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/DataLayout.h"
19 #include "llvm/IR/IntrinsicInst.h"
20 #include "llvm/Support/ConstantRange.h"
21 #include "llvm/Support/GetElementPtrTypeIterator.h"
22 #include "llvm/Support/PatternMatch.h"
23 #include "llvm/Target/TargetLibraryInfo.h"
25 using namespace PatternMatch;
27 static ConstantInt *getOne(Constant *C) {
28 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
31 /// AddOne - Add one to a ConstantInt
32 static Constant *AddOne(Constant *C) {
33 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
35 /// SubOne - Subtract one from a ConstantInt
36 static Constant *SubOne(Constant *C) {
37 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
40 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
41 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
44 static bool HasAddOverflow(ConstantInt *Result,
45 ConstantInt *In1, ConstantInt *In2,
48 return Result->getValue().ult(In1->getValue());
50 if (In2->isNegative())
51 return Result->getValue().sgt(In1->getValue());
52 return Result->getValue().slt(In1->getValue());
55 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
56 /// overflowed for this type.
57 static bool AddWithOverflow(Constant *&Result, Constant *In1,
58 Constant *In2, bool IsSigned = false) {
59 Result = ConstantExpr::getAdd(In1, In2);
61 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
62 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
63 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
64 if (HasAddOverflow(ExtractElement(Result, Idx),
65 ExtractElement(In1, Idx),
66 ExtractElement(In2, Idx),
73 return HasAddOverflow(cast<ConstantInt>(Result),
74 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
78 static bool HasSubOverflow(ConstantInt *Result,
79 ConstantInt *In1, ConstantInt *In2,
82 return Result->getValue().ugt(In1->getValue());
84 if (In2->isNegative())
85 return Result->getValue().slt(In1->getValue());
87 return Result->getValue().sgt(In1->getValue());
90 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
91 /// overflowed for this type.
92 static bool SubWithOverflow(Constant *&Result, Constant *In1,
93 Constant *In2, bool IsSigned = false) {
94 Result = ConstantExpr::getSub(In1, In2);
96 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
97 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
98 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
99 if (HasSubOverflow(ExtractElement(Result, Idx),
100 ExtractElement(In1, Idx),
101 ExtractElement(In2, Idx),
108 return HasSubOverflow(cast<ConstantInt>(Result),
109 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
113 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
114 /// comparison only checks the sign bit. If it only checks the sign bit, set
115 /// TrueIfSigned if the result of the comparison is true when the input value is
117 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
118 bool &TrueIfSigned) {
120 case ICmpInst::ICMP_SLT: // True if LHS s< 0
122 return RHS->isZero();
123 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
125 return RHS->isAllOnesValue();
126 case ICmpInst::ICMP_SGT: // True if LHS s> -1
127 TrueIfSigned = false;
128 return RHS->isAllOnesValue();
129 case ICmpInst::ICMP_UGT:
130 // True if LHS u> RHS and RHS == high-bit-mask - 1
132 return RHS->isMaxValue(true);
133 case ICmpInst::ICMP_UGE:
134 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
136 return RHS->getValue().isSignBit();
142 /// Returns true if the exploded icmp can be expressed as a comparison to zero
143 /// and update the predicate accordingly. The signedness of the comparison is
144 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
145 if (!ICmpInst::isSigned(pred))
153 case ICmpInst::ICMP_SGE:
154 pred = ICmpInst::ICMP_SGT;
156 case ICmpInst::ICMP_SLT:
157 pred = ICmpInst::ICMP_SLE;
163 if (RHS->isAllOnesValue())
165 case ICmpInst::ICMP_SLE:
166 pred = ICmpInst::ICMP_SLT;
168 case ICmpInst::ICMP_SGT:
169 pred = ICmpInst::ICMP_SGE;
178 // isHighOnes - Return true if the constant is of the form 1+0+.
179 // This is the same as lowones(~X).
180 static bool isHighOnes(const ConstantInt *CI) {
181 return (~CI->getValue() + 1).isPowerOf2();
184 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
185 /// set of known zero and one bits, compute the maximum and minimum values that
186 /// could have the specified known zero and known one bits, returning them in
188 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
189 const APInt& KnownOne,
190 APInt& Min, APInt& Max) {
191 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
192 KnownZero.getBitWidth() == Min.getBitWidth() &&
193 KnownZero.getBitWidth() == Max.getBitWidth() &&
194 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
195 APInt UnknownBits = ~(KnownZero|KnownOne);
197 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
198 // bit if it is unknown.
