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 "InstCombine.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/IR/ConstantRange.h"
19 #include "llvm/IR/DataLayout.h"
20 #include "llvm/IR/GetElementPtrTypeIterator.h"
21 #include "llvm/IR/IntrinsicInst.h"
22 #include "llvm/IR/PatternMatch.h"
23 #include "llvm/Target/TargetLibraryInfo.h"
25 using namespace PatternMatch;
27 #define DEBUG_TYPE "instcombine"
29 static ConstantInt *getOne(Constant *C) {
30 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
33 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
34 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
37 static bool HasAddOverflow(ConstantInt *Result,
38 ConstantInt *In1, ConstantInt *In2,
41 return Result->getValue().ult(In1->getValue());
43 if (In2->isNegative())
44 return Result->getValue().sgt(In1->getValue());
45 return Result->getValue().slt(In1->getValue());
48 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
49 /// overflowed for this type.
50 static bool AddWithOverflow(Constant *&Result, Constant *In1,
51 Constant *In2, bool IsSigned = false) {
52 Result = ConstantExpr::getAdd(In1, In2);
54 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
55 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
56 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
57 if (HasAddOverflow(ExtractElement(Result, Idx),
58 ExtractElement(In1, Idx),
59 ExtractElement(In2, Idx),
66 return HasAddOverflow(cast<ConstantInt>(Result),
67 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
71 static bool HasSubOverflow(ConstantInt *Result,
72 ConstantInt *In1, ConstantInt *In2,
75 return Result->getValue().ugt(In1->getValue());
77 if (In2->isNegative())
78 return Result->getValue().slt(In1->getValue());
80 return Result->getValue().sgt(In1->getValue());
83 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
84 /// overflowed for this type.
85 static bool SubWithOverflow(Constant *&Result, Constant *In1,
86 Constant *In2, bool IsSigned = false) {
87 Result = ConstantExpr::getSub(In1, In2);
89 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
90 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
91 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
92 if (HasSubOverflow(ExtractElement(Result, Idx),
93 ExtractElement(In1, Idx),
94 ExtractElement(In2, Idx),
101 return HasSubOverflow(cast<ConstantInt>(Result),
102 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
106 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
107 /// comparison only checks the sign bit. If it only checks the sign bit, set
108 /// TrueIfSigned if the result of the comparison is true when the input value is
110 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
111 bool &TrueIfSigned) {
113 case ICmpInst::ICMP_SLT: // True if LHS s< 0
115 return RHS->isZero();
116 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
118 return RHS->isAllOnesValue();
119 case ICmpInst::ICMP_SGT: // True if LHS s> -1
120 TrueIfSigned = false;
121 return RHS->isAllOnesValue();
122 case ICmpInst::ICMP_UGT:
123 // True if LHS u> RHS and RHS == high-bit-mask - 1
125 return RHS->isMaxValue(true);
126 case ICmpInst::ICMP_UGE:
127 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
129 return RHS->getValue().isSignBit();
135 /// Returns true if the exploded icmp can be expressed as a signed comparison
136 /// to zero and updates the predicate accordingly.
137 /// The signedness of the comparison is preserved.
138 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
139 if (!ICmpInst::isSigned(pred))
143 return ICmpInst::isRelational(pred);
146 if (pred == ICmpInst::ICMP_SLT) {
147 pred = ICmpInst::ICMP_SLE;
150 } else if (RHS->isAllOnesValue()) {
151 if (pred == ICmpInst::ICMP_SGT) {
152 pred = ICmpInst::ICMP_SGE;
160 // isHighOnes - Return true if the constant is of the form 1+0+.
161 // This is the same as lowones(~X).
162 static bool isHighOnes(const ConstantInt *CI) {
163 return (~CI->getValue() + 1).isPowerOf2();
166 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
167 /// set of known zero and one bits, compute the maximum and minimum values that
168 /// could have the specified known zero and known one bits, returning them in
170 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
171 const APInt& KnownOne,
172 APInt& Min, APInt& Max) {
173 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
174 KnownZero.getBitWidth() == Min.getBitWidth() &&
175 KnownZero.getBitWidth() == Max.getBitWidth() &&
176 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
177 APInt UnknownBits = ~(KnownZero|KnownOne);
179 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
180 // bit if it is unknown.
182 Max = KnownOne|UnknownBits;
184 if (UnknownBits.isNegative()) { // Sign bit is unknown
185 Min.setBit(Min.getBitWidth()-1);
186 Max.clearBit(Max.getBitWidth()-1);
190 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
191 // a set of known zero and one bits, compute the maximum and minimum values that
192 // could have the specified known zero and known one bits, returning them in
194 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
195 const APInt &KnownOne,
196 APInt &Min, APInt &Max) {
197 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
198 KnownZero.getBitWidth() == Min.getBitWidth() &&
199 KnownZero.getBitWidth() == Max.getBitWidth() &&
200 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
201 APInt UnknownBits = ~(KnownZero|KnownOne);
203 // The minimum value is when the unknown bits are all zeros.
205 // The maximum value is when the unknown bits are all ones.
206 Max = KnownOne|UnknownBits;
211 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
212 /// cmp pred (load (gep GV, ...)), cmpcst
213 /// where GV is a global variable with a constant initializer. Try to simplify
214 /// this into some simple computation that does not need the load. For example
215 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
217 /// If AndCst is non-null, then the loaded value is masked with that constant
218 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
219 Instruction *InstCombiner::
220 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
221 CmpInst &ICI, ConstantInt *AndCst) {
222 // We need TD information to know the pointer size unless this is inbounds.
223 if (!GEP->isInBounds() && !DL)
226 Constant *Init = GV->getInitializer();
227 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
230 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
231 if (ArrayElementCount > 1024) return nullptr; // Don't blow up on huge arrays.
233 // There are many forms of this optimization we can handle, for now, just do
234 // the simple index into a single-dimensional array.
236 // Require: GEP GV, 0, i {{, constant indices}}
237 if (GEP->getNumOperands() < 3 ||
238 !isa<ConstantInt>(GEP->getOperand(1)) ||
239 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
240 isa<Constant>(GEP->getOperand(2)))
243 // Check that indices after the variable are constants and in-range for the
244 // type they index. Collect the indices. This is typically for arrays of
246 SmallVector<unsigned, 4> LaterIndices;
248 Type *EltTy = Init->getType()->getArrayElementType();
249 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
250 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
251 if (!Idx) return nullptr; // Variable index.
253 uint64_t IdxVal = Idx->getZExtValue();
254 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
256 if (StructType *STy = dyn_cast<StructType>(EltTy))
257 EltTy = STy->getElementType(IdxVal);
258 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
259 if (IdxVal >= ATy->getNumElements()) return nullptr;
260 EltTy = ATy->getElementType();
262 return nullptr; // Unknown type.
265 LaterIndices.push_back(IdxVal);
268 enum { Overdefined = -3, Undefined = -2 };
270 // Variables for our state machines.
272 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
273 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
274 // and 87 is the second (and last) index. FirstTrueElement is -2 when
275 // undefined, otherwise set to the first true element. SecondTrueElement is
276 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
277 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
279 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
280 // form "i != 47 & i != 87". Same state transitions as for true elements.
281 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
283 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
284 /// define a state machine that triggers for ranges of values that the index
285 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
286 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
287 /// index in the range (inclusive). We use -2 for undefined here because we
288 /// use relative comparisons and don't want 0-1 to match -1.
289 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
291 // MagicBitvector - This is a magic bitvector where we set a bit if the
292 // comparison is true for element 'i'. If there are 64 elements or less in
293 // the array, this will fully represent all the comparison results.
294 uint64_t MagicBitvector = 0;
297 // Scan the array and see if one of our patterns matches.
298 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
299 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
300 Constant *Elt = Init->getAggregateElement(i);
301 if (!Elt) return nullptr;
303 // If this is indexing an array of structures, get the structure element.
304 if (!LaterIndices.empty())
305 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
307 // If the element is masked, handle it.
308 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
310 // Find out if the comparison would be true or false for the i'th element.
311 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
312 CompareRHS, DL, TLI);
313 // If the result is undef for this element, ignore it.
314 if (isa<UndefValue>(C)) {
315 // Extend range state machines to cover this element in case there is an
316 // undef in the middle of the range.
317 if (TrueRangeEnd == (int)i-1)
319 if (FalseRangeEnd == (int)i-1)
324 // If we can't compute the result for any of the elements, we have to give
325 // up evaluating the entire conditional.
326 if (!isa<ConstantInt>(C)) return nullptr;
328 // Otherwise, we know if the comparison is true or false for this element,
329 // update our state machines.
330 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
332 // State machine for single/double/range index comparison.
334 // Update the TrueElement state machine.
335 if (FirstTrueElement == Undefined)
336 FirstTrueElement = TrueRangeEnd = i; // First true element.
338 // Update double-compare state machine.
339 if (SecondTrueElement == Undefined)
340 SecondTrueElement = i;
342 SecondTrueElement = Overdefined;
344 // Update range state machine.
345 if (TrueRangeEnd == (int)i-1)
348 TrueRangeEnd = Overdefined;
351 // Update the FalseElement state machine.
352 if (FirstFalseElement == Undefined)
353 FirstFalseElement = FalseRangeEnd = i; // First false element.
355 // Update double-compare state machine.
356 if (SecondFalseElement == Undefined)
357 SecondFalseElement = i;
359 SecondFalseElement = Overdefined;
361 // Update range state machine.
362 if (FalseRangeEnd == (int)i-1)
365 FalseRangeEnd = Overdefined;
370 // If this element is in range, update our magic bitvector.
371 if (i < 64 && IsTrueForElt)
372 MagicBitvector |= 1ULL << i;
374 // If all of our states become overdefined, bail out early. Since the
375 // predicate is expensive, only check it every 8 elements. This is only
376 // really useful for really huge arrays.
377 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
378 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
379 FalseRangeEnd == Overdefined)
383 // Now that we've scanned the entire array, emit our new comparison(s). We
384 // order the state machines in complexity of the generated code.
385 Value *Idx = GEP->getOperand(2);
387 // If the index is larger than the pointer size of the target, truncate the
388 // index down like the GEP would do implicitly. We don't have to do this for
389 // an inbounds GEP because the index can't be out of range.
390 if (!GEP->isInBounds()) {
391 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
392 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
393 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
394 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
397 // If the comparison is only true for one or two elements, emit direct
399 if (SecondTrueElement != Overdefined) {
400 // None true -> false.
401 if (FirstTrueElement == Undefined)
402 return ReplaceInstUsesWith(ICI, Builder->getFalse());
404 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
406 // True for one element -> 'i == 47'.
407 if (SecondTrueElement == Undefined)
408 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
410 // True for two elements -> 'i == 47 | i == 72'.
411 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
412 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
413 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
414 return BinaryOperator::CreateOr(C1, C2);
417 // If the comparison is only false for one or two elements, emit direct
419 if (SecondFalseElement != Overdefined) {
420 // None false -> true.
421 if (FirstFalseElement == Undefined)
422 return ReplaceInstUsesWith(ICI, Builder->getTrue());
424 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
426 // False for one element -> 'i != 47'.
427 if (SecondFalseElement == Undefined)
428 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
430 // False for two elements -> 'i != 47 & i != 72'.
431 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
432 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
433 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
434 return BinaryOperator::CreateAnd(C1, C2);
437 // If the comparison can be replaced with a range comparison for the elements
438 // where it is true, emit the range check.
439 if (TrueRangeEnd != Overdefined) {
440 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
442 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
443 if (FirstTrueElement) {
444 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
445 Idx = Builder->CreateAdd(Idx, Offs);
448 Value *End = ConstantInt::get(Idx->getType(),
449 TrueRangeEnd-FirstTrueElement+1);
450 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
453 // False range check.
454 if (FalseRangeEnd != Overdefined) {
455 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
456 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
457 if (FirstFalseElement) {
458 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
459 Idx = Builder->CreateAdd(Idx, Offs);
462 Value *End = ConstantInt::get(Idx->getType(),
463 FalseRangeEnd-FirstFalseElement);
464 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
468 // If a magic bitvector captures the entire comparison state
469 // of this load, replace it with computation that does:
470 // ((magic_cst >> i) & 1) != 0
474 // Look for an appropriate type:
475 // - The type of Idx if the magic fits
476 // - The smallest fitting legal type if we have a DataLayout
478 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
481 Ty = DL->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
482 else if (ArrayElementCount <= 32)
483 Ty = Type::getInt32Ty(Init->getContext());
486 Value *V = Builder->CreateIntCast(Idx, Ty, false);
487 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
488 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
489 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
497 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
498 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
499 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
500 /// be complex, and scales are involved. The above expression would also be
501 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
502 /// This later form is less amenable to optimization though, and we are allowed
503 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
505 /// If we can't emit an optimized form for this expression, this returns null.
