icmp.
2691 if (Op0->hasOneUse()) {
2692 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2693 if (I.isEquality() && CI->isZero() &&
2694 match(Op0, m_Sub(m_Value(A), m_Value(B))))
2695 return new ICmpInst(I.getPredicate(), A, B);
2697 // (icmp sgt (sub nsw A B), -1) -> (icmp sge A, B)
2698 if (I.getPredicate() == ICmpInst::ICMP_SGT && CI->isAllOnesValue() &&
2699 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2700 return new ICmpInst(ICmpInst::ICMP_SGE, A, B);
2702 // (icmp sgt (sub nsw A B), 0) -> (icmp sgt A, B)
2703 if (I.getPredicate() == ICmpInst::ICMP_SGT && CI->isZero() &&
2704 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2705 return new ICmpInst(ICmpInst::ICMP_SGT, A, B);
2707 // (icmp slt (sub nsw A B), 0) -> (icmp slt A, B)
2708 if (I.getPredicate() == ICmpInst::ICMP_SLT && CI->isZero() &&
2709 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2710 return new ICmpInst(ICmpInst::ICMP_SLT, A, B);
2712 // (icmp slt (sub nsw A B), 1) -> (icmp sle A, B)
2713 if (I.getPredicate() == ICmpInst::ICMP_SLT && CI->isOne() &&
2714 match(Op0, m_NSWSub(m_Value(A), m_Value(B))))
2715 return new ICmpInst(ICmpInst::ICMP_SLE, A, B);
2718 // If we have an icmp le or icmp ge instruction, turn it into the
2719 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2720 // them being folded in the code below. The SimplifyICmpInst code has
2721 // already handled the edge cases for us, so we just assert on them.
2722 switch (I.getPredicate()) {
2724 case ICmpInst::ICMP_ULE:
2725 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2726 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2727 Builder->getInt(CI->getValue()+1));
2728 case ICmpInst::ICMP_SLE:
2729 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2730 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2731 Builder->getInt(CI->getValue()+1));
2732 case ICmpInst::ICMP_UGE:
2733 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2734 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2735 Builder->getInt(CI->getValue()-1));
2736 case ICmpInst::ICMP_SGE:
2737 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2738 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2739 Builder->getInt(CI->getValue()-1));
2742 if (I.isEquality()) {
2744 if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
2745 match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
2746 // (icmp eq/ne (ashr/lshr const2, A), const1)
2747 if (Instruction *Inst = FoldICmpCstShrCst(I, Op0, A, CI, CI2))
2750 if (match(Op0, m_Shl(m_ConstantInt(CI2), m_Value(A)))) {
2751 // (icmp eq/ne (shl const2, A), const1)
2752 if (Instruction *Inst = FoldICmpCstShlCst(I, Op0, A, CI, CI2))
2757 // If this comparison is a normal comparison, it demands all
2758 // bits, if it is a sign bit comparison, it only demands the sign bit.
2760 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2763 // See if we can fold the comparison based on range information we can get
2764 // by checking whether bits are known to be zero or one in the input.
2765 if (BitWidth != 0) {
2766 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2767 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2769 if (SimplifyDemandedBits(I.getOperandUse(0),
2770 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2771 Op0KnownZero, Op0KnownOne, 0))
2773 if (SimplifyDemandedBits(I.getOperandUse(1),
2774 APInt::getAllOnesValue(BitWidth),
2775 Op1KnownZero, Op1KnownOne, 0))
2778 // Given the known and unknown bits, compute a range that the LHS could be
2779 // in. Compute the Min, Max and RHS values based on the known bits. For the
2780 // EQ and NE we use unsigned values.
2781 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2782 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2784 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2786 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2789 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2791 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2795 // If Min and Max are known to be the same, then SimplifyDemandedBits
2796 // figured out that the LHS is a constant. Just constant fold this now so
2797 // that code below can assume that Min != Max.
2798 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2799 return new ICmpInst(I.getPredicate(),
2800 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2801 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2802 return new ICmpInst(I.getPredicate(), Op0,
2803 ConstantInt::get(Op1->getType(), Op1Min));
2805 // Based on the range information we know about the LHS, see if we can
2806 // simplify this comparison. For example, (x&4) < 8 is always true.
2807 switch (I.getPredicate()) {
2808 default: llvm_unreachable("Unknown icmp opcode!");
2809 case ICmpInst::ICMP_EQ: {
2810 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2811 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2813 // If all bits are known zero except for one, then we know at most one
2814 // bit is set. If the comparison is against zero, then this is a check
2815 // to see if *that* bit is set.
2816 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2817 if (~Op1KnownZero == 0) {
2818 // If the LHS is an AND with the same constant, look through it.