200 Max = KnownOne|UnknownBits;
202 if (UnknownBits.isNegative()) { // Sign bit is unknown
203 Min.setBit(Min.getBitWidth()-1);
204 Max.clearBit(Max.getBitWidth()-1);
208 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
209 // a set of known zero and one bits, compute the maximum and minimum values that
210 // could have the specified known zero and known one bits, returning them in
212 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
213 const APInt &KnownOne,
214 APInt &Min, APInt &Max) {
215 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
216 KnownZero.getBitWidth() == Min.getBitWidth() &&
217 KnownZero.getBitWidth() == Max.getBitWidth() &&
218 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
219 APInt UnknownBits = ~(KnownZero|KnownOne);
221 // The minimum value is when the unknown bits are all zeros.
223 // The maximum value is when the unknown bits are all ones.
224 Max = KnownOne|UnknownBits;
229 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
230 /// cmp pred (load (gep GV, ...)), cmpcst
231 /// where GV is a global variable with a constant initializer. Try to simplify
232 /// this into some simple computation that does not need the load. For example
233 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
235 /// If AndCst is non-null, then the loaded value is masked with that constant
236 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
237 Instruction *InstCombiner::
238 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
239 CmpInst &ICI, ConstantInt *AndCst) {
240 // We need TD information to know the pointer size unless this is inbounds.
241 if (!GEP->isInBounds() && TD == 0) return 0;
243 Constant *Init = GV->getInitializer();
244 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
247 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
248 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays.
250 // There are many forms of this optimization we can handle, for now, just do
251 // the simple index into a single-dimensional array.
253 // Require: GEP GV, 0, i {{, constant indices}}
254 if (GEP->getNumOperands() < 3 ||
255 !isa<ConstantInt>(GEP->getOperand(1)) ||
256 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
257 isa<Constant>(GEP->getOperand(2)))
260 // Check that indices after the variable are constants and in-range for the
261 // type they index. Collect the indices. This is typically for arrays of
263 SmallVector<unsigned, 4> LaterIndices;
265 Type *EltTy = Init->getType()->getArrayElementType();
266 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
267 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
268 if (Idx == 0) return 0; // Variable index.
270 uint64_t IdxVal = Idx->getZExtValue();
271 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
273 if (StructType *STy = dyn_cast<StructType>(EltTy))
274 EltTy = STy->getElementType(IdxVal);
275 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
276 if (IdxVal >= ATy->getNumElements()) return 0;
277 EltTy = ATy->getElementType();
279 return 0; // Unknown type.
282 LaterIndices.push_back(IdxVal);
285 enum { Overdefined = -3, Undefined = -2 };
287 // Variables for our state machines.
289 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
290 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
291 // and 87 is the second (and last) index. FirstTrueElement is -2 when
292 // undefined, otherwise set to the first true element. SecondTrueElement is
293 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
294 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
296 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
297 // form "i != 47 & i != 87". Same state transitions as for true elements.
298 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
300 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
301 /// define a state machine that triggers for ranges of values that the index
302 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
303 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
304 /// index in the range (inclusive). We use -2 for undefined here because we
305 /// use relative comparisons and don't want 0-1 to match -1.
306 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
308 // MagicBitvector - This is a magic bitvector where we set a bit if the
309 // comparison is true for element 'i'. If there are 64 elements or less in
310 // the array, this will fully represent all the comparison results.
311 uint64_t MagicBitvector = 0;
314 // Scan the array and see if one of our patterns matches.
315 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
316 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
317 Constant *Elt = Init->getAggregateElement(i);
318 if (Elt == 0) return 0;
320 // If this is indexing an array of structures, get the structure element.
321 if (!LaterIndices.empty())
322 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
324 // If the element is masked, handle it.
325 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
327 // Find out if the comparison would be true or false for the i'th element.
328 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
329 CompareRHS, TD, TLI);
330 // If the result is undef for this element, ignore it.
331 if (isa<UndefValue>(C)) {
332 // Extend range state machines to cover this element in case there is an
333 // undef in the middle of the range.
334 if (TrueRangeEnd == (int)i-1)
336 if (FalseRangeEnd == (int)i-1)
341 // If we can't compute the result for any of the elements, we have to give
342 // up evaluating the entire conditional.
343 if (!isa<ConstantInt>(C)) return 0;
345 // Otherwise, we know if the comparison is true or false for this element,
346 // update our state machines.
347 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
349 // State machine for single/double/range index comparison.
351 // Update the TrueElement state machine.