507 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
508 const DataLayout &DL = *IC.getDataLayout();
509 gep_type_iterator GTI = gep_type_begin(GEP);
511 // Check to see if this gep only has a single variable index. If so, and if
512 // any constant indices are a multiple of its scale, then we can compute this
513 // in terms of the scale of the variable index. For example, if the GEP
514 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
515 // because the expression will cross zero at the same point.
516 unsigned i, e = GEP->getNumOperands();
518 for (i = 1; i != e; ++i, ++GTI) {
519 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
520 // Compute the aggregate offset of constant indices.
521 if (CI->isZero()) continue;
523 // Handle a struct index, which adds its field offset to the pointer.
524 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
525 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
527 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
528 Offset += Size*CI->getSExtValue();
531 // Found our variable index.
536 // If there are no variable indices, we must have a constant offset, just
537 // evaluate it the general way.
538 if (i == e) return nullptr;
540 Value *VariableIdx = GEP->getOperand(i);
541 // Determine the scale factor of the variable element. For example, this is
542 // 4 if the variable index is into an array of i32.
543 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
545 // Verify that there are no other variable indices. If so, emit the hard way.
546 for (++i, ++GTI; i != e; ++i, ++GTI) {
547 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
548 if (!CI) return nullptr;
550 // Compute the aggregate offset of constant indices.
551 if (CI->isZero()) continue;
553 // Handle a struct index, which adds its field offset to the pointer.
554 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
555 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
557 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
558 Offset += Size*CI->getSExtValue();
564 // Okay, we know we have a single variable index, which must be a
565 // pointer/array/vector index. If there is no offset, life is simple, return
567 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
568 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
570 // Cast to intptrty in case a truncation occurs. If an extension is needed,
571 // we don't need to bother extending: the extension won't affect where the
572 // computation crosses zero.
573 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
574 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
579 // Otherwise, there is an index. The computation we will do will be modulo
580 // the pointer size, so get it.
581 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
583 Offset &= PtrSizeMask;
584 VariableScale &= PtrSizeMask;
586 // To do this transformation, any constant index must be a multiple of the
587 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
588 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
589 // multiple of the variable scale.
590 int64_t NewOffs = Offset / (int64_t)VariableScale;
591 if (Offset != NewOffs*(int64_t)VariableScale)
594 // Okay, we can do this evaluation. Start by converting the index to intptr.
595 if (VariableIdx->getType() != IntPtrTy)
596 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
598 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
599 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
602 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
603 /// else. At this point we know that the GEP is on the LHS of the comparison.
604 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
605 ICmpInst::Predicate Cond,
607 // Don't transform signed compares of GEPs into index compares. Even if the
608 // GEP is inbounds, the final add of the base pointer can have signed overflow
609 // and would change the result of the icmp.
610 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
611 // the maximum signed value for the pointer type.
612 if (ICmpInst::isSigned(Cond))
615 // Look through bitcasts and addrspacecasts. We do not however want to remove
617 if (!isa<GetElementPtrInst>(RHS))
618 RHS = RHS->stripPointerCasts();
620 Value *PtrBase = GEPLHS->getOperand(0);
621 if (DL && PtrBase == RHS && GEPLHS->isInBounds()) {
622 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
623 // This transformation (ignoring the base and scales) is valid because we
624 // know pointers can't overflow since the gep is inbounds. See if we can
625 // output an optimized form.
626 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
628 // If not, synthesize the offset the hard way.
630 Offset = EmitGEPOffset(GEPLHS);
631 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
632 Constant::getNullValue(Offset->getType()));
633 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
634 // If the base pointers are different, but the indices are the same, just
635 // compare the base pointer.
636 if (PtrBase != GEPRHS->getOperand(0)) {
637 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
638 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
639 GEPRHS->getOperand(0)->getType();
641 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
642 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
643 IndicesTheSame = false;
647 // If all indices are the same, just compare the base pointers.
649 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
651 // If we're comparing GEPs with two base pointers that only differ in type
652 // and both GEPs have only constant indices or just one use, then fold
653 // the compare with the adjusted indices.
654 if (DL && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
655 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
656 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
657 PtrBase->stripPointerCasts() ==
658 GEPRHS->getOperand(0)->stripPointerCasts()) {
659 Value *LOffset = EmitGEPOffset(GEPLHS);
660 Value *ROffset = EmitGEPOffset(GEPRHS);
662 // If we looked through an addrspacecast between different sized address
663 // spaces, the LHS and RHS pointers are different sized
664 // integers. Truncate to the smaller one.
665 Type *LHSIndexTy = LOffset->getType();
666 Type *RHSIndexTy = ROffset->getType();
667 if (LHSIndexTy != RHSIndexTy) {
668 if (LHSIndexTy->getPrimitiveSizeInBits() <
669 RHSIndexTy->getPrimitiveSizeInBits()) {
670 ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy);
672 LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy);
675 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
677 return ReplaceInstUsesWith(I, Cmp);
680 // Otherwise, the base pointers are different and the indices are
681 // different, bail out.
685 // If one of the GEPs has all zero indices, recurse.
686 if (GEPLHS->hasAllZeroIndices())
687 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
688 ICmpInst::getSwappedPredicate(Cond), I);
690 // If the other GEP has all zero indices, recurse.
691 if (GEPRHS->hasAllZeroIndices())
692 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
694 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
695 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
696 // If the GEPs only differ by one index, compare it.
697 unsigned NumDifferences = 0; // Keep track of # differences.
698 unsigned DiffOperand = 0; // The operand that differs.
699 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
700 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
701 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
702 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
703 // Irreconcilable differences.
707 if (NumDifferences++) break;
712 if (NumDifferences == 0) // SAME GEP?
713 return ReplaceInstUsesWith(I, // No comparison is needed here.
714 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
716 else if (NumDifferences == 1 && GEPsInBounds) {
717 Value *LHSV = GEPLHS->getOperand(DiffOperand);
718 Value *RHSV = GEPRHS->getOperand(DiffOperand);
719 // Make sure we do a signed comparison here.
720 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
724 // Only lower this if the icmp is the only user of the GEP or if we expect
725 // the result to fold to a constant!
728 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
729 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
730 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
731 Value *L = EmitGEPOffset(GEPLHS);
732 Value *R = EmitGEPOffset(GEPRHS);
733 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
739 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
740 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
741 Value *X, ConstantInt *CI,
742 ICmpInst::Predicate Pred) {
743 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
744 // so the values can never be equal. Similarly for all other "or equals"
747 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
748 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
749 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
750 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
752 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
753 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
756 // (X+1) >u X --> X <u (0-1) --> X != 255
757 // (X+2) >u X --> X <u (0-2) --> X <u 254
758 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
759 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
760 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
762 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
763 ConstantInt *SMax = ConstantInt::get(X->getContext(),
764 APInt::getSignedMaxValue(BitWidth));
766 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
767 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
768 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
769 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
770 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
771 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
772 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
773 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
775 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
776 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
777 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
778 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
779 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
780 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
782 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
783 Constant *C = Builder->getInt(CI->getValue()-1);
784 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
787 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
788 /// and CmpRHS are both known to be integer constants.
789 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
790 ConstantInt *DivRHS) {
791 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
792 const APInt &CmpRHSV = CmpRHS->getValue();
794 // FIXME: If the operand types don't match the type of the divide
795 // then don't attempt this transform. The code below doesn't have the
796 // logic to deal with a signed divide and an unsigned compare (and
797 // vice versa). This is because (x /s C1) <s C2 produces different
798 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
799 // (x /u C1) <u C2. Simply casting the operands and result won't
800 // work. :( The if statement below tests that condition and bails
802 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
803 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
805 if (DivRHS->isZero())
806 return nullptr; // The ProdOV computation fails on divide by zero.
807 if (DivIsSigned && DivRHS->isAllOnesValue())
808 return nullptr; // The overflow computation also screws up here
809 if (DivRHS->isOne()) {
810 // This eliminates some funny cases with INT_MIN.
811 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
815 // Compute Prod = CI * DivRHS. We are essentially solving an equation
816 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
817 // C2 (CI). By solving for X we can turn this into a range check
818 // instead of computing a divide.
819 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
821 // Determine if the product overflows by seeing if the product is
822 // not equal to the divide. Make sure we do the same kind of divide
823 // as in the LHS instruction that we're folding.
824 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
825 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
827 // Get the ICmp opcode
828 ICmpInst::Predicate Pred = ICI.getPredicate();
830 /// If the division is known to be exact, then there is no remainder from the
831 /// divide, so the covered range size is unit, otherwise it is the divisor.
832 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
834 // Figure out the interval that is being checked. For example, a comparison
835 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
836 // Compute this interval based on the constants involved and the signedness of
837 // the compare/divide. This computes a half-open interval, keeping track of
838 // whether either value in the interval overflows. After analysis each
839 // overflow variable is set to 0 if it's corresponding bound variable is valid
840 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
841 int LoOverflow = 0, HiOverflow = 0;
842 Constant *LoBound = nullptr, *HiBound = nullptr;
844 if (!DivIsSigned) { // udiv
845 // e.g. X/5 op 3 --> [15, 20)
847 HiOverflow = LoOverflow = ProdOV;
849 // If this is not an exact divide, then many values in the range collapse
850 // to the same result value.
851 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
854 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
855 if (CmpRHSV == 0) { // (X / pos) op 0
856 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
857 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
859 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
860 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
861 HiOverflow = LoOverflow = ProdOV;
863 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
864 } else { // (X / pos) op neg
865 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
866 HiBound = AddOne(Prod);
867 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
869 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
870 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
873 } else if (DivRHS->isNegative()) { // Divisor is < 0.
875 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
876 if (CmpRHSV == 0) { // (X / neg) op 0
877 // e.g. X/-5 op 0 --> [-4, 5)
878 LoBound = AddOne(RangeSize);
879 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
880 if (HiBound == DivRHS) { // -INTMIN = INTMIN
881 HiOverflow = 1; // [INTMIN+1, overflow)
882 HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
884 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
885 // e.g. X/-5 op 3 --> [-19, -14)
886 HiBound = AddOne(Prod);
887 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
889 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
890 } else { // (X / neg) op neg
891 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
892 LoOverflow = HiOverflow = ProdOV;
894 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
897 // Dividing by a negative swaps the condition. LT <-> GT
898 Pred = ICmpInst::getSwappedPredicate(Pred);
901 Value *X = DivI->getOperand(0);
903 default: llvm_unreachable("Unhandled icmp opcode!");
904 case ICmpInst::ICMP_EQ:
905 if (LoOverflow && HiOverflow)
906 return ReplaceInstUsesWith(ICI, Builder->getFalse());
908 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
909 ICmpInst::ICMP_UGE, X, LoBound);
911 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
912 ICmpInst::ICMP_ULT, X, HiBound);
913 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
915 case ICmpInst::ICMP_NE:
916 if (LoOverflow && HiOverflow)
917 return ReplaceInstUsesWith(ICI, Builder->getTrue());
919 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
920 ICmpInst::ICMP_ULT, X, LoBound);
922 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
923 ICmpInst::ICMP_UGE, X, HiBound);
924 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
925 DivIsSigned, false));
926 case ICmpInst::ICMP_ULT:
927 case ICmpInst::ICMP_SLT:
928 if (LoOverflow == +1) // Low bound is greater than input range.
929 return ReplaceInstUsesWith(ICI, Builder->getTrue());
930 if (LoOverflow == -1) // Low bound is less than input range.
931 return ReplaceInstUsesWith(ICI, Builder->getFalse());
932 return new ICmpInst(Pred, X, LoBound);
933 case ICmpInst::ICMP_UGT:
934 case ICmpInst::ICMP_SGT:
935 if (HiOverflow == +1) // High bound greater than input range.
936 return ReplaceInstUsesWith(ICI, Builder->getFalse());
937 if (HiOverflow == -1) // High bound less than input range.
938 return ReplaceInstUsesWith(ICI, Builder->getTrue());
939 if (Pred == ICmpInst::ICMP_UGT)
940 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
941 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
945 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
946 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
947 ConstantInt *ShAmt) {
948 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
950 // Check that the shift amount is in range. If not, don't perform
951 // undefined shifts. When the shift is visited it will be
953 uint32_t TypeBits = CmpRHSV.getBitWidth();
954 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
955 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
958 if (!ICI.isEquality()) {
959 // If we have an unsigned comparison and an ashr, we can't simplify this.
960 // Similarly for signed comparisons with lshr.
961 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
964 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
965 // by a power of 2. Since we already have logic to simplify these,
966 // transform to div and then simplify the resultant comparison.
967 if (Shr->getOpcode() == Instruction::AShr &&
968 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
971 // Revisit the shift (to delete it).