2819 Value *LHS = nullptr;
2820 ConstantInt *LHSC = nullptr;
2821 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2822 LHSC->getValue() != Op0KnownZeroInverted)
2825 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2826 // then turn "((1 << x)&8) == 0" into "x != 3".
2827 // or turn "((1 << x)&7) == 0" into "x > 2".
2829 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2830 APInt ValToCheck = Op0KnownZeroInverted;
2831 if (ValToCheck.isPowerOf2()) {
2832 unsigned CmpVal = ValToCheck.countTrailingZeros();
2833 return new ICmpInst(ICmpInst::ICMP_NE, X,
2834 ConstantInt::get(X->getType(), CmpVal));
2835 } else if ((++ValToCheck).isPowerOf2()) {
2836 unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
2837 return new ICmpInst(ICmpInst::ICMP_UGT, X,
2838 ConstantInt::get(X->getType(), CmpVal));
2842 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2843 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2845 if (Op0KnownZeroInverted == 1 &&
2846 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2847 return new ICmpInst(ICmpInst::ICMP_NE, X,
2848 ConstantInt::get(X->getType(),
2849 CI->countTrailingZeros()));
2854 case ICmpInst::ICMP_NE: {
2855 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2856 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2858 // If all bits are known zero except for one, then we know at most one
2859 // bit is set. If the comparison is against zero, then this is a check
2860 // to see if *that* bit is set.
2861 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2862 if (~Op1KnownZero == 0) {
2863 // If the LHS is an AND with the same constant, look through it.
2864 Value *LHS = nullptr;
2865 ConstantInt *LHSC = nullptr;
2866 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2867 LHSC->getValue() != Op0KnownZeroInverted)
2870 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2871 // then turn "((1 << x)&8) != 0" into "x == 3".
2872 // or turn "((1 << x)&7) != 0" into "x < 3".
2874 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2875 APInt ValToCheck = Op0KnownZeroInverted;
2876 if (ValToCheck.isPowerOf2()) {
2877 unsigned CmpVal = ValToCheck.countTrailingZeros();
2878 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2879 ConstantInt::get(X->getType(), CmpVal));
2880 } else if ((++ValToCheck).isPowerOf2()) {
2881 unsigned CmpVal = ValToCheck.countTrailingZeros();
2882 return new ICmpInst(ICmpInst::ICMP_ULT, X,
2883 ConstantInt::get(X->getType(), CmpVal));
2887 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2888 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2890 if (Op0KnownZeroInverted == 1 &&
2891 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2892 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2893 ConstantInt::get(X->getType(),
2894 CI->countTrailingZeros()));
2899 case ICmpInst::ICMP_ULT:
2900 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2901 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2902 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2903 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2904 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2905 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2906 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2907 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2908 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2909 Builder->getInt(CI->getValue()-1));
2911 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2912 if (CI->isMinValue(true))
2913 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2914 Constant::getAllOnesValue(Op0->getType()));
2917 case ICmpInst::ICMP_UGT:
2918 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2919 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2920 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2921 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2923 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2924 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2925 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2926 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2927 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2928 Builder->getInt(CI->getValue()+1));
2930 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2931 if (CI->isMaxValue(true))
2932 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2933 Constant::getNullValue(Op0->getType()));
2936 case ICmpInst::ICMP_SLT:
2937 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2938 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2939 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2940 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2941 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2942 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2943 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2944 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2945 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2946 Builder->getInt(CI->getValue()-1));
2949 case ICmpInst::ICMP_SGT:
2950 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2951 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2952 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2953 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2955 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2956 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2957 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2958 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2959 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2960 Builder->getInt(CI->getValue()+1));
2963 case ICmpInst::ICMP_SGE:
2964 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2965 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2966 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2967 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2968 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2970 case ICmpInst::ICMP_SLE:
2971 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2972 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2973 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2974 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2975 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2977 case ICmpInst::ICMP_UGE:
2978 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2979 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2980 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2981 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2982 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2984 case ICmpInst::ICMP_ULE:
2985 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2986 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2987 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2988 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2989 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2993 // Turn a signed comparison into an unsigned one if both operands
2994 // are known to have the same sign.
2996 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2997 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2998 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
3001 // Test if the ICmpInst instruction is used exclusively by a select as
3002 // part of a minimum or maximum operation. If so, refrain from doing
3003 // any other folding. This helps out other analyses which understand
3004 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
3005 // and CodeGen. And in this case, at least one of the comparison
3006 // operands has at least one user besides the compare (the select),
3007 // which would often largely negate the benefit of folding anyway.
3009 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
3010 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
3011 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
3014 // See if we are doing a comparison between a constant and an instruction that
3015 // can be folded into the comparison.