352 if (FirstTrueElement == Undefined)
353 FirstTrueElement = TrueRangeEnd = i; // First true element.
355 // Update double-compare state machine.
356 if (SecondTrueElement == Undefined)
357 SecondTrueElement = i;
359 SecondTrueElement = Overdefined;
361 // Update range state machine.
362 if (TrueRangeEnd == (int)i-1)
365 TrueRangeEnd = Overdefined;
368 // Update the FalseElement state machine.
369 if (FirstFalseElement == Undefined)
370 FirstFalseElement = FalseRangeEnd = i; // First false element.
372 // Update double-compare state machine.
373 if (SecondFalseElement == Undefined)
374 SecondFalseElement = i;
376 SecondFalseElement = Overdefined;
378 // Update range state machine.
379 if (FalseRangeEnd == (int)i-1)
382 FalseRangeEnd = Overdefined;
387 // If this element is in range, update our magic bitvector.
388 if (i < 64 && IsTrueForElt)
389 MagicBitvector |= 1ULL << i;
391 // If all of our states become overdefined, bail out early. Since the
392 // predicate is expensive, only check it every 8 elements. This is only
393 // really useful for really huge arrays.
394 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
395 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
396 FalseRangeEnd == Overdefined)
400 // Now that we've scanned the entire array, emit our new comparison(s). We
401 // order the state machines in complexity of the generated code.
402 Value *Idx = GEP->getOperand(2);
404 // If the index is larger than the pointer size of the target, truncate the
405 // index down like the GEP would do implicitly. We don't have to do this for
406 // an inbounds GEP because the index can't be out of range.
407 if (!GEP->isInBounds() &&
408 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
409 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
411 // If the comparison is only true for one or two elements, emit direct
413 if (SecondTrueElement != Overdefined) {
414 // None true -> false.
415 if (FirstTrueElement == Undefined)
416 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
418 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
420 // True for one element -> 'i == 47'.
421 if (SecondTrueElement == Undefined)
422 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
424 // True for two elements -> 'i == 47 | i == 72'.
425 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
426 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
427 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
428 return BinaryOperator::CreateOr(C1, C2);
431 // If the comparison is only false for one or two elements, emit direct
433 if (SecondFalseElement != Overdefined) {
434 // None false -> true.
435 if (FirstFalseElement == Undefined)
436 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
438 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
440 // False for one element -> 'i != 47'.
441 if (SecondFalseElement == Undefined)
442 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
444 // False for two elements -> 'i != 47 & i != 72'.
445 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
446 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
447 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
448 return BinaryOperator::CreateAnd(C1, C2);
451 // If the comparison can be replaced with a range comparison for the elements
452 // where it is true, emit the range check.
453 if (TrueRangeEnd != Overdefined) {
454 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
456 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
457 if (FirstTrueElement) {
458 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
459 Idx = Builder->CreateAdd(Idx, Offs);
462 Value *End = ConstantInt::get(Idx->getType(),
463 TrueRangeEnd-FirstTrueElement+1);
464 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
467 // False range check.
468 if (FalseRangeEnd != Overdefined) {
469 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
470 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
471 if (FirstFalseElement) {
472 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
473 Idx = Builder->CreateAdd(Idx, Offs);
476 Value *End = ConstantInt::get(Idx->getType(),
477 FalseRangeEnd-FirstFalseElement);
478 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
482 // If a magic bitvector captures the entire comparison state
483 // of this load, replace it with computation that does:
484 // ((magic_cst >> i) & 1) != 0
488 // Look for an appropriate type:
489 // - The type of Idx if the magic fits
490 // - The smallest fitting legal type if we have a DataLayout
492 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
495 Ty = TD->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
496 else if (ArrayElementCount <= 32)
497 Ty = Type::getInt32Ty(Init->getContext());
500 Value *V = Builder->CreateIntCast(Idx, Ty, false);
501 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
502 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
503 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
511 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
512 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
513 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
514 /// be complex, and scales are involved. The above expression would also be
515 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
516 /// This later form is less amenable to optimization though, and we are allowed
517 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
519 /// If we can't emit an optimized form for this expression, this returns null.
521 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
522 DataLayout &TD = *IC.getDataLayout();
523 gep_type_iterator GTI = gep_type_begin(GEP);
525 // Check to see if this gep only has a single variable index. If so, and if
526 // any constant indices are a multiple of its scale, then we can compute this
527 // in terms of the scale of the variable index. For example, if the GEP
528 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
529 // because the expression will cross zero at the same point.