975 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
978 Shr->getOpcode() == Instruction::AShr ?
979 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
980 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
982 ICI.setOperand(0, Tmp);
984 // If the builder folded the binop, just return it.
985 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
989 // Otherwise, fold this div/compare.
990 assert(TheDiv->getOpcode() == Instruction::SDiv ||
991 TheDiv->getOpcode() == Instruction::UDiv);
993 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
994 assert(Res && "This div/cst should have folded!");
999 // If we are comparing against bits always shifted out, the
1000 // comparison cannot succeed.
1001 APInt Comp = CmpRHSV << ShAmtVal;
1002 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1003 if (Shr->getOpcode() == Instruction::LShr)
1004 Comp = Comp.lshr(ShAmtVal);
1006 Comp = Comp.ashr(ShAmtVal);
1008 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1009 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1010 Constant *Cst = Builder->getInt1(IsICMP_NE);
1011 return ReplaceInstUsesWith(ICI, Cst);
1014 // Otherwise, check to see if the bits shifted out are known to be zero.
1015 // If so, we can compare against the unshifted value:
1016 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1017 if (Shr->hasOneUse() && Shr->isExact())
1018 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1020 if (Shr->hasOneUse()) {
1021 // Otherwise strength reduce the shift into an and.
1022 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1023 Constant *Mask = Builder->getInt(Val);
1025 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1026 Mask, Shr->getName()+".mask");
1027 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1032 /// FoldICmpCstShrCst - Handle "(icmp eq/ne (ashr/lshr const2, A), const1)" ->
1033 /// (icmp eq/ne A, Log2(const2/const1)) ->
1034 /// (icmp eq/ne A, Log2(const2) - Log2(const1)).
1035 Instruction *InstCombiner::FoldICmpCstShrCst(ICmpInst &I, Value *Op, Value *A,
1038 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1040 auto getConstant = [&I, this](bool IsTrue) {
1041 if (I.getPredicate() == I.ICMP_NE)
1043 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1046 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1047 if (I.getPredicate() == I.ICMP_NE)
1048 Pred = CmpInst::getInversePredicate(Pred);
1049 return new ICmpInst(Pred, LHS, RHS);
1052 APInt AP1 = CI1->getValue();
1053 APInt AP2 = CI2->getValue();
1057 // Both Constants are 0.
1058 return getConstant(true);
1061 if (cast<BinaryOperator>(Op)->isExact())
1062 return getConstant(false);
1064 if (AP2.isNegative()) {
1065 // MSB is set, so a lshr with a large enough 'A' would be undefined.
1066 return getConstant(false);
1069 // 'A' must be large enough to shift out the highest set bit.
1070 return getICmp(I.ICMP_UGT, A,
1071 ConstantInt::get(A->getType(), AP2.logBase2()));
1075 // Shifting 0 by any value gives 0.
1076 return getConstant(false);
1079 bool IsAShr = isa<AShrOperator>(Op);
1081 if (AP1.isAllOnesValue() && IsAShr) {
1082 // Arithmatic shift of -1 is always -1.
1083 return getConstant(true);
1085 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1088 bool IsNegative = false;
1090 if (AP1.isNegative() != AP2.isNegative()) {
1091 // Arithmetic shift will never change the sign.
1092 return getConstant(false);
1094 // Both the constants are negative, take their positive to calculate log.
1095 if (AP1.isNegative()) {
1097 // Right-shifting won't increase the magnitude.
1098 return getConstant(false);
1103 if (!IsNegative && AP1.ugt(AP2))
1104 // Right-shifting will not increase the value.
1105 return getConstant(false);
1107 // Get the distance between the highest bit that's set.
1110 // Get the ones' complement of AP2 and AP1 when computing the distance.
1111 Shift = (~AP2).logBase2() - (~AP1).logBase2();
1113 Shift = AP2.logBase2() - AP1.logBase2();
1115 if (IsAShr ? AP1 == AP2.ashr(Shift) : AP1 == AP2.lshr(Shift))
1116 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1118 // Shifting const2 will never be equal to const1.
1119 return getConstant(false);
1122 /// FoldICmpCstShlCst - Handle "(icmp eq/ne (shl const2, A), const1)" ->
1123 /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)).
1124 Instruction *InstCombiner::FoldICmpCstShlCst(ICmpInst &I, Value *Op, Value *A,
1127 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1129 auto getConstant = [&I, this](bool IsTrue) {
1130 if (I.getPredicate() == I.ICMP_NE)
1132 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1135 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1136 if (I.getPredicate() == I.ICMP_NE)
1137 Pred = CmpInst::getInversePredicate(Pred);
1138 return new ICmpInst(Pred, LHS, RHS);
1141 APInt AP1 = CI1->getValue();
1142 APInt AP2 = CI2->getValue();
1144 assert(AP2 != 0 && "Handled in InstSimplify");
1146 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1148 if (!AP1 && AP2TrailingZeros != 0)
1149 return getICmp(I.ICMP_UGE, A,
1150 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1153 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1155 // Get the distance between the lowest bits that are set.
1156 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1158 if (Shift > 0 && AP2.shl(Shift) == AP1)
1159 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1161 // Shifting const2 will never be equal to const1.
1162 return getConstant(false);
1165 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1167 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1170 const APInt &RHSV = RHS->getValue();
1172 switch (LHSI->getOpcode()) {
1173 case Instruction::Trunc:
1174 if (ICI.isEquality() && LHSI->hasOneUse()) {
1175 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1176 // of the high bits truncated out of x are known.
1177 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1178 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1179 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1180 computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne, 0, &ICI);
1182 // If all the high bits are known, we can do this xform.
1183 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1184 // Pull in the high bits from known-ones set.
1185 APInt NewRHS = RHS->getValue().zext(SrcBits);
1186 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1187 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1188 Builder->getInt(NewRHS));
1193 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1194 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1195 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1197 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1198 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1199 Value *CompareVal = LHSI->getOperand(0);
1201 // If the sign bit of the XorCst is not set, there is no change to
1202 // the operation, just stop using the Xor.
1203 if (!XorCst->isNegative()) {
1204 ICI.setOperand(0, CompareVal);
1209 // Was the old condition true if the operand is positive?
1210 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1212 // If so, the new one isn't.
1213 isTrueIfPositive ^= true;
1215 if (isTrueIfPositive)
1216 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1219 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1223 if (LHSI->hasOneUse()) {
1224 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1225 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1226 const APInt &SignBit = XorCst->getValue();
1227 ICmpInst::Predicate Pred = ICI.isSigned()
1228 ? ICI.getUnsignedPredicate()
1229 : ICI.getSignedPredicate();
1230 return new ICmpInst(Pred, LHSI->getOperand(0),
1231 Builder->getInt(RHSV ^ SignBit));
1234 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1235 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1236 const APInt &NotSignBit = XorCst->getValue();
1237 ICmpInst::Predicate Pred = ICI.isSigned()
1238 ? ICI.getUnsignedPredicate()
1239 : ICI.getSignedPredicate();
1240 Pred = ICI.getSwappedPredicate(Pred);
1241 return new ICmpInst(Pred, LHSI->getOperand(0),
1242 Builder->getInt(RHSV ^ NotSignBit));
1246 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1247 // iff -C is a power of 2
1248 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1249 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1250 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1252 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1253 // iff -C is a power of 2
1254 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1255 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1256 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1259 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1260 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1261 LHSI->getOperand(0)->hasOneUse()) {
1262 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1264 // If the LHS is an AND of a truncating cast, we can widen the
1265 // and/compare to be the input width without changing the value
1266 // produced, eliminating a cast.
1267 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1268 // We can do this transformation if either the AND constant does not
1269 // have its sign bit set or if it is an equality comparison.
1270 // Extending a relational comparison when we're checking the sign
1271 // bit would not work.
1272 if (ICI.isEquality() ||
1273 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1275 Builder->CreateAnd(Cast->getOperand(0),
1276 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1277 NewAnd->takeName(LHSI);
1278 return new ICmpInst(ICI.getPredicate(), NewAnd,
1279 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1283 // If the LHS is an AND of a zext, and we have an equality compare, we can
1284 // shrink the and/compare to the smaller type, eliminating the cast.
1285 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1286 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1287 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1288 // should fold the icmp to true/false in that case.
1289 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1291 Builder->CreateAnd(Cast->getOperand(0),
1292 ConstantExpr::getTrunc(AndCst, Ty));
1293 NewAnd->takeName(LHSI);
1294 return new ICmpInst(ICI.getPredicate(), NewAnd,
1295 ConstantExpr::getTrunc(RHS, Ty));
1299 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1300 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1301 // happens a LOT in code produced by the C front-end, for bitfield
1303 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1304 if (Shift && !Shift->isShift())
1308 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
1310 // This seemingly simple opportunity to fold away a shift turns out to
1311 // be rather complicated. See PR17827
1312 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1314 bool CanFold = false;
1315 unsigned ShiftOpcode = Shift->getOpcode();
1316 if (ShiftOpcode == Instruction::AShr) {
1317 // There may be some constraints that make this possible,
1318 // but nothing simple has been discovered yet.
1320 } else if (ShiftOpcode == Instruction::Shl) {
1321 // For a left shift, we can fold if the comparison is not signed.
1322 // We can also fold a signed comparison if the mask value and
1323 // comparison value are not negative. These constraints may not be
1324 // obvious, but we can prove that they are correct using an SMT
1326 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1328 } else if (ShiftOpcode == Instruction::LShr) {
1329 // For a logical right shift, we can fold if the comparison is not
1330 // signed. We can also fold a signed comparison if the shifted mask
1331 // value and the shifted comparison value are not negative.
1332 // These constraints may not be obvious, but we can prove that they
1333 // are correct using an SMT solver.
1334 if (!ICI.isSigned())
1337 ConstantInt *ShiftedAndCst =
1338 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1339 ConstantInt *ShiftedRHSCst =
1340 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1342 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1349 if (ShiftOpcode == Instruction::Shl)
1350 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1352 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1354 // Check to see if we are shifting out any of the bits being
1356 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1357 // If we shifted bits out, the fold is not going to work out.
1358 // As a special case, check to see if this means that the
1359 // result is always true or false now.
1360 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1361 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1362 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1363 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1365 ICI.setOperand(1, NewCst);
1366 Constant *NewAndCst;
1367 if (ShiftOpcode == Instruction::Shl)
1368 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1370 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1371 LHSI->setOperand(1, NewAndCst);
1372 LHSI->setOperand(0, Shift->getOperand(0));
1373 Worklist.Add(Shift); // Shift is dead.
1379 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1380 // preferable because it allows the C<<Y expression to be hoisted out
1381 // of a loop if Y is invariant and X is not.
1382 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1383 ICI.isEquality() && !Shift->isArithmeticShift() &&
1384 !isa<Constant>(Shift->getOperand(0))) {
1387 if (Shift->getOpcode() == Instruction::LShr) {
1388 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1390 // Insert a logical shift.
1391 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1394 // Compute X & (C << Y).
1396 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1398 ICI.setOperand(0, NewAnd);
1402 // (icmp pred (and (or (lshr X, Y), X), 1), 0) -->
1403 // (icmp pred (and X, (or (shl 1, Y), 1), 0))
1405 // iff pred isn't signed
1407 Value *X, *Y, *LShr;
1408 if (!ICI.isSigned() && RHSV == 0) {
1409 if (match(LHSI->getOperand(1), m_One())) {
1410 Constant *One = cast<Constant>(LHSI->getOperand(1));
1411 Value *Or = LHSI->getOperand(0);
1412 if (match(Or, m_Or(m_Value(LShr), m_Value(X))) &&
1413 match(LShr, m_LShr(m_Specific(X), m_Value(Y)))) {
1414 unsigned UsesRemoved = 0;
1415 if (LHSI->hasOneUse())
1417 if (Or->hasOneUse())
1419 if (LShr->hasOneUse())
1421 Value *NewOr = nullptr;
1422 // Compute X & ((1 << Y) | 1)
1423 if (auto *C = dyn_cast<Constant>(Y)) {
1424 if (UsesRemoved >= 1)
1426 ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1428 if (UsesRemoved >= 3)
1429 NewOr = Builder->CreateOr(Builder->CreateShl(One, Y,
1432 One, Or->getName());
1435 Value *NewAnd = Builder->CreateAnd(X, NewOr, LHSI->getName());
1436 ICI.setOperand(0, NewAnd);
1444 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1445 // bit set in (X & AndCst) will produce a result greater than RHSV.
1446 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1447 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1448 if ((NTZ < AndCst->getBitWidth()) &&
1449 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1450 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1451 Constant::getNullValue(RHS->getType()));
1455 // Try to optimize things like "A[i]&42 == 0" to index computations.