3016 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3017 // Since the RHS is a ConstantInt (CI), if the left hand side is an
3018 // instruction, see if that instruction also has constants so that the
3019 // instruction can be folded into the icmp
3020 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3021 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
3025 // Handle icmp with constant (but not simple integer constant) RHS
3026 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3027 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3028 switch (LHSI->getOpcode()) {
3029 case Instruction::GetElementPtr:
3030 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3031 if (RHSC->isNullValue() &&
3032 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3033 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
3034 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3036 case Instruction::PHI:
3037 // Only fold icmp into the PHI if the phi and icmp are in the same
3038 // block. If in the same block, we're encouraging jump threading. If
3039 // not, we are just pessimizing the code by making an i1 phi.
3040 if (LHSI->getParent() == I.getParent())
3041 if (Instruction *NV = FoldOpIntoPhi(I))
3044 case Instruction::Select: {
3045 // If either operand of the select is a constant, we can fold the
3046 // comparison into the select arms, which will cause one to be
3047 // constant folded and the select turned into a bitwise or.
3048 Value *Op1 = nullptr, *Op2 = nullptr;
3049 ConstantInt *CI = 0;
3050 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3051 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3052 CI = dyn_cast<ConstantInt>(Op1);
3054 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3055 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3056 CI = dyn_cast<ConstantInt>(Op2);
3059 // We only want to perform this transformation if it will not lead to
3060 // additional code. This is true if either both sides of the select
3061 // fold to a constant (in which case the icmp is replaced with a select
3062 // which will usually simplify) or this is the only user of the
3063 // select (in which case we are trading a select+icmp for a simpler
3064 // select+icmp) or all uses of the select can be replaced based on
3065 // dominance information ("Global cases").
3066 bool Transform = false;
3069 else if (Op1 || Op2) {
3071 if (LHSI->hasOneUse())
3074 else if (CI && !CI->isZero())
3075 // When Op1 is constant try replacing select with second operand.
3076 // Otherwise Op2 is constant and try replacing select with first
3078 Transform = replacedSelectWithOperand(cast<SelectInst>(LHSI), &I,
3083 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
3086 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
3088 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3092 case Instruction::IntToPtr:
3093 // icmp pred inttoptr(X), null -> icmp pred X, 0
3094 if (RHSC->isNullValue() && DL &&
3095 DL->getIntPtrType(RHSC->getType()) ==
3096 LHSI->getOperand(0)->getType())
3097 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
3098 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3101 case Instruction::Load:
3102 // Try to optimize things like "A[i] > 4" to index computations.
3103 if (GetElementPtrInst *GEP =
3104 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3105 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3106 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3107 !cast<LoadInst>(LHSI)->isVolatile())
3108 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3115 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
3116 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
3117 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
3119 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
3120 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
3121 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
3124 // Test to see if the operands of the icmp are casted versions of other
3125 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
3127 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
3128 if (Op0->getType()->isPointerTy() &&
3129 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
3130 // We keep moving the cast from the left operand over to the right
3131 // operand, where it can often be eliminated completely.
3132 Op0 = CI->getOperand(0);
3134 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
3135 // so eliminate it as well.
3136 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
3137 Op1 = CI2->getOperand(0);
3139 // If Op1 is a constant, we can fold the cast into the constant.
3140 if (Op0->getType() != Op1->getType()) {
3141 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3142 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
3144 // Otherwise, cast the RHS right before the icmp
3145 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
3148 return new ICmpInst(I.getPredicate(), Op0, Op1);
3152 if (isa<CastInst>(Op0)) {
3153 // Handle the special case of: icmp (cast bool to X), <cst>
3154 // This comes up when you have code like
3157 // For generality, we handle any zero-extension of any operand comparison
3158 // with a constant or another cast from the same type.
3159 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
3160 if (Instruction *R = visitICmpInstWithCastAndCast(I))
3164 // Special logic for binary operators.
3165 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3166 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3168 CmpInst::Predicate Pred = I.getPredicate();
3169 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3170 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3171 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
3172 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3173 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3174 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3175 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
3176 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3177 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3179 // Analyze the case when either Op0 or Op1 is an add instruction.
3180 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3181 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3182 if (BO0 && BO0->getOpcode() == Instruction::Add)
3183 A = BO0->getOperand(0), B = BO0->getOperand(1);
3184 if (BO1 && BO1->getOpcode() == Instruction::Add)
3185 C = BO1->getOperand(0), D = BO1->getOperand(1);
3187 // icmp (X+cst) < 0 --> X < -cst
3188 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred) && match(Op1, m_Zero()))
3189 if (ConstantInt *RHSC = dyn_cast_or_null<ConstantInt>(B))
3190 if (!RHSC->isMinValue(/*isSigned=*/true))
3191 return new ICmpInst(Pred, A, ConstantExpr::getNeg(RHSC));
3193 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3194 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3195 return new ICmpInst(Pred, A == Op1 ? B : A,
3196 Constant::getNullValue(Op1->getType()));
3198 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3199 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3200 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3203 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3204 if (A && C && (A == C || A == D || B == C || B == D) &&
3205 NoOp0WrapProblem && NoOp1WrapProblem &&
3206 // Try not to increase register pressure.