530 unsigned i, e = GEP->getNumOperands();
532 for (i = 1; i != e; ++i, ++GTI) {
533 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
534 // Compute the aggregate offset of constant indices.
535 if (CI->isZero()) continue;
537 // Handle a struct index, which adds its field offset to the pointer.
538 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
539 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
541 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
542 Offset += Size*CI->getSExtValue();
545 // Found our variable index.
550 // If there are no variable indices, we must have a constant offset, just
551 // evaluate it the general way.
552 if (i == e) return 0;
554 Value *VariableIdx = GEP->getOperand(i);
555 // Determine the scale factor of the variable element. For example, this is
556 // 4 if the variable index is into an array of i32.
557 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
559 // Verify that there are no other variable indices. If so, emit the hard way.
560 for (++i, ++GTI; i != e; ++i, ++GTI) {
561 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
564 // Compute the aggregate offset of constant indices.
565 if (CI->isZero()) continue;
567 // Handle a struct index, which adds its field offset to the pointer.
568 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
569 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
571 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
572 Offset += Size*CI->getSExtValue();
576 // Okay, we know we have a single variable index, which must be a
577 // pointer/array/vector index. If there is no offset, life is simple, return
579 unsigned IntPtrWidth = TD.getPointerSizeInBits();
581 // Cast to intptrty in case a truncation occurs. If an extension is needed,
582 // we don't need to bother extending: the extension won't affect where the
583 // computation crosses zero.
584 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
585 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
586 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
591 // Otherwise, there is an index. The computation we will do will be modulo
592 // the pointer size, so get it.
593 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
595 Offset &= PtrSizeMask;
596 VariableScale &= PtrSizeMask;
598 // To do this transformation, any constant index must be a multiple of the
599 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
600 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
601 // multiple of the variable scale.
602 int64_t NewOffs = Offset / (int64_t)VariableScale;
603 if (Offset != NewOffs*(int64_t)VariableScale)
606 // Okay, we can do this evaluation. Start by converting the index to intptr.
607 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
608 if (VariableIdx->getType() != IntPtrTy)
609 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
611 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
612 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
615 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
616 /// else. At this point we know that the GEP is on the LHS of the comparison.
617 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
618 ICmpInst::Predicate Cond,
620 // Don't transform signed compares of GEPs into index compares. Even if the
621 // GEP is inbounds, the final add of the base pointer can have signed overflow
622 // and would change the result of the icmp.
623 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
624 // the maximum signed value for the pointer type.
625 if (ICmpInst::isSigned(Cond))
628 // Look through bitcasts.
629 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
630 RHS = BCI->getOperand(0);
632 Value *PtrBase = GEPLHS->getOperand(0);
633 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
634 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
635 // This transformation (ignoring the base and scales) is valid because we
636 // know pointers can't overflow since the gep is inbounds. See if we can
637 // output an optimized form.
638 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
640 // If not, synthesize the offset the hard way.
642 Offset = EmitGEPOffset(GEPLHS);
643 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
644 Constant::getNullValue(Offset->getType()));
645 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
646 // If the base pointers are different, but the indices are the same, just
647 // compare the base pointer.
648 if (PtrBase != GEPRHS->getOperand(0)) {
649 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
650 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
651 GEPRHS->getOperand(0)->getType();
653 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
654 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
655 IndicesTheSame = false;
659 // If all indices are the same, just compare the base pointers.
661 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
662 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
664 // If we're comparing GEPs with two base pointers that only differ in type
665 // and both GEPs have only constant indices or just one use, then fold
666 // the compare with the adjusted indices.
667 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
668 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
669 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
670 PtrBase->stripPointerCasts() ==
671 GEPRHS->getOperand(0)->stripPointerCasts()) {
672 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
673 EmitGEPOffset(GEPLHS),
674 EmitGEPOffset(GEPRHS));
675 return ReplaceInstUsesWith(I, Cmp);
678 // Otherwise, the base pointers are different and the indices are
679 // different, bail out.
683 // If one of the GEPs has all zero indices, recurse.
684 bool AllZeros = true;
685 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
686 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
687 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
692 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
693 ICmpInst::getSwappedPredicate(Cond), I);
695 // If the other GEP has all zero indices, recurse.
697 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
698 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
699 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
704 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
706 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
707 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
708 // If the GEPs only differ by one index, compare it.
709 unsigned NumDifferences = 0; // Keep track of # differences.
710 unsigned DiffOperand = 0; // The operand that differs.