1456 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1457 if (GetElementPtrInst *GEP =
1458 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1459 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1460 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1461 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1462 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1463 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1468 // X & -C == -C -> X > u ~C
1469 // X & -C != -C -> X <= u ~C
1470 // iff C is a power of 2
1471 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1472 return new ICmpInst(
1473 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1474 : ICmpInst::ICMP_ULE,
1475 LHSI->getOperand(0), SubOne(RHS));
1478 case Instruction::Or: {
1479 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1482 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1483 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1484 // -> and (icmp eq P, null), (icmp eq Q, null).
1485 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1486 Constant::getNullValue(P->getType()));
1487 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1488 Constant::getNullValue(Q->getType()));
1490 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1491 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1493 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1499 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1500 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1503 // If this is a signed comparison to 0 and the mul is sign preserving,
1504 // use the mul LHS operand instead.
1505 ICmpInst::Predicate pred = ICI.getPredicate();
1506 if (isSignTest(pred, RHS) && !Val->isZero() &&
1507 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1508 return new ICmpInst(Val->isNegative() ?
1509 ICmpInst::getSwappedPredicate(pred) : pred,
1510 LHSI->getOperand(0),
1511 Constant::getNullValue(RHS->getType()));
1516 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1517 uint32_t TypeBits = RHSV.getBitWidth();
1518 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1521 // (1 << X) pred P2 -> X pred Log2(P2)
1522 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1523 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1524 ICmpInst::Predicate Pred = ICI.getPredicate();
1525 if (ICI.isUnsigned()) {
1526 if (!RHSVIsPowerOf2) {
1527 // (1 << X) < 30 -> X <= 4
1528 // (1 << X) <= 30 -> X <= 4
1529 // (1 << X) >= 30 -> X > 4
1530 // (1 << X) > 30 -> X > 4
1531 if (Pred == ICmpInst::ICMP_ULT)
1532 Pred = ICmpInst::ICMP_ULE;
1533 else if (Pred == ICmpInst::ICMP_UGE)
1534 Pred = ICmpInst::ICMP_UGT;
1536 unsigned RHSLog2 = RHSV.logBase2();
1538 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1539 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1540 if (RHSLog2 == TypeBits-1) {
1541 if (Pred == ICmpInst::ICMP_UGE)
1542 Pred = ICmpInst::ICMP_EQ;
1543 else if (Pred == ICmpInst::ICMP_ULT)
1544 Pred = ICmpInst::ICMP_NE;
1547 return new ICmpInst(Pred, X,
1548 ConstantInt::get(RHS->getType(), RHSLog2));
1549 } else if (ICI.isSigned()) {
1550 if (RHSV.isAllOnesValue()) {
1551 // (1 << X) <= -1 -> X == 31
1552 if (Pred == ICmpInst::ICMP_SLE)
1553 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1554 ConstantInt::get(RHS->getType(), TypeBits-1));
1556 // (1 << X) > -1 -> X != 31
1557 if (Pred == ICmpInst::ICMP_SGT)
1558 return new ICmpInst(ICmpInst::ICMP_NE, X,
1559 ConstantInt::get(RHS->getType(), TypeBits-1));
1561 // (1 << X) < 0 -> X == 31
1562 // (1 << X) <= 0 -> X == 31
1563 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1564 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1565 ConstantInt::get(RHS->getType(), TypeBits-1));
1567 // (1 << X) >= 0 -> X != 31
1568 // (1 << X) > 0 -> X != 31
1569 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1570 return new ICmpInst(ICmpInst::ICMP_NE, X,
1571 ConstantInt::get(RHS->getType(), TypeBits-1));
1573 } else if (ICI.isEquality()) {
1575 return new ICmpInst(
1576 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1582 // Check that the shift amount is in range. If not, don't perform
1583 // undefined shifts. When the shift is visited it will be
1585 if (ShAmt->uge(TypeBits))
1588 if (ICI.isEquality()) {
1589 // If we are comparing against bits always shifted out, the
1590 // comparison cannot succeed.
1592 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1594 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1595 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1596 Constant *Cst = Builder->getInt1(IsICMP_NE);
1597 return ReplaceInstUsesWith(ICI, Cst);
1600 // If the shift is NUW, then it is just shifting out zeros, no need for an
1602 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1603 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1604 ConstantExpr::getLShr(RHS, ShAmt));
1606 // If the shift is NSW and we compare to 0, then it is just shifting out
1607 // sign bits, no need for an AND either.
1608 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1609 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1610 ConstantExpr::getLShr(RHS, ShAmt));
1612 if (LHSI->hasOneUse()) {
1613 // Otherwise strength reduce the shift into an and.
1614 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1615 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1616 TypeBits - ShAmtVal));
1619 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1620 return new ICmpInst(ICI.getPredicate(), And,
1621 ConstantExpr::getLShr(RHS, ShAmt));
1625 // If this is a signed comparison to 0 and the shift is sign preserving,
1626 // use the shift LHS operand instead.
1627 ICmpInst::Predicate pred = ICI.getPredicate();
1628 if (isSignTest(pred, RHS) &&
1629 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1630 return new ICmpInst(pred,
1631 LHSI->getOperand(0),
1632 Constant::getNullValue(RHS->getType()));
1634 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1635 bool TrueIfSigned = false;
1636 if (LHSI->hasOneUse() &&
1637 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1638 // (X << 31) <s 0 --> (X&1) != 0
1639 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1640 APInt::getOneBitSet(TypeBits,
1641 TypeBits-ShAmt->getZExtValue()-1));
1643 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1644 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1645 And, Constant::getNullValue(And->getType()));
1648 // Transform (icmp pred iM (shl iM %v, N), CI)
1649 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1650 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1651 // This enables to get rid of the shift in favor of a trunc which can be
1652 // free on the target. It has the additional benefit of comparing to a
1653 // smaller constant, which will be target friendly.
1654 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1655 if (LHSI->hasOneUse() &&
1656 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1657 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1658 Constant *NCI = ConstantExpr::getTrunc(
1659 ConstantExpr::getAShr(RHS,
1660 ConstantInt::get(RHS->getType(), Amt)),
1662 return new ICmpInst(ICI.getPredicate(),
1663 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1670 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1671 case Instruction::AShr: {
1672 // Handle equality comparisons of shift-by-constant.
1673 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1674 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1675 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1679 // Handle exact shr's.
1680 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1681 if (RHSV.isMinValue())
1682 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1687 case Instruction::SDiv:
1688 case Instruction::UDiv:
1689 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1690 // Fold this div into the comparison, producing a range check.
1691 // Determine, based on the divide type, what the range is being
1692 // checked. If there is an overflow on the low or high side, remember
1693 // it, otherwise compute the range [low, hi) bounding the new value.
1694 // See: InsertRangeTest above for the kinds of replacements possible.
1695 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1696 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1701 case Instruction::Sub: {
1702 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1704 const APInt &LHSV = LHSC->getValue();
1706 // C1-X <u C2 -> (X|(C2-1)) == C1
1707 // iff C1 & (C2-1) == C2-1
1708 // C2 is a power of 2
1709 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1710 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1711 return new ICmpInst(ICmpInst::ICMP_EQ,
1712 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1715 // C1-X >u C2 -> (X|C2) != C1
1716 // iff C1 & C2 == C2
1717 // C2+1 is a power of 2
1718 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1719 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1720 return new ICmpInst(ICmpInst::ICMP_NE,
1721 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1725 case Instruction::Add:
1726 // Fold: icmp pred (add X, C1), C2
1727 if (!ICI.isEquality()) {
1728 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1730 const APInt &LHSV = LHSC->getValue();
1732 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1735 if (ICI.isSigned()) {
1736 if (CR.getLower().isSignBit()) {
1737 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1738 Builder->getInt(CR.getUpper()));
1739 } else if (CR.getUpper().isSignBit()) {
1740 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1741 Builder->getInt(CR.getLower()));
1744 if (CR.getLower().isMinValue()) {
1745 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1746 Builder->getInt(CR.getUpper()));
1747 } else if (CR.getUpper().isMinValue()) {
1748 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1749 Builder->getInt(CR.getLower()));
1753 // X-C1 <u C2 -> (X & -C2) == C1
1754 // iff C1 & (C2-1) == 0
1755 // C2 is a power of 2
1756 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1757 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1758 return new ICmpInst(ICmpInst::ICMP_EQ,
1759 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1760 ConstantExpr::getNeg(LHSC));
1762 // X-C1 >u C2 -> (X & ~C2) != C1
1764 // C2+1 is a power of 2
1765 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1766 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1767 return new ICmpInst(ICmpInst::ICMP_NE,
1768 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1769 ConstantExpr::getNeg(LHSC));
1774 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1775 if (ICI.isEquality()) {
1776 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1778 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1779 // the second operand is a constant, simplify a bit.
1780 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1781 switch (BO->getOpcode()) {
1782 case Instruction::SRem:
1783 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1784 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1785 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1786 if (V.sgt(1) && V.isPowerOf2()) {
1788 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1790 return new ICmpInst(ICI.getPredicate(), NewRem,
1791 Constant::getNullValue(BO->getType()));
1795 case Instruction::Add:
1796 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1797 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1798 if (BO->hasOneUse())
1799 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1800 ConstantExpr::getSub(RHS, BOp1C));
1801 } else if (RHSV == 0) {
1802 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1803 // efficiently invertible, or if the add has just this one use.
1804 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1806 if (Value *NegVal = dyn_castNegVal(BOp1))
1807 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1808 if (Value *NegVal = dyn_castNegVal(BOp0))
1809 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1810 if (BO->hasOneUse()) {
1811 Value *Neg = Builder->CreateNeg(BOp1);
1813 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1817 case Instruction::Xor:
1818 // For the xor case, we can xor two constants together, eliminating
1819 // the explicit xor.
1820 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1821 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1822 ConstantExpr::getXor(RHS, BOC));
1823 } else if (RHSV == 0) {
1824 // Replace ((xor A, B) != 0) with (A != B)
1825 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1829 case Instruction::Sub:
1830 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1831 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1832 if (BO->hasOneUse())
1833 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1834 ConstantExpr::getSub(BOp0C, RHS));
1835 } else if (RHSV == 0) {
1836 // Replace ((sub A, B) != 0) with (A != B)
1837 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1841 case Instruction::Or:
1842 // If bits are being or'd in that are not present in the constant we
1843 // are comparing against, then the comparison could never succeed!
1844 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1845 Constant *NotCI = ConstantExpr::getNot(RHS);
1846 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1847 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1851 case Instruction::And:
1852 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1853 // If bits are being compared against that are and'd out, then the
1854 // comparison can never succeed!
1855 if ((RHSV & ~BOC->getValue()) != 0)
1856 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1858 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1859 if (RHS == BOC && RHSV.isPowerOf2())
1860 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1861 ICmpInst::ICMP_NE, LHSI,
1862 Constant::getNullValue(RHS->getType()));
1864 // Don't perform the following transforms if the AND has multiple uses
1865 if (!BO->hasOneUse())
1868 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1869 if (BOC->getValue().isSignBit()) {
1870 Value *X = BO->getOperand(0);
1871 Constant *Zero = Constant::getNullValue(X->getType());
1872 ICmpInst::Predicate pred = isICMP_NE ?
1873 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1874 return new ICmpInst(pred, X, Zero);
1877 // ((X & ~7) == 0) --> X < 8
1878 if (RHSV == 0 && isHighOnes(BOC)) {
1879 Value *X = BO->getOperand(0);
1880 Constant *NegX = ConstantExpr::getNeg(BOC);
1881 ICmpInst::Predicate pred = isICMP_NE ?
1882 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1883 return new ICmpInst(pred, X, NegX);
1887 case Instruction::Mul:
1888 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1889 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1890 // The trivial case (mul X, 0) is handled by InstSimplify
1891 // General case : (mul X, C) != 0 iff X != 0
1892 // (mul X, C) == 0 iff X == 0
1894 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1895 Constant::getNullValue(RHS->getType()));
1901 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1902 // Handle icmp {eq|ne} <intrinsic>, intcst.
1903 switch (II->getIntrinsicID()) {
1904 case Intrinsic::bswap:
1906 ICI.setOperand(0, II->getArgOperand(0));
1907 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1909 case Intrinsic::ctlz:
1910 case Intrinsic::cttz:
1911 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1912 if (RHSV == RHS->getType()->getBitWidth()) {
1914 ICI.setOperand(0, II->getArgOperand(0));
1915 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1919 case Intrinsic::ctpop:
1920 // popcount(A) == 0 -> A == 0 and likewise for !=
1921 if (RHS->isZero()) {
1923 ICI.setOperand(0, II->getArgOperand(0));
1924 ICI.setOperand(1, RHS);
1936 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1937 /// We only handle extending casts so far.