3207 BO0->hasOneUse() && BO1->hasOneUse()) {
3208 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3211 // C + B == C + D -> B == D
3214 } else if (A == D) {
3215 // D + B == C + D -> B == C
3218 } else if (B == C) {
3219 // A + C == C + D -> A == D
3224 // A + D == C + D -> A == C
3228 return new ICmpInst(Pred, Y, Z);
3231 // icmp slt (X + -1), Y -> icmp sle X, Y
3232 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3233 match(B, m_AllOnes()))
3234 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3236 // icmp sge (X + -1), Y -> icmp sgt X, Y
3237 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3238 match(B, m_AllOnes()))
3239 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3241 // icmp sle (X + 1), Y -> icmp slt X, Y
3242 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
3244 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3246 // icmp sgt (X + 1), Y -> icmp sge X, Y
3247 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
3249 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3251 // if C1 has greater magnitude than C2:
3252 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3253 // s.t. C3 = C1 - C2
3255 // if C2 has greater magnitude than C1:
3256 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3257 // s.t. C3 = C2 - C1
3258 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3259 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3260 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3261 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3262 const APInt &AP1 = C1->getValue();
3263 const APInt &AP2 = C2->getValue();
3264 if (AP1.isNegative() == AP2.isNegative()) {
3265 APInt AP1Abs = C1->getValue().abs();
3266 APInt AP2Abs = C2->getValue().abs();
3267 if (AP1Abs.uge(AP2Abs)) {
3268 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
3269 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
3270 return new ICmpInst(Pred, NewAdd, C);
3272 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
3273 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
3274 return new ICmpInst(Pred, A, NewAdd);
3280 // Analyze the case when either Op0 or Op1 is a sub instruction.
3281 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3282 A = nullptr; B = nullptr; C = nullptr; D = nullptr;
3283 if (BO0 && BO0->getOpcode() == Instruction::Sub)
3284 A = BO0->getOperand(0), B = BO0->getOperand(1);
3285 if (BO1 && BO1->getOpcode() == Instruction::Sub)
3286 C = BO1->getOperand(0), D = BO1->getOperand(1);
3288 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3289 if (A == Op1 && NoOp0WrapProblem)
3290 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3292 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3293 if (C == Op0 && NoOp1WrapProblem)
3294 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3296 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3297 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3298 // Try not to increase register pressure.
3299 BO0->hasOneUse() && BO1->hasOneUse())
3300 return new ICmpInst(Pred, A, C);
3302 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3303 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3304 // Try not to increase register pressure.
3305 BO0->hasOneUse() && BO1->hasOneUse())
3306 return new ICmpInst(Pred, D, B);
3308 // icmp (0-X) < cst --> x > -cst
3309 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3311 if (match(BO0, m_Neg(m_Value(X))))
3312 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3313 if (!RHSC->isMinValue(/*isSigned=*/true))
3314 return new ICmpInst(I.getSwappedPredicate(), X,
3315 ConstantExpr::getNeg(RHSC));
3318 BinaryOperator *SRem = nullptr;
3319 // icmp (srem X, Y), Y
3320 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
3321 Op1 == BO0->getOperand(1))
3323 // icmp Y, (srem X, Y)
3324 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3325 Op0 == BO1->getOperand(1))
3328 // We don't check hasOneUse to avoid increasing register pressure because
3329 // the value we use is the same value this instruction was already using.
3330 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3332 case ICmpInst::ICMP_EQ:
3333 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3334 case ICmpInst::ICMP_NE:
3335 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3336 case ICmpInst::ICMP_SGT:
3337 case ICmpInst::ICMP_SGE:
3338 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3339 Constant::getAllOnesValue(SRem->getType()));
3340 case ICmpInst::ICMP_SLT:
3341 case ICmpInst::ICMP_SLE:
3342 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3343 Constant::getNullValue(SRem->getType()));
3347 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
3348 BO0->hasOneUse() && BO1->hasOneUse() &&
3349 BO0->getOperand(1) == BO1->getOperand(1)) {
3350 switch (BO0->getOpcode()) {
3352 case Instruction::Add:
3353 case Instruction::Sub:
3354 case Instruction::Xor:
3355 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3356 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3357 BO1->getOperand(0));
3358 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
3359 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3360 if (CI->getValue().isSignBit()) {
3361 ICmpInst::Predicate Pred = I.isSigned()
3362 ? I.getUnsignedPredicate()
3363 : I.getSignedPredicate();
3364 return new ICmpInst(Pred, BO0->getOperand(0),
3365 BO1->getOperand(0));
3368 if (CI->isMaxValue(true)) {
3369 ICmpInst::Predicate Pred = I.isSigned()
3370 ? I.getUnsignedPredicate()
3371 : I.getSignedPredicate();
3372 Pred = I.getSwappedPredicate(Pred);
3373 return new ICmpInst(Pred, BO0->getOperand(0),
3374 BO1->getOperand(0));
3378 case Instruction::Mul:
3379 if (!I.isEquality())
3382 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3383 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
3384 // Mask = -1 >> count-trailing-zeros(Cst).
3385 if (!CI->isZero() && !CI->isOne()) {
3386 const APInt &AP = CI->getValue();
3387 ConstantInt *Mask = ConstantInt::get(I.getContext(),
3388 APInt::getLowBitsSet(AP.getBitWidth(),
3390 AP.countTrailingZeros()));
3391 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
3392 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
3393 return new ICmpInst(I.getPredicate(), And1, And2);
3397 case Instruction::UDiv:
3398 case Instruction::LShr:
3402 case Instruction::SDiv:
3403 case Instruction::AShr:
3404 if (!BO0->isExact() || !BO1->isExact())
3406 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3407 BO1->getOperand(0));
3408 case Instruction::Shl: {
3409 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3410 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3413 if (!NSW && I.isSigned())
3415 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3416 BO1->getOperand(0));
3423 // Transform (A & ~B) == 0 --> (A & B) != 0
3424 // and (A & ~B) != 0 --> (A & B) == 0
3425 // if A is a power of 2.
3426 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
3427 match(Op1, m_Zero()) &&
3428 isKnownToBeAPowerOfTwo(A, false, 0, AC, &I, DT) && I.isEquality())
3429 return new ICmpInst(I.getInversePredicate(),
3430 Builder->CreateAnd(A, B),
3433 // ~x < ~y --> y < x
3434 // ~x < cst --> ~cst < x
3435 if (match(Op0, m_Not(m_Value(A)))) {
3436 if (match(Op1, m_Not(m_Value(B))))
3437 return new ICmpInst(I.getPredicate(), B, A);
3438 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3439 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
3442 // (a+b) <u a --> llvm.uadd.with.overflow.
3443 // (a+b) <u b --> llvm.uadd.with.overflow.
3444 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
3445 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
3446 (Op1 == A || Op1 == B))
3447 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
3450 // a >u (a+b) --> llvm.uadd.with.overflow.
3451 // b >u (a+b) --> llvm.uadd.with.overflow.
3452 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
3453 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
3454 (Op0 == A || Op0 == B))
3455 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
3458 // (zext a) * (zext b) --> llvm.umul.with.overflow.
3459 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3460 if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
3463 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3464 if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
3469 if (I.isEquality()) {
3470 Value *A, *B, *C, *D;
3472 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3473 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3474 Value *OtherVal = A == Op1 ? B : A;
3475 return new ICmpInst(I.getPredicate(), OtherVal,
3476 Constant::getNullValue(A->getType()));
3479 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3480 // A^c1 == C^c2 --> A == C^(c1^c2)
3481 ConstantInt *C1, *C2;
3482 if (match(B, m_ConstantInt(C1)) &&
3483 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
3484 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3485 Value *Xor = Builder->CreateXor(C, NC);
3486 return new ICmpInst(I.getPredicate(), A, Xor);
3489 // A^B == A^D -> B == D
3490 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
3491 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
3492 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
3493 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
3497 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
3498 (A == Op0 || B == Op0)) {
3499 // A == (A^B) -> B == 0
3500 Value *OtherVal = A == Op0 ? B : A;
3501 return new ICmpInst(I.getPredicate(), OtherVal,
3502 Constant::getNullValue(A->getType()));
3505 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3506 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3507 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3508 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3511 X = B; Y = D; Z = A;
3512 } else if (A == D) {
3513 X = B; Y = C; Z = A;
3514 } else if (B == C) {
3515 X = A; Y = D; Z = B;
3516 } else if (B == D) {
3517 X = A; Y = C; Z = B;
3520 if (X) { // Build (X^Y) & Z
3521 Op1 = Builder->CreateXor(X, Y);
3522 Op1 = Builder->CreateAnd(Op1, Z);
3523 I.setOperand(0, Op1);
3524 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3529 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3530 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3532 if ((Op0->hasOneUse() &&
3533 match(Op0, m_ZExt(m_Value(A))) &&
3534 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3535 (Op1->hasOneUse() &&
3536 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3537 match(Op1, m_ZExt(m_Value(A))))) {
3538 APInt Pow2 = Cst1->getValue() + 1;
3539 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3540 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3541 return new ICmpInst(I.getPredicate(), A,
3542 Builder->CreateTrunc(B, A->getType()));
3545 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3546 // For lshr and ashr pairs.