1939 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1940 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1941 Value *LHSCIOp = LHSCI->getOperand(0);
1942 Type *SrcTy = LHSCIOp->getType();
1943 Type *DestTy = LHSCI->getType();
1946 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1947 // integer type is the same size as the pointer type.
1948 if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
1949 DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1950 Value *RHSOp = nullptr;
1951 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1952 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1953 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1954 RHSOp = RHSC->getOperand(0);
1955 // If the pointer types don't match, insert a bitcast.
1956 if (LHSCIOp->getType() != RHSOp->getType())
1957 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1961 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1964 // The code below only handles extension cast instructions, so far.
1966 if (LHSCI->getOpcode() != Instruction::ZExt &&
1967 LHSCI->getOpcode() != Instruction::SExt)
1970 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1971 bool isSignedCmp = ICI.isSigned();
1973 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1974 // Not an extension from the same type?
1975 RHSCIOp = CI->getOperand(0);
1976 if (RHSCIOp->getType() != LHSCIOp->getType())
1979 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1980 // and the other is a zext), then we can't handle this.
1981 if (CI->getOpcode() != LHSCI->getOpcode())
1984 // Deal with equality cases early.
1985 if (ICI.isEquality())
1986 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1988 // A signed comparison of sign extended values simplifies into a
1989 // signed comparison.
1990 if (isSignedCmp && isSignedExt)
1991 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1993 // The other three cases all fold into an unsigned comparison.
1994 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1997 // If we aren't dealing with a constant on the RHS, exit early
1998 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
2002 // Compute the constant that would happen if we truncated to SrcTy then
2003 // reextended to DestTy.
2004 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
2005 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
2008 // If the re-extended constant didn't change...
2010 // Deal with equality cases early.
2011 if (ICI.isEquality())
2012 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
2014 // A signed comparison of sign extended values simplifies into a
2015 // signed comparison.
2016 if (isSignedExt && isSignedCmp)
2017 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
2019 // The other three cases all fold into an unsigned comparison.
2020 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
2023 // The re-extended constant changed so the constant cannot be represented
2024 // in the shorter type. Consequently, we cannot emit a simple comparison.
2025 // All the cases that fold to true or false will have already been handled
2026 // by SimplifyICmpInst, so only deal with the tricky case.
2028 if (isSignedCmp || !isSignedExt)
2031 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
2032 // should have been folded away previously and not enter in here.
2034 // We're performing an unsigned comp with a sign extended value.
2035 // This is true if the input is >= 0. [aka >s -1]
2036 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
2037 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
2039 // Finally, return the value computed.
2040 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
2041 return ReplaceInstUsesWith(ICI, Result);
2043 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
2044 return BinaryOperator::CreateNot(Result);
2047 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
2048 /// I = icmp ugt (add (add A, B), CI2), CI1
2049 /// If this is of the form:
2051 /// if (sum+128 >u 255)
2052 /// Then replace it with llvm.sadd.with.overflow.i8.
2054 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
2055 ConstantInt *CI2, ConstantInt *CI1,
2057 // The transformation we're trying to do here is to transform this into an
2058 // llvm.sadd.with.overflow. To do this, we have to replace the original add
2059 // with a narrower add, and discard the add-with-constant that is part of the
2060 // range check (if we can't eliminate it, this isn't profitable).
2062 // In order to eliminate the add-with-constant, the compare can be its only
2064 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
2065 if (!AddWithCst->hasOneUse()) return nullptr;
2067 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
2068 if (!CI2->getValue().isPowerOf2()) return nullptr;
2069 unsigned NewWidth = CI2->getValue().countTrailingZeros();
2070 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
2072 // The width of the new add formed is 1 more than the bias.
2075 // Check to see that CI1 is an all-ones value with NewWidth bits.
2076 if (CI1->getBitWidth() == NewWidth ||
2077 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
2080 // This is only really a signed overflow check if the inputs have been
2081 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
2082 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
2083 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
2084 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
2085 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
2088 // In order to replace the original add with a narrower
2089 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
2090 // and truncates that discard the high bits of the add. Verify that this is
2092 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
2093 for (User *U : OrigAdd->users()) {
2094 if (U == AddWithCst) continue;
2096 // Only accept truncates for now. We would really like a nice recursive
2097 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
2098 // chain to see which bits of a value are actually demanded. If the
2099 // original add had another add which was then immediately truncated, we
2100 // could still do the transformation.
2101 TruncInst *TI = dyn_cast<TruncInst>(U);
2102 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
2106 // If the pattern matches, truncate the inputs to the narrower type and
2107 // use the sadd_with_overflow intrinsic to efficiently compute both the
2108 // result and the overflow bit.
2109 Module *M = I.getParent()->getParent()->getParent();
2111 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
2112 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
2115 InstCombiner::BuilderTy *Builder = IC.Builder;
2117 // Put the new code above the original add, in case there are any uses of the
2118 // add between the add and the compare.
2119 Builder->SetInsertPoint(OrigAdd);
2121 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
2122 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
2123 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
2124 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
2125 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
2127 // The inner add was the result of the narrow add, zero extended to the
2128 // wider type. Replace it with the result computed by the intrinsic.
2129 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
2131 // The original icmp gets replaced with the overflow value.
2132 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
2135 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
2137 // Don't bother doing this transformation for pointers, don't do it for
2139 if (!isa<IntegerType>(OrigAddV->getType())) return nullptr;
2141 // If the add is a constant expr, then we don't bother transforming it.
2142 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
2143 if (!OrigAdd) return nullptr;
2145 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
2147 // Put the new code above the original add, in case there are any uses of the
2148 // add between the add and the compare.
2149 InstCombiner::BuilderTy *Builder = IC.Builder;
2150 Builder->SetInsertPoint(OrigAdd);
2152 Module *M = I.getParent()->getParent()->getParent();
2153 Type *Ty = LHS->getType();
2154 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
2155 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
2156 Value *Add = Builder->CreateExtractValue(Call, 0);
2158 IC.ReplaceInstUsesWith(*OrigAdd, Add);
2160 // The original icmp gets replaced with the overflow value.
2161 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2164 /// \brief Recognize and process idiom involving test for multiplication
2167 /// The caller has matched a pattern of the form:
2168 /// I = cmp u (mul(zext A, zext B), V
2169 /// The function checks if this is a test for overflow and if so replaces
2170 /// multiplication with call to 'mul.with.overflow' intrinsic.
2172 /// \param I Compare instruction.
2173 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
2174 /// the compare instruction. Must be of integer type.
2175 /// \param OtherVal The other argument of compare instruction.
2176 /// \returns Instruction which must replace the compare instruction, NULL if no
2177 /// replacement required.
2178 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
2179 Value *OtherVal, InstCombiner &IC) {
2180 // Don't bother doing this transformation for pointers, don't do it for
2182 if (!isa<IntegerType>(MulVal->getType()))
2185 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
2186 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
2187 Instruction *MulInstr = cast<Instruction>(MulVal);
2188 assert(MulInstr->getOpcode() == Instruction::Mul);
2190 Instruction *LHS = cast<Instruction>(MulInstr->getOperand(0)),
2191 *RHS = cast<Instruction>(MulInstr->getOperand(1));
2192 assert(LHS->getOpcode() == Instruction::ZExt);
2193 assert(RHS->getOpcode() == Instruction::ZExt);
2194 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
2196 // Calculate type and width of the result produced by mul.with.overflow.
2197 Type *TyA = A->getType(), *TyB = B->getType();
2198 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
2199 WidthB = TyB->getPrimitiveSizeInBits();
2202 if (WidthB > WidthA) {
2210 // In order to replace the original mul with a narrower mul.with.overflow,
2211 // all uses must ignore upper bits of the product. The number of used low
2212 // bits must be not greater than the width of mul.with.overflow.
2213 if (MulVal->hasNUsesOrMore(2))
2214 for (User *U : MulVal->users()) {
2217 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2218 // Check if truncation ignores bits above MulWidth.
2219 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
2220 if (TruncWidth > MulWidth)
2222 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2223 // Check if AND ignores bits above MulWidth.
2224 if (BO->getOpcode() != Instruction::And)
2226 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
2227 const APInt &CVal = CI->getValue();
2228 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
2232 // Other uses prohibit this transformation.
2237 // Recognize patterns
2238 switch (I.getPredicate()) {
2239 case ICmpInst::ICMP_EQ:
2240 case ICmpInst::ICMP_NE:
2241 // Recognize pattern:
2242 // mulval = mul(zext A, zext B)
2243 // cmp eq/neq mulval, zext trunc mulval
2244 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
2245 if (Zext->hasOneUse()) {
2246 Value *ZextArg = Zext->getOperand(0);
2247 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
2248 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
2252 // Recognize pattern:
2253 // mulval = mul(zext A, zext B)
2254 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
2257 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
2258 if (ValToMask != MulVal)
2260 const APInt &CVal = CI->getValue() + 1;
2261 if (CVal.isPowerOf2()) {
2262 unsigned MaskWidth = CVal.logBase2();
2263 if (MaskWidth == MulWidth)
2264 break; // Recognized
2269 case ICmpInst::ICMP_UGT:
2270 // Recognize pattern:
2271 // mulval = mul(zext A, zext B)
2272 // cmp ugt mulval, max
2273 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2274 APInt MaxVal = APInt::getMaxValue(MulWidth);
2275 MaxVal = MaxVal.zext(CI->getBitWidth());
2276 if (MaxVal.eq(CI->getValue()))
2277 break; // Recognized
2281 case ICmpInst::ICMP_UGE:
2282 // Recognize pattern:
2283 // mulval = mul(zext A, zext B)
2284 // cmp uge mulval, max+1
2285 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2286 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2287 if (MaxVal.eq(CI->getValue()))
2288 break; // Recognized
2292 case ICmpInst::ICMP_ULE:
2293 // Recognize pattern:
2294 // mulval = mul(zext A, zext B)
2295 // cmp ule mulval, max
2296 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2297 APInt MaxVal = APInt::getMaxValue(MulWidth);
2298 MaxVal = MaxVal.zext(CI->getBitWidth());
2299 if (MaxVal.eq(CI->getValue()))
2300 break; // Recognized
2304 case ICmpInst::ICMP_ULT:
2305 // Recognize pattern:
2306 // mulval = mul(zext A, zext B)
2307 // cmp ule mulval, max + 1
2308 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2309 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2310 if (MaxVal.eq(CI->getValue()))
2311 break; // Recognized
2319 InstCombiner::BuilderTy *Builder = IC.Builder;
2320 Builder->SetInsertPoint(MulInstr);
2321 Module *M = I.getParent()->getParent()->getParent();
2323 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
2324 Value *MulA = A, *MulB = B;
2325 if (WidthA < MulWidth)
2326 MulA = Builder->CreateZExt(A, MulType);
2327 if (WidthB < MulWidth)
2328 MulB = Builder->CreateZExt(B, MulType);
2330 Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
2331 CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
2332 IC.Worklist.Add(MulInstr);
2334 // If there are uses of mul result other than the comparison, we know that
2335 // they are truncation or binary AND. Change them to use result of
2336 // mul.with.overflow and adjust properly mask/size.
2337 if (MulVal->hasNUsesOrMore(2)) {
2338 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
2339 for (User *U : MulVal->users()) {
2340 if (U == &I || U == OtherVal)
2342 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2343 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
2344 IC.ReplaceInstUsesWith(*TI, Mul);
2346 TI->setOperand(0, Mul);
2347 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2348 assert(BO->getOpcode() == Instruction::And);
2349 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
2350 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
2351 APInt ShortMask = CI->getValue().trunc(MulWidth);
2352 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
2354 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
2355 IC.Worklist.Add(Zext);
2356 IC.ReplaceInstUsesWith(*BO, Zext);
2358 llvm_unreachable("Unexpected Binary operation");
2360 IC.Worklist.Add(cast<Instruction>(U));
2363 if (isa<Instruction>(OtherVal))
2364 IC.Worklist.Add(cast<Instruction>(OtherVal));
2366 // The original icmp gets replaced with the overflow value, maybe inverted
2367 // depending on predicate.
2368 bool Inverse = false;
2369 switch (I.getPredicate()) {
2370 case ICmpInst::ICMP_NE:
2372 case ICmpInst::ICMP_EQ:
2375 case ICmpInst::ICMP_UGT:
2376 case ICmpInst::ICMP_UGE:
2377 if (I.getOperand(0) == MulVal)
2381 case ICmpInst::ICMP_ULT:
2382 case ICmpInst::ICMP_ULE:
2383 if (I.getOperand(1) == MulVal)
2388 llvm_unreachable("Unexpected predicate");
2391 Value *Res = Builder->CreateExtractValue(Call, 1);
2392 return BinaryOperator::CreateNot(Res);
2395 return ExtractValueInst::Create(Call, 1);
2398 // DemandedBitsLHSMask - When performing a comparison against a constant,
2399 // it is possible that not all the bits in the LHS are demanded. This helper
2400 // method computes the mask that IS demanded.