3547 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3548 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3549 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3550 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3551 unsigned TypeBits = Cst1->getBitWidth();
3552 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3553 if (ShAmt < TypeBits && ShAmt != 0) {
3554 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
3555 ? ICmpInst::ICMP_UGE
3556 : ICmpInst::ICMP_ULT;
3557 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3558 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3559 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
3563 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3564 // "icmp (and X, mask), cst"
3566 if (Op0->hasOneUse() &&
3567 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
3568 m_ConstantInt(ShAmt))))) &&
3569 match(Op1, m_ConstantInt(Cst1)) &&
3570 // Only do this when A has multiple uses. This is most important to do
3571 // when it exposes other optimizations.
3573 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3575 if (ShAmt < ASize) {
3577 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3580 APInt CmpV = Cst1->getValue().zext(ASize);
3583 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3584 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3589 // The 'cmpxchg' instruction returns an aggregate containing the old value and
3590 // an i1 which indicates whether or not we successfully did the swap.
3592 // Replace comparisons between the old value and the expected value with the
3593 // indicator that 'cmpxchg' returns.
3595 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
3596 // spuriously fail. In those cases, the old value may equal the expected
3597 // value but it is possible for the swap to not occur.
3598 if (I.getPredicate() == ICmpInst::ICMP_EQ)
3599 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
3600 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
3601 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
3603 return ExtractValueInst::Create(ACXI, 1);
3606 Value *X; ConstantInt *Cst;
3608 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3609 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3612 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3613 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3615 return Changed ? &I : nullptr;
3618 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3619 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3622 if (!isa<ConstantFP>(RHSC)) return nullptr;
3623 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3625 // Get the width of the mantissa. We don't want to hack on conversions that
3626 // might lose information from the integer, e.g. "i64 -> float"
3627 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3628 if (MantissaWidth == -1) return nullptr; // Unknown.
3630 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3632 // Check to see that the input is converted from an integer type that is small
3633 // enough that preserves all bits. TODO: check here for "known" sign bits.
3634 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3635 unsigned InputSize = IntTy->getScalarSizeInBits();
3637 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3638 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3642 if (I.isEquality()) {
3643 FCmpInst::Predicate P = I.getPredicate();
3644 bool IsExact = false;
3645 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
3646 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
3648 // If the floating point constant isn't an integer value, we know if we will
3649 // ever compare equal / not equal to it.
3651 // TODO: Can never be -0.0 and other non-representable values
3652 APFloat RHSRoundInt(RHS);
3653 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
3654 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
3655 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
3656 return ReplaceInstUsesWith(I, Builder->getFalse());
3658 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
3659 return ReplaceInstUsesWith(I, Builder->getTrue());
3663 // TODO: If the constant is exactly representable, is it always OK to do
3664 // equality compares as integer?
3667 // Comparisons with zero are a special case where we know we won't lose
3669 bool IsCmpZero = RHS.isPosZero();
3671 // If the conversion would lose info, don't hack on this.
3672 if ((int)InputSize > MantissaWidth && !IsCmpZero)
3675 // Otherwise, we can potentially simplify the comparison. We know that it
3676 // will always come through as an integer value and we know the constant is
3677 // not a NAN (it would have been previously simplified).
3678 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3680 ICmpInst::Predicate Pred;
3681 switch (I.getPredicate()) {
3682 default: llvm_unreachable("Unexpected predicate!");
3683 case FCmpInst::FCMP_UEQ:
3684 case FCmpInst::FCMP_OEQ:
3685 Pred = ICmpInst::ICMP_EQ;
3687 case FCmpInst::FCMP_UGT:
3688 case FCmpInst::FCMP_OGT:
3689 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3691 case FCmpInst::FCMP_UGE:
3692 case FCmpInst::FCMP_OGE:
3693 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3695 case FCmpInst::FCMP_ULT:
3696 case FCmpInst::FCMP_OLT:
3697 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3699 case FCmpInst::FCMP_ULE:
3700 case FCmpInst::FCMP_OLE:
3701 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3703 case FCmpInst::FCMP_UNE:
3704 case FCmpInst::FCMP_ONE:
3705 Pred = ICmpInst::ICMP_NE;
3707 case FCmpInst::FCMP_ORD:
3708 return ReplaceInstUsesWith(I, Builder->getTrue());
3709 case FCmpInst::FCMP_UNO:
3710 return ReplaceInstUsesWith(I, Builder->getFalse());
3713 // Now we know that the APFloat is a normal number, zero or inf.
3715 // See if the FP constant is too large for the integer. For example,
3716 // comparing an i8 to 300.0.
3717 unsigned IntWidth = IntTy->getScalarSizeInBits();
3720 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3721 // and large values.
3722 APFloat SMax(RHS.getSemantics());
3723 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3724 APFloat::rmNearestTiesToEven);
3725 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3726 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3727 Pred == ICmpInst::ICMP_SLE)
3728 return ReplaceInstUsesWith(I, Builder->getTrue());
3729 return ReplaceInstUsesWith(I, Builder->getFalse());
3732 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3733 // +INF and large values.
3734 APFloat UMax(RHS.getSemantics());
3735 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3736 APFloat::rmNearestTiesToEven);
3737 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3738 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3739 Pred == ICmpInst::ICMP_ULE)
3740 return ReplaceInstUsesWith(I, Builder->getTrue());
3741 return ReplaceInstUsesWith(I, Builder->getFalse());
3746 // See if the RHS value is < SignedMin.
3747 APFloat SMin(RHS.getSemantics());
3748 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3749 APFloat::rmNearestTiesToEven);
3750 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3751 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3752 Pred == ICmpInst::ICMP_SGE)
3753 return ReplaceInstUsesWith(I, Builder->getTrue());
3754 return ReplaceInstUsesWith(I, Builder->getFalse());
3757 // See if the RHS value is < UnsignedMin.
3758 APFloat SMin(RHS.getSemantics());
3759 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3760 APFloat::rmNearestTiesToEven);
3761 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3762 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3763 Pred == ICmpInst::ICMP_UGE)
3764 return ReplaceInstUsesWith(I, Builder->getTrue());
3765 return ReplaceInstUsesWith(I, Builder->getFalse());
3769 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3770 // [0, UMAX], but it may still be fractional. See if it is fractional by
3771 // casting the FP value to the integer value and back, checking for equality.
3772 // Don't do this for zero, because -0.0 is not fractional.
3773 Constant *RHSInt = LHSUnsigned
3774 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3775 : ConstantExpr::getFPToSI(RHSC, IntTy);
3776 if (!RHS.isZero()) {
3777 bool Equal = LHSUnsigned
3778 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3779 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3781 // If we had a comparison against a fractional value, we have to adjust
3782 // the compare predicate and sometimes the value. RHSC is rounded towards
3783 // zero at this point.
3785 default: llvm_unreachable("Unexpected integer comparison!");
3786 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3787 return ReplaceInstUsesWith(I, Builder->getTrue());
3788 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3789 return ReplaceInstUsesWith(I, Builder->getFalse());
3790 case ICmpInst::ICMP_ULE:
3791 // (float)int <= 4.4 --> int <= 4
3792 // (float)int <= -4.4 --> false
3793 if (RHS.isNegative())
3794 return ReplaceInstUsesWith(I, Builder->getFalse());
3796 case ICmpInst::ICMP_SLE:
3797 // (float)int <= 4.4 --> int <= 4
3798 // (float)int <= -4.4 --> int < -4
3799 if (RHS.isNegative())
3800 Pred = ICmpInst::ICMP_SLT;
3802 case ICmpInst::ICMP_ULT:
3803 // (float)int < -4.4 --> false
3804 // (float)int < 4.4 --> int <= 4
3805 if (RHS.isNegative())
3806 return ReplaceInstUsesWith(I, Builder->getFalse());
3807 Pred = ICmpInst::ICMP_ULE;
3809 case ICmpInst::ICMP_SLT:
3810 // (float)int < -4.4 --> int < -4
3811 // (float)int < 4.4 --> int <= 4
3812 if (!RHS.isNegative())
3813 Pred = ICmpInst::ICMP_SLE;
3815 case ICmpInst::ICMP_UGT:
3816 // (float)int > 4.4 --> int > 4
3817 // (float)int > -4.4 --> true
3818 if (RHS.isNegative())
3819 return ReplaceInstUsesWith(I, Builder->getTrue());
3821 case ICmpInst::ICMP_SGT:
3822 // (float)int > 4.4 --> int > 4
3823 // (float)int > -4.4 --> int >= -4
3824 if (RHS.isNegative())
3825 Pred = ICmpInst::ICMP_SGE;
3827 case ICmpInst::ICMP_UGE:
3828 // (float)int >= -4.4 --> true
3829 // (float)int >= 4.4 --> int > 4
3830 if (RHS.isNegative())
3831 return ReplaceInstUsesWith(I, Builder->getTrue());
3832 Pred = ICmpInst::ICMP_UGT;
3834 case ICmpInst::ICMP_SGE:
3835 // (float)int >= -4.4 --> int >= -4
3836 // (float)int >= 4.4 --> int > 4
3837 if (!RHS.isNegative())
3838 Pred = ICmpInst::ICMP_SGT;
3844 // Lower this FP comparison into an appropriate integer version of the
3846 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3849 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3850 bool Changed = false;
3852 /// Orders the operands of the compare so that they are listed from most
3853 /// complex to least complex. This puts constants before unary operators,
3854 /// before binary operators.