2401 static APInt DemandedBitsLHSMask(ICmpInst &I,
2402 unsigned BitWidth, bool isSignCheck) {
2404 return APInt::getSignBit(BitWidth);
2406 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2407 if (!CI) return APInt::getAllOnesValue(BitWidth);
2408 const APInt &RHS = CI->getValue();
2410 switch (I.getPredicate()) {
2411 // For a UGT comparison, we don't care about any bits that
2412 // correspond to the trailing ones of the comparand. The value of these
2413 // bits doesn't impact the outcome of the comparison, because any value
2414 // greater than the RHS must differ in a bit higher than these due to carry.
2415 case ICmpInst::ICMP_UGT: {
2416 unsigned trailingOnes = RHS.countTrailingOnes();
2417 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2421 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2422 // Any value less than the RHS must differ in a higher bit because of carries.
2423 case ICmpInst::ICMP_ULT: {
2424 unsigned trailingZeros = RHS.countTrailingZeros();
2425 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2430 return APInt::getAllOnesValue(BitWidth);
2435 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2436 /// should be swapped.
2437 /// The decision is based on how many times these two operands are reused
2438 /// as subtract operands and their positions in those instructions.
2439 /// The rational is that several architectures use the same instruction for
2440 /// both subtract and cmp, thus it is better if the order of those operands
2442 /// \return true if Op0 and Op1 should be swapped.
2443 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2444 const Value * Op1) {
2445 // Filter out pointer value as those cannot appears directly in subtract.
2446 // FIXME: we may want to go through inttoptrs or bitcasts.
2447 if (Op0->getType()->isPointerTy())
2449 // Count every uses of both Op0 and Op1 in a subtract.
2450 // Each time Op0 is the first operand, count -1: swapping is bad, the
2451 // subtract has already the same layout as the compare.
2452 // Each time Op0 is the second operand, count +1: swapping is good, the
2453 // subtract has a different layout as the compare.
2454 // At the end, if the benefit is greater than 0, Op0 should come second to
2455 // expose more CSE opportunities.
2456 int GlobalSwapBenefits = 0;
2457 for (const User *U : Op0->users()) {
2458 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
2459 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2461 // If Op0 is the first argument, this is not beneficial to swap the
2463 int LocalSwapBenefits = -1;
2464 unsigned Op1Idx = 1;
2465 if (BinOp->getOperand(Op1Idx) == Op0) {
2467 LocalSwapBenefits = 1;
2469 if (BinOp->getOperand(Op1Idx) != Op1)
2471 GlobalSwapBenefits += LocalSwapBenefits;
2473 return GlobalSwapBenefits > 0;
2476 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2477 bool Changed = false;
2478 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2479 unsigned Op0Cplxity = getComplexity(Op0);
2480 unsigned Op1Cplxity = getComplexity(Op1);
2482 /// Orders the operands of the compare so that they are listed from most
2483 /// complex to least complex. This puts constants before unary operators,
2484 /// before binary operators.
2485 if (Op0Cplxity < Op1Cplxity ||
2486 (Op0Cplxity == Op1Cplxity &&
2487 swapMayExposeCSEOpportunities(Op0, Op1))) {
2489 std::swap(Op0, Op1);
2493 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
2494 return ReplaceInstUsesWith(I, V);
2496 // comparing -val or val with non-zero is the same as just comparing val
2497 // ie, abs(val) != 0 -> val != 0
2498 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2500 Value *Cond, *SelectTrue, *SelectFalse;
2501 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2502 m_Value(SelectFalse)))) {
2503 if (Value *V = dyn_castNegVal(SelectTrue)) {
2504 if (V == SelectFalse)
2505 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2507 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2508 if (V == SelectTrue)
2509 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2514 Type *Ty = Op0->getType();
2516 // icmp's with boolean values can always be turned into bitwise operations
2517 if (Ty->isIntegerTy(1)) {
2518 switch (I.getPredicate()) {
2519 default: llvm_unreachable("Invalid icmp instruction!");
2520 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2521 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2522 return BinaryOperator::CreateNot(Xor);
2524 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2525 return BinaryOperator::CreateXor(Op0, Op1);
2527 case ICmpInst::ICMP_UGT:
2528 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2530 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2531 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2532 return BinaryOperator::CreateAnd(Not, Op1);
2534 case ICmpInst::ICMP_SGT:
2535 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2537 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2538 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2539 return BinaryOperator::CreateAnd(Not, Op0);
2541 case ICmpInst::ICMP_UGE:
2542 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2544 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2545 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2546 return BinaryOperator::CreateOr(Not, Op1);
2548 case ICmpInst::ICMP_SGE:
2549 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2551 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2552 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2553 return BinaryOperator::CreateOr(Not, Op0);
2558 unsigned BitWidth = 0;
2559 if (Ty->isIntOrIntVectorTy())
2560 BitWidth = Ty->getScalarSizeInBits();
2561 else if (DL) // Pointers require DL info to get their size.
2562 BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
2564 bool isSignBit = false;
2566 // See if we are doing a comparison with a constant.
2567 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2568 Value *A = nullptr, *B = nullptr;
2570 // Match the following pattern, which is a common idiom when writing
2571 // overflow-safe integer arithmetic function. The source performs an
2572 // addition in wider type, and explicitly checks for overflow using
2573 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2574 // sadd_with_overflow intrinsic.
2576 // TODO: This could probably be generalized to handle other overflow-safe
2577 // operations if we worked out the formulas to compute the appropriate
2581 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2583 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2584 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2585 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2586 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2590 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2591 if (I.isEquality() && CI->isZero() &&
2592 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2593 // (icmp cond A B) if cond is equality
2594 return new ICmpInst(I.getPredicate(), A, B);
2597 // If we have an icmp le or icmp ge instruction, turn it into the
2598 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2599 // them being folded in the code below. The SimplifyICmpInst code has
2600 // already handled the edge cases for us, so we just assert on them.
2601 switch (I.getPredicate()) {
2603 case ICmpInst::ICMP_ULE:
2604 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2605 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2606 Builder->getInt(CI->getValue()+1));
2607 case ICmpInst::ICMP_SLE:
2608 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2609 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2610 Builder->getInt(CI->getValue()+1));
2611 case ICmpInst::ICMP_UGE:
2612 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2613 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2614 Builder->getInt(CI->getValue()-1));
2615 case ICmpInst::ICMP_SGE:
2616 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2617 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2618 Builder->getInt(CI->getValue()-1));
2621 if (I.isEquality()) {
2623 if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
2624 match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
2625 // (icmp eq/ne (ashr/lshr const2, A), const1)
2626 return FoldICmpCstShrCst(I, Op0, A, CI, CI2);
2628 if (match(Op0, m_Shl(m_ConstantInt(CI2), m_Value(A)))) {
2629 // (icmp eq/ne (shl const2, A), const1)
2630 return FoldICmpCstShlCst(I, Op0, A, CI, CI2);
2634 // If this comparison is a normal comparison, it demands all
2635 // bits, if it is a sign bit comparison, it only demands the sign bit.
2637 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2640 // See if we can fold the comparison based on range information we can get
2641 // by checking whether bits are known to be zero or one in the input.
2642 if (BitWidth != 0) {
2643 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2644 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2646 if (SimplifyDemandedBits(I.getOperandUse(0),
2647 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2648 Op0KnownZero, Op0KnownOne, 0))
2650 if (SimplifyDemandedBits(I.getOperandUse(1),
2651 APInt::getAllOnesValue(BitWidth),
2652 Op1KnownZero, Op1KnownOne, 0))
2655 // Given the known and unknown bits, compute a range that the LHS could be
2656 // in. Compute the Min, Max and RHS values based on the known bits. For the
2657 // EQ and NE we use unsigned values.
2658 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2659 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2661 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2663 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2666 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2668 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2672 // If Min and Max are known to be the same, then SimplifyDemandedBits
2673 // figured out that the LHS is a constant. Just constant fold this now so
2674 // that code below can assume that Min != Max.
2675 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2676 return new ICmpInst(I.getPredicate(),
2677 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2678 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2679 return new ICmpInst(I.getPredicate(), Op0,
2680 ConstantInt::get(Op1->getType(), Op1Min));
2682 // Based on the range information we know about the LHS, see if we can
2683 // simplify this comparison. For example, (x&4) < 8 is always true.
2684 switch (I.getPredicate()) {
2685 default: llvm_unreachable("Unknown icmp opcode!");
2686 case ICmpInst::ICMP_EQ: {
2687 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2688 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2690 // If all bits are known zero except for one, then we know at most one
2691 // bit is set. If the comparison is against zero, then this is a check
2692 // to see if *that* bit is set.
2693 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2694 if (~Op1KnownZero == 0) {
2695 // If the LHS is an AND with the same constant, look through it.
2696 Value *LHS = nullptr;
2697 ConstantInt *LHSC = nullptr;
2698 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2699 LHSC->getValue() != Op0KnownZeroInverted)
2702 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2703 // then turn "((1 << x)&8) == 0" into "x != 3".
2704 // or turn "((1 << x)&7) == 0" into "x > 2".
2706 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2707 APInt ValToCheck = Op0KnownZeroInverted;
2708 if (ValToCheck.isPowerOf2()) {
2709 unsigned CmpVal = ValToCheck.countTrailingZeros();
2710 return new ICmpInst(ICmpInst::ICMP_NE, X,
2711 ConstantInt::get(X->getType(), CmpVal));
2712 } else if ((++ValToCheck).isPowerOf2()) {
2713 unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
2714 return new ICmpInst(ICmpInst::ICMP_UGT, X,
2715 ConstantInt::get(X->getType(), CmpVal));
2719 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2720 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2722 if (Op0KnownZeroInverted == 1 &&
2723 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2724 return new ICmpInst(ICmpInst::ICMP_NE, X,
2725 ConstantInt::get(X->getType(),
2726 CI->countTrailingZeros()));
2731 case ICmpInst::ICMP_NE: {
2732 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2733 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2735 // If all bits are known zero except for one, then we know at most one
2736 // bit is set. If the comparison is against zero, then this is a check
2737 // to see if *that* bit is set.
2738 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2739 if (~Op1KnownZero == 0) {
2740 // If the LHS is an AND with the same constant, look through it.
2741 Value *LHS = nullptr;
2742 ConstantInt *LHSC = nullptr;
2743 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2744 LHSC->getValue() != Op0KnownZeroInverted)
2747 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2748 // then turn "((1 << x)&8) != 0" into "x == 3".
2749 // or turn "((1 << x)&7) != 0" into "x < 3".
2751 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2752 APInt ValToCheck = Op0KnownZeroInverted;
2753 if (ValToCheck.isPowerOf2()) {
2754 unsigned CmpVal = ValToCheck.countTrailingZeros();
2755 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2756 ConstantInt::get(X->getType(), CmpVal));
2757 } else if ((++ValToCheck).isPowerOf2()) {
2758 unsigned CmpVal = ValToCheck.countTrailingZeros();
2759 return new ICmpInst(ICmpInst::ICMP_ULT, X,
2760 ConstantInt::get(X->getType(), CmpVal));
2764 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2765 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2767 if (Op0KnownZeroInverted == 1 &&
2768 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2769 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2770 ConstantInt::get(X->getType(),
2771 CI->countTrailingZeros()));
2776 case ICmpInst::ICMP_ULT:
2777 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2778 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2779 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2780 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2781 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2782 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2783 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2784 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2785 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2786 Builder->getInt(CI->getValue()-1));
2788 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2789 if (CI->isMinValue(true))
2790 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2791 Constant::getAllOnesValue(Op0->getType()));
2794 case ICmpInst::ICMP_UGT:
2795 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2796 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2797 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2798 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2800 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2801 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2802 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2803 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2804 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2805 Builder->getInt(CI->getValue()+1));
2807 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2808 if (CI->isMaxValue(true))
2809 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2810 Constant::getNullValue(Op0->getType()));
2813 case ICmpInst::ICMP_SLT:
2814 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2815 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2816 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2817 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2818 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2819 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2820 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2821 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2822 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2823 Builder->getInt(CI->getValue()-1));
2826 case ICmpInst::ICMP_SGT:
2827 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2828 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2829 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2830 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2832 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2833 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2834 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2835 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2836 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2837 Builder->getInt(CI->getValue()+1));
2840 case ICmpInst::ICMP_SGE:
2841 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2842 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2843 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2844 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2845 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2847 case ICmpInst::ICMP_SLE:
2848 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2849 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2850 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2851 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2852 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2854 case ICmpInst::ICMP_UGE:
2855 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2856 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2857 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2858 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2859 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2861 case ICmpInst::ICMP_ULE:
2862 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2863 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2864 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2865 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2866 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2870 // Turn a signed comparison into an unsigned one if both operands
2871 // are known to have the same sign.