3855 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3860 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3862 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AC))
3863 return ReplaceInstUsesWith(I, V);
3865 // Simplify 'fcmp pred X, X'
3867 switch (I.getPredicate()) {
3868 default: llvm_unreachable("Unknown predicate!");
3869 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3870 case FCmpInst::FCMP_ULT: // True if unordered or less than
3871 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3872 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3873 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3874 I.setPredicate(FCmpInst::FCMP_UNO);
3875 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3878 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3879 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3880 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3881 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3882 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3883 I.setPredicate(FCmpInst::FCMP_ORD);
3884 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3889 // Handle fcmp with constant RHS
3890 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3891 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3892 switch (LHSI->getOpcode()) {
3893 case Instruction::FPExt: {
3894 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3895 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3896 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3900 const fltSemantics *Sem;
3901 // FIXME: This shouldn't be here.
3902 if (LHSExt->getSrcTy()->isHalfTy())
3903 Sem = &APFloat::IEEEhalf;
3904 else if (LHSExt->getSrcTy()->isFloatTy())
3905 Sem = &APFloat::IEEEsingle;
3906 else if (LHSExt->getSrcTy()->isDoubleTy())
3907 Sem = &APFloat::IEEEdouble;
3908 else if (LHSExt->getSrcTy()->isFP128Ty())
3909 Sem = &APFloat::IEEEquad;
3910 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3911 Sem = &APFloat::x87DoubleExtended;
3912 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3913 Sem = &APFloat::PPCDoubleDouble;
3918 APFloat F = RHSF->getValueAPF();
3919 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3921 // Avoid lossy conversions and denormals. Zero is a special case
3922 // that's OK to convert.
3926 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3927 APFloat::cmpLessThan) || Fabs.isZero()))
3929 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3930 ConstantFP::get(RHSC->getContext(), F));
3933 case Instruction::PHI:
3934 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3935 // block. If in the same block, we're encouraging jump threading. If
3936 // not, we are just pessimizing the code by making an i1 phi.
3937 if (LHSI->getParent() == I.getParent())
3938 if (Instruction *NV = FoldOpIntoPhi(I))
3941 case Instruction::SIToFP:
3942 case Instruction::UIToFP:
3943 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3946 case Instruction::FSub: {
3947 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3949 if (match(LHSI, m_FNeg(m_Value(Op))))
3950 return new FCmpInst(I.getSwappedPredicate(), Op,
3951 ConstantExpr::getFNeg(RHSC));
3954 case Instruction::Load:
3955 if (GetElementPtrInst *GEP =
3956 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3957 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3958 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3959 !cast<LoadInst>(LHSI)->isVolatile())
3960 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3964 case Instruction::Call: {
3965 if (!RHSC->isNullValue())
3968 CallInst *CI = cast<CallInst>(LHSI);
3969 const Function *F = CI->getCalledFunction();
3973 // Various optimization for fabs compared with zero.
3975 if (F->getIntrinsicID() == Intrinsic::fabs ||
3976 (TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
3977 (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3978 Func == LibFunc::fabsl))) {
3979 switch (I.getPredicate()) {
3982 // fabs(x) < 0 --> false
3983 case FCmpInst::FCMP_OLT:
3984 return ReplaceInstUsesWith(I, Builder->getFalse());
3985 // fabs(x) > 0 --> x != 0
3986 case FCmpInst::FCMP_OGT:
3987 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), RHSC);
3988 // fabs(x) <= 0 --> x == 0
3989 case FCmpInst::FCMP_OLE:
3990 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), RHSC);
3991 // fabs(x) >= 0 --> !isnan(x)
3992 case FCmpInst::FCMP_OGE:
3993 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), RHSC);
3994 // fabs(x) == 0 --> x == 0
3995 // fabs(x) != 0 --> x != 0
3996 case FCmpInst::FCMP_OEQ:
3997 case FCmpInst::FCMP_UEQ:
3998 case FCmpInst::FCMP_ONE:
3999 case FCmpInst::FCMP_UNE:
4000 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), RHSC);
4007 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
4009 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
4010 return new FCmpInst(I.getSwappedPredicate(), X, Y);
4012 // fcmp (fpext x), (fpext y) -> fcmp x, y
4013 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
4014 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
4015 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
4016 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
4017 RHSExt->getOperand(0));
4019 return Changed ? &I : nullptr;