2873 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2874 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2875 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2878 // Test if the ICmpInst instruction is used exclusively by a select as
2879 // part of a minimum or maximum operation. If so, refrain from doing
2880 // any other folding. This helps out other analyses which understand
2881 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2882 // and CodeGen. And in this case, at least one of the comparison
2883 // operands has at least one user besides the compare (the select),
2884 // which would often largely negate the benefit of folding anyway.
2886 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
2887 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2888 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2891 // See if we are doing a comparison between a constant and an instruction that
2892 // can be folded into the comparison.
2893 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2894 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2895 // instruction, see if that instruction also has constants so that the
2896 // instruction can be folded into the icmp
2897 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2898 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2902 // Handle icmp with constant (but not simple integer constant) RHS
2903 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2904 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2905 switch (LHSI->getOpcode()) {
2906 case Instruction::GetElementPtr:
2907 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2908 if (RHSC->isNullValue() &&
2909 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2910 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2911 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2913 case Instruction::PHI:
2914 // Only fold icmp into the PHI if the phi and icmp are in the same
2915 // block. If in the same block, we're encouraging jump threading. If
2916 // not, we are just pessimizing the code by making an i1 phi.
2917 if (LHSI->getParent() == I.getParent())
2918 if (Instruction *NV = FoldOpIntoPhi(I))
2921 case Instruction::Select: {
2922 // If either operand of the select is a constant, we can fold the
2923 // comparison into the select arms, which will cause one to be
2924 // constant folded and the select turned into a bitwise or.
2925 Value *Op1 = nullptr, *Op2 = nullptr;
2926 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2927 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2928 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2929 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2931 // We only want to perform this transformation if it will not lead to
2932 // additional code. This is true if either both sides of the select
2933 // fold to a constant (in which case the icmp is replaced with a select
2934 // which will usually simplify) or this is the only user of the
2935 // select (in which case we are trading a select+icmp for a simpler
2937 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2939 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2942 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2944 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2948 case Instruction::IntToPtr:
2949 // icmp pred inttoptr(X), null -> icmp pred X, 0
2950 if (RHSC->isNullValue() && DL &&
2951 DL->getIntPtrType(RHSC->getType()) ==
2952 LHSI->getOperand(0)->getType())
2953 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2954 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2957 case Instruction::Load:
2958 // Try to optimize things like "A[i] > 4" to index computations.
2959 if (GetElementPtrInst *GEP =
2960 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2961 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2962 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2963 !cast<LoadInst>(LHSI)->isVolatile())
2964 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2971 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2972 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2973 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2975 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2976 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2977 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2980 // Test to see if the operands of the icmp are casted versions of other
2981 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2983 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2984 if (Op0->getType()->isPointerTy() &&
2985 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2986 // We keep moving the cast from the left operand over to the right
2987 // operand, where it can often be eliminated completely.
2988 Op0 = CI->getOperand(0);
2990 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2991 // so eliminate it as well.
2992 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2993 Op1 = CI2->getOperand(0);
2995 // If Op1 is a constant, we can fold the cast into the constant.
2996 if (Op0->getType() != Op1->getType()) {
2997 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2998 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
3000 // Otherwise, cast the RHS right before the icmp
3001 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
3004 return new ICmpInst(I.getPredicate(), Op0, Op1);
3008 if (isa<CastInst>(Op0)) {
3009 // Handle the special case of: icmp (cast bool to X), <cst>
3010 // This comes up when you have code like
3013 // For generality, we handle any zero-extension of any operand comparison
3014 // with a constant or another cast from the same type.
3015 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
3016 if (Instruction *R = visitICmpInstWithCastAndCast(I))
3020 // Special logic for binary operators.
3021 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3022 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3024 CmpInst::Predicate Pred = I.getPredicate();
3025 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3026 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3027 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
3028 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3029 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3030 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3031 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
3032 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3033 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3035 // Analyze the case when either Op0 or Op1 is an add instruction.
3036 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3037 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3038 if (BO0 && BO0->getOpcode() == Instruction::Add)
3039 A = BO0->getOperand(0), B = BO0->getOperand(1);
3040 if (BO1 && BO1->getOpcode() == Instruction::Add)
3041 C = BO1->getOperand(0), D = BO1->getOperand(1);
3043 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3044 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3045 return new ICmpInst(Pred, A == Op1 ? B : A,
3046 Constant::getNullValue(Op1->getType()));
3048 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3049 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3050 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3053 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3054 if (A && C && (A == C || A == D || B == C || B == D) &&
3055 NoOp0WrapProblem && NoOp1WrapProblem &&
3056 // Try not to increase register pressure.
3057 BO0->hasOneUse() && BO1->hasOneUse()) {
3058 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3061 // C + B == C + D -> B == D
3064 } else if (A == D) {
3065 // D + B == C + D -> B == C
3068 } else if (B == C) {
3069 // A + C == C + D -> A == D
3074 // A + D == C + D -> A == C
3078 return new ICmpInst(Pred, Y, Z);
3081 // icmp slt (X + -1), Y -> icmp sle X, Y
3082 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3083 match(B, m_AllOnes()))
3084 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3086 // icmp sge (X + -1), Y -> icmp sgt X, Y
3087 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3088 match(B, m_AllOnes()))
3089 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3091 // icmp sle (X + 1), Y -> icmp slt X, Y
3092 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
3094 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3096 // icmp sgt (X + 1), Y -> icmp sge X, Y
3097 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
3099 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3101 // if C1 has greater magnitude than C2:
3102 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3103 // s.t. C3 = C1 - C2
3105 // if C2 has greater magnitude than C1:
3106 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3107 // s.t. C3 = C2 - C1
3108 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3109 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3110 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3111 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3112 const APInt &AP1 = C1->getValue();
3113 const APInt &AP2 = C2->getValue();
3114 if (AP1.isNegative() == AP2.isNegative()) {
3115 APInt AP1Abs = C1->getValue().abs();
3116 APInt AP2Abs = C2->getValue().abs();
3117 if (AP1Abs.uge(AP2Abs)) {
3118 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
3119 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
3120 return new ICmpInst(Pred, NewAdd, C);
3122 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
3123 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
3124 return new ICmpInst(Pred, A, NewAdd);
3130 // Analyze the case when either Op0 or Op1 is a sub instruction.
3131 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3132 A = nullptr; B = nullptr; C = nullptr; D = nullptr;
3133 if (BO0 && BO0->getOpcode() == Instruction::Sub)
3134 A = BO0->getOperand(0), B = BO0->getOperand(1);
3135 if (BO1 && BO1->getOpcode() == Instruction::Sub)
3136 C = BO1->getOperand(0), D = BO1->getOperand(1);
3138 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3139 if (A == Op1 && NoOp0WrapProblem)
3140 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3142 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3143 if (C == Op0 && NoOp1WrapProblem)
3144 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3146 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3147 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3148 // Try not to increase register pressure.
3149 BO0->hasOneUse() && BO1->hasOneUse())
3150 return new ICmpInst(Pred, A, C);
3152 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3153 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3154 // Try not to increase register pressure.
3155 BO0->hasOneUse() && BO1->hasOneUse())
3156 return new ICmpInst(Pred, D, B);
3158 // icmp (0-X) < cst --> x > -cst
3159 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3161 if (match(BO0, m_Neg(m_Value(X))))
3162 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3163 if (!RHSC->isMinValue(/*isSigned=*/true))
3164 return new ICmpInst(I.getSwappedPredicate(), X,
3165 ConstantExpr::getNeg(RHSC));
3168 BinaryOperator *SRem = nullptr;
3169 // icmp (srem X, Y), Y
3170 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
3171 Op1 == BO0->getOperand(1))
3173 // icmp Y, (srem X, Y)
3174 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3175 Op0 == BO1->getOperand(1))
3178 // We don't check hasOneUse to avoid increasing register pressure because
3179 // the value we use is the same value this instruction was already using.
3180 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3182 case ICmpInst::ICMP_EQ:
3183 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3184 case ICmpInst::ICMP_NE:
3185 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3186 case ICmpInst::ICMP_SGT:
3187 case ICmpInst::ICMP_SGE:
3188 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3189 Constant::getAllOnesValue(SRem->getType()));
3190 case ICmpInst::ICMP_SLT:
3191 case ICmpInst::ICMP_SLE:
3192 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3193 Constant::getNullValue(SRem->getType()));
3197 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
3198 BO0->hasOneUse() && BO1->hasOneUse() &&
3199 BO0->getOperand(1) == BO1->getOperand(1)) {
3200 switch (BO0->getOpcode()) {
3202 case Instruction::Add:
3203 case Instruction::Sub:
3204 case Instruction::Xor:
3205 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3206 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3207 BO1->getOperand(0));
3208 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
3209 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3210 if (CI->getValue().isSignBit()) {
3211 ICmpInst::Predicate Pred = I.isSigned()
3212 ? I.getUnsignedPredicate()
3213 : I.getSignedPredicate();
3214 return new ICmpInst(Pred, BO0->getOperand(0),
3215 BO1->getOperand(0));
3218 if (CI->isMaxValue(true)) {
3219 ICmpInst::Predicate Pred = I.isSigned()
3220 ? I.getUnsignedPredicate()
3221 : I.getSignedPredicate();
3222 Pred = I.getSwappedPredicate(Pred);
3223 return new ICmpInst(Pred, BO0->getOperand(0),
3224 BO1->getOperand(0));
3228 case Instruction::Mul:
3229 if (!I.isEquality())
3232 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3233 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
3234 // Mask = -1 >> count-trailing-zeros(Cst).
3235 if (!CI->isZero() && !CI->isOne()) {
3236 const APInt &AP = CI->getValue();
3237 ConstantInt *Mask = ConstantInt::get(I.getContext(),
3238 APInt::getLowBitsSet(AP.getBitWidth(),
3240 AP.countTrailingZeros()));
3241 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
3242 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
3243 return new ICmpInst(I.getPredicate(), And1, And2);
3247 case Instruction::UDiv:
3248 case Instruction::LShr:
3252 case Instruction::SDiv:
3253 case Instruction::AShr:
3254 if (!BO0->isExact() || !BO1->isExact())
3256 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3257 BO1->getOperand(0));
3258 case Instruction::Shl: {
3259 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3260 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3263 if (!NSW && I.isSigned())
3265 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3266 BO1->getOperand(0));
3273 // Transform (A & ~B) == 0 --> (A & B) != 0
3274 // and (A & ~B) != 0 --> (A & B) == 0
3275 // if A is a power of 2.
3276 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
3277 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A, false,
3280 return new ICmpInst(I.getInversePredicate(),
3281 Builder->CreateAnd(A, B),
3284 // ~x < ~y --> y < x
3285 // ~x < cst --> ~cst < x
3286 if (match(Op0, m_Not(m_Value(A)))) {
3287 if (match(Op1, m_Not(m_Value(B))))
3288 return new ICmpInst(I.getPredicate(), B, A);
3289 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3290 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
3293 // (a+b) <u a --> llvm.uadd.with.overflow.
3294 // (a+b) <u b --> llvm.uadd.with.overflow.
3295 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
3296 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
3297 (Op1 == A || Op1 == B))
3298 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
3301 // a >u (a+b) --> llvm.uadd.with.overflow.
3302 // b >u (a+b) --> llvm.uadd.with.overflow.
3303 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
3304 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
3305 (Op0 == A || Op0 == B))
3306 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
3309 // (zext a) * (zext b) --> llvm.umul.with.overflow.
3310 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3311 if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
3314 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3315 if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
3320 if (I.isEquality()) {
3321 Value *A, *B, *C, *D;
3323 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3324 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3325 Value *OtherVal = A == Op1 ? B : A;
3326 return new ICmpInst(I.getPredicate(), OtherVal,
3327 Constant::getNullValue(A->getType()));
3330 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3331 // A^c1 == C^c2 --> A == C^(c1^c2)
3332 ConstantInt *C1, *C2;
3333 if (match(B, m_ConstantInt(C1)) &&
3334 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
3335 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3336 Value *Xor = Builder->CreateXor(C, NC);
3337 return new ICmpInst(I.getPredicate(), A, Xor);
3340 // A^B == A^D -> B == D
3341 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
3342 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
3343 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
3344 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
3348 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
3349 (A == Op0 || B == Op0)) {
3350 // A == (A^B) -> B == 0
3351 Value *OtherVal = A == Op0 ? B : A;
3352 return new ICmpInst(I.getPredicate(), OtherVal,
3353 Constant::getNullValue(A->getType()));
3356 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3357 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3358 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3359 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3362 X = B; Y = D; Z = A;
3363 } else if (A == D) {
3364 X = B; Y = C; Z = A;
3365 } else if (B == C) {
3366 X = A; Y = D; Z = B;
3367 } else if (B == D) {
3368 X = A; Y = C; Z = B;
3371 if (X) { // Build (X^Y) & Z
3372 Op1 = Builder->CreateXor(X, Y);
3373 Op1 = Builder->CreateAnd(Op1, Z);
3374 I.setOperand(0, Op1);
3375 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3380 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3381 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3383 if ((Op0->hasOneUse() &&
3384 match(Op0, m_ZExt(m_Value(A))) &&
3385 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3386 (Op1->hasOneUse() &&
3387 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3388 match(Op1, m_ZExt(m_Value(A))))) {
3389 APInt Pow2 = Cst1->getValue() + 1;
3390 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3391 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3392 return new ICmpInst(I.getPredicate(), A,
3393 Builder->CreateTrunc(B, A->getType()));
3396 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3397 // For lshr and ashr pairs.
3398 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3399 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3400 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3401 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3402 unsigned TypeBits = Cst1->getBitWidth();
3403 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3404 if (ShAmt < TypeBits && ShAmt != 0) {
3405 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
3406 ? ICmpInst::ICMP_UGE
3407 : ICmpInst::ICMP_ULT;
3408 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3409 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3410 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
3414 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3415 // "icmp (and X, mask), cst"
3417 if (Op0->hasOneUse() &&
3418 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
3419 m_ConstantInt(ShAmt))))) &&
3420 match(Op1, m_ConstantInt(Cst1)) &&
3421 // Only do this when A has multiple uses. This is most important to do
3422 // when it exposes other optimizations.
3424 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3426 if (ShAmt < ASize) {
3428 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3431 APInt CmpV = Cst1->getValue().zext(ASize);
3434 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3435 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3441 Value *X; ConstantInt *Cst;
3443 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3444 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3447 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3448 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3450 return Changed ? &I : nullptr;
3453 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3455 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3458 if (!isa<ConstantFP>(RHSC)) return nullptr;
3459 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3461 // Get the width of the mantissa. We don't want to hack on conversions that
3462 // might lose information from the integer, e.g. "i64 -> float"
3463 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3464 if (MantissaWidth == -1) return nullptr; // Unknown.
3466 // Check to see that the input is converted from an integer type that is small
3467 // enough that preserves all bits. TODO: check here for "known" sign bits.
3468 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3469 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
3471 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3472 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3476 // If the conversion would lose info, don't hack on this.
3477 if ((int)InputSize > MantissaWidth)
3480 // Otherwise, we can potentially simplify the comparison. We know that it
3481 // will always come through as an integer value and we know the constant is
3482 // not a NAN (it would have been previously simplified).
3483 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3485 ICmpInst::Predicate Pred;
3486 switch (I.getPredicate()) {
3487 default: llvm_unreachable("Unexpected predicate!");
3488 case FCmpInst::FCMP_UEQ:
3489 case FCmpInst::FCMP_OEQ:
3490 Pred = ICmpInst::ICMP_EQ;
3492 case FCmpInst::FCMP_UGT:
3493 case FCmpInst::FCMP_OGT:
3494 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3496 case FCmpInst::FCMP_UGE:
3497 case FCmpInst::FCMP_OGE:
3498 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3500 case FCmpInst::FCMP_ULT:
3501 case FCmpInst::FCMP_OLT:
3502 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3504 case FCmpInst::FCMP_ULE:
3505 case FCmpInst::FCMP_OLE:
3506 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3508 case FCmpInst::FCMP_UNE:
3509 case FCmpInst::FCMP_ONE:
3510 Pred = ICmpInst::ICMP_NE;
3512 case FCmpInst::FCMP_ORD:
3513 return ReplaceInstUsesWith(I, Builder->getTrue());
3514 case FCmpInst::FCMP_UNO:
3515 return ReplaceInstUsesWith(I, Builder->getFalse());
3518 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3520 // Now we know that the APFloat is a normal number, zero or inf.
3522 // See if the FP constant is too large for the integer. For example,
3523 // comparing an i8 to 300.0.
3524 unsigned IntWidth = IntTy->getScalarSizeInBits();
3527 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3528 // and large values.
3529 APFloat SMax(RHS.getSemantics());
3530 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3531 APFloat::rmNearestTiesToEven);
3532 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3533 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3534 Pred == ICmpInst::ICMP_SLE)
3535 return ReplaceInstUsesWith(I, Builder->getTrue());
3536 return ReplaceInstUsesWith(I, Builder->getFalse());
3539 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3540 // +INF and large values.
3541 APFloat UMax(RHS.getSemantics());
3542 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3543 APFloat::rmNearestTiesToEven);
3544 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3545 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3546 Pred == ICmpInst::ICMP_ULE)
3547 return ReplaceInstUsesWith(I, Builder->getTrue());
3548 return ReplaceInstUsesWith(I, Builder->getFalse());
3553 // See if the RHS value is < SignedMin.
3554 APFloat SMin(RHS.getSemantics());
3555 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3556 APFloat::rmNearestTiesToEven);
3557 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3558 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3559 Pred == ICmpInst::ICMP_SGE)
3560 return ReplaceInstUsesWith(I, Builder->getTrue());
3561 return ReplaceInstUsesWith(I, Builder->getFalse());
3564 // See if the RHS value is < UnsignedMin.
3565 APFloat SMin(RHS.getSemantics());
3566 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3567 APFloat::rmNearestTiesToEven);
3568 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3569 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3570 Pred == ICmpInst::ICMP_UGE)
3571 return ReplaceInstUsesWith(I, Builder->getTrue());
3572 return ReplaceInstUsesWith(I, Builder->getFalse());
3576 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3577 // [0, UMAX], but it may still be fractional. See if it is fractional by
3578 // casting the FP value to the integer value and back, checking for equality.
3579 // Don't do this for zero, because -0.0 is not fractional.
3580 Constant *RHSInt = LHSUnsigned
3581 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3582 : ConstantExpr::getFPToSI(RHSC, IntTy);
3583 if (!RHS.isZero()) {
3584 bool Equal = LHSUnsigned
3585 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3586 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3588 // If we had a comparison against a fractional value, we have to adjust
3589 // the compare predicate and sometimes the value. RHSC is rounded towards
3590 // zero at this point.
3592 default: llvm_unreachable("Unexpected integer comparison!");
3593 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3594 return ReplaceInstUsesWith(I, Builder->getTrue());
3595 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3596 return ReplaceInstUsesWith(I, Builder->getFalse());
3597 case ICmpInst::ICMP_ULE:
3598 // (float)int <= 4.4 --> int <= 4
3599 // (float)int <= -4.4 --> false
3600 if (RHS.isNegative())
3601 return ReplaceInstUsesWith(I, Builder->getFalse());
3603 case ICmpInst::ICMP_SLE:
3604 // (float)int <= 4.4 --> int <= 4
3605 // (float)int <= -4.4 --> int < -4
3606 if (RHS.isNegative())
3607 Pred = ICmpInst::ICMP_SLT;
3609 case ICmpInst::ICMP_ULT:
3610 // (float)int < -4.4 --> false
3611 // (float)int < 4.4 --> int <= 4
3612 if (RHS.isNegative())
3613 return ReplaceInstUsesWith(I, Builder->getFalse());
3614 Pred = ICmpInst::ICMP_ULE;
3616 case ICmpInst::ICMP_SLT:
3617 // (float)int < -4.4 --> int < -4
3618 // (float)int < 4.4 --> int <= 4
3619 if (!RHS.isNegative())
3620 Pred = ICmpInst::ICMP_SLE;
3622 case ICmpInst::ICMP_UGT:
3623 // (float)int > 4.4 --> int > 4
3624 // (float)int > -4.4 --> true
3625 if (RHS.isNegative())
3626 return ReplaceInstUsesWith(I, Builder->getTrue());
3628 case ICmpInst::ICMP_SGT:
3629 // (float)int > 4.4 --> int > 4
3630 // (float)int > -4.4 --> int >= -4
3631 if (RHS.isNegative())
3632 Pred = ICmpInst::ICMP_SGE;
3634 case ICmpInst::ICMP_UGE:
3635 // (float)int >= -4.4 --> true
3636 // (float)int >= 4.4 --> int > 4
3637 if (RHS.isNegative())
3638 return ReplaceInstUsesWith(I, Builder->getTrue());
3639 Pred = ICmpInst::ICMP_UGT;
3641 case ICmpInst::ICMP_SGE:
3642 // (float)int >= -4.4 --> int >= -4
3643 // (float)int >= 4.4 --> int > 4
3644 if (!RHS.isNegative())
3645 Pred = ICmpInst::ICMP_SGT;
3651 // Lower this FP comparison into an appropriate integer version of the
3653 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3656 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3657 bool Changed = false;
3659 /// Orders the operands of the compare so that they are listed from most
3660 /// complex to least complex. This puts constants before unary operators,
3661 /// before binary operators.
3662 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3667 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3669 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
3670 return ReplaceInstUsesWith(I, V);
3672 // Simplify 'fcmp pred X, X'
3674 switch (I.getPredicate()) {
3675 default: llvm_unreachable("Unknown predicate!");
3676 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3677 case FCmpInst::FCMP_ULT: // True if unordered or less than
3678 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3679 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3680 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3681 I.setPredicate(FCmpInst::FCMP_UNO);
3682 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3685 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3686 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3687 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3688 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3689 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3690 I.setPredicate(FCmpInst::FCMP_ORD);
3691 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3696 // Handle fcmp with constant RHS
3697 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3698 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3699 switch (LHSI->getOpcode()) {
3700 case Instruction::FPExt: {
3701 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3702 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3703 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3707 const fltSemantics *Sem;
3708 // FIXME: This shouldn't be here.
3709 if (LHSExt->getSrcTy()->isHalfTy())
3710 Sem = &APFloat::IEEEhalf;
3711 else if (LHSExt->getSrcTy()->isFloatTy())
3712 Sem = &APFloat::IEEEsingle;
3713 else if (LHSExt->getSrcTy()->isDoubleTy())
3714 Sem = &APFloat::IEEEdouble;
3715 else if (LHSExt->getSrcTy()->isFP128Ty())
3716 Sem = &APFloat::IEEEquad;
3717 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3718 Sem = &APFloat::x87DoubleExtended;
3719 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3720 Sem = &APFloat::PPCDoubleDouble;
3725 APFloat F = RHSF->getValueAPF();
3726 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3728 // Avoid lossy conversions and denormals. Zero is a special case
3729 // that's OK to convert.
3733 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3734 APFloat::cmpLessThan) || Fabs.isZero()))
3736 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3737 ConstantFP::get(RHSC->getContext(), F));
3740 case Instruction::PHI:
3741 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3742 // block. If in the same block, we're encouraging jump threading. If
3743 // not, we are just pessimizing the code by making an i1 phi.
3744 if (LHSI->getParent() == I.getParent())
3745 if (Instruction *NV = FoldOpIntoPhi(I))
3748 case Instruction::SIToFP:
3749 case Instruction::UIToFP:
3750 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3753 case Instruction::FSub: {
3754 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3756 if (match(LHSI, m_FNeg(m_Value(Op))))
3757 return new FCmpInst(I.getSwappedPredicate(), Op,
3758 ConstantExpr::getFNeg(RHSC));
3761 case Instruction::Load:
3762 if (GetElementPtrInst *GEP =
3763 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3764 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3765 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3766 !cast<LoadInst>(LHSI)->isVolatile())
3767 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3771 case Instruction::Call: {
3772 CallInst *CI = cast<CallInst>(LHSI);
3774 // Various optimization for fabs compared with zero.
3775 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3776 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3778 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3779 Func == LibFunc::fabsl) {
3780 switch (I.getPredicate()) {
3782 // fabs(x) < 0 --> false
3783 case FCmpInst::FCMP_OLT:
3784 return ReplaceInstUsesWith(I, Builder->getFalse());
3785 // fabs(x) > 0 --> x != 0
3786 case FCmpInst::FCMP_OGT:
3787 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3789 // fabs(x) <= 0 --> x == 0
3790 case FCmpInst::FCMP_OLE:
3791 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3793 // fabs(x) >= 0 --> !isnan(x)
3794 case FCmpInst::FCMP_OGE:
3795 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3797 // fabs(x) == 0 --> x == 0
3798 // fabs(x) != 0 --> x != 0
3799 case FCmpInst::FCMP_OEQ:
3800 case FCmpInst::FCMP_UEQ:
3801 case FCmpInst::FCMP_ONE:
3802 case FCmpInst::FCMP_UNE:
3803 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3812 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3814 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3815 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3817 // fcmp (fpext x), (fpext y) -> fcmp x, y
3818 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3819 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3820 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3821 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3822 RHSExt->getOperand(0));
3824 return Changed ? &I : nullptr;