711 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
712 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
713 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
714 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
715 // Irreconcilable differences.
719 if (NumDifferences++) break;
724 if (NumDifferences == 0) // SAME GEP?
725 return ReplaceInstUsesWith(I, // No comparison is needed here.
726 ConstantInt::get(Type::getInt1Ty(I.getContext()),
727 ICmpInst::isTrueWhenEqual(Cond)));
729 else if (NumDifferences == 1 && GEPsInBounds) {
730 Value *LHSV = GEPLHS->getOperand(DiffOperand);
731 Value *RHSV = GEPRHS->getOperand(DiffOperand);
732 // Make sure we do a signed comparison here.
733 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
737 // Only lower this if the icmp is the only user of the GEP or if we expect
738 // the result to fold to a constant!
741 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
742 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
743 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
744 Value *L = EmitGEPOffset(GEPLHS);
745 Value *R = EmitGEPOffset(GEPRHS);
746 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
752 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
753 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
754 Value *X, ConstantInt *CI,
755 ICmpInst::Predicate Pred,
757 // If we have X+0, exit early (simplifying logic below) and let it get folded
758 // elsewhere. icmp X+0, X -> icmp X, X
760 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
761 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
764 // (X+4) == X -> false.
765 if (Pred == ICmpInst::ICMP_EQ)
766 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
768 // (X+4) != X -> true.
769 if (Pred == ICmpInst::ICMP_NE)
770 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
772 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
773 // so the values can never be equal. Similarly for all other "or equals"
776 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
777 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
778 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
779 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
781 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
782 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
785 // (X+1) >u X --> X <u (0-1) --> X != 255
786 // (X+2) >u X --> X <u (0-2) --> X <u 254
787 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
788 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
789 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
791 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
792 ConstantInt *SMax = ConstantInt::get(X->getContext(),
793 APInt::getSignedMaxValue(BitWidth));
795 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
796 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
797 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
798 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
799 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
800 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
801 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
802 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
804 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
805 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
806 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
807 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
808 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
809 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
811 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
812 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
813 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
816 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
817 /// and CmpRHS are both known to be integer constants.
818 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
819 ConstantInt *DivRHS) {
820 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
821 const APInt &CmpRHSV = CmpRHS->getValue();
823 // FIXME: If the operand types don't match the type of the divide
824 // then don't attempt this transform. The code below doesn't have the
825 // logic to deal with a signed divide and an unsigned compare (and
826 // vice versa). This is because (x /s C1) <s C2 produces different
827 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
828 // (x /u C1) <u C2. Simply casting the operands and result won't
829 // work. :( The if statement below tests that condition and bails
831 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
832 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
834 if (DivRHS->isZero())
835 return 0; // The ProdOV computation fails on divide by zero.
836 if (DivIsSigned && DivRHS->isAllOnesValue())
837 return 0; // The overflow computation also screws up here
838 if (DivRHS->isOne()) {
839 // This eliminates some funny cases with INT_MIN.
840 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
844 // Compute Prod = CI * DivRHS. We are essentially solving an equation
845 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
846 // C2 (CI). By solving for X we can turn this into a range check
847 // instead of computing a divide.
848 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
850 // Determine if the product overflows by seeing if the product is
851 // not equal to the divide. Make sure we do the same kind of divide
852 // as in the LHS instruction that we're folding.
853 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
854 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
856 // Get the ICmp opcode
857 ICmpInst::Predicate Pred = ICI.getPredicate();
859 /// If the division is known to be exact, then there is no remainder from the
860 /// divide, so the covered range size is unit, otherwise it is the divisor.
861 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
863 // Figure out the interval that is being checked. For example, a comparison
864 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
865 // Compute this interval based on the constants involved and the signedness of
866 // the compare/divide. This computes a half-open interval, keeping track of
867 // whether either value in the interval overflows. After analysis each
868 // overflow variable is set to 0 if it's corresponding bound variable is valid
869 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
870 int LoOverflow = 0, HiOverflow = 0;
871 Constant *LoBound = 0, *HiBound = 0;
873 if (!DivIsSigned) { // udiv
874 // e.g. X/5 op 3 --> [15, 20)
876 HiOverflow = LoOverflow = ProdOV;
878 // If this is not an exact divide, then many values in the range collapse
879 // to the same result value.
880 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
883 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
884 if (CmpRHSV == 0) { // (X / pos) op 0
885 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
886 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
888 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
889 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
890 HiOverflow = LoOverflow = ProdOV;
892 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
893 } else { // (X / pos) op neg
894 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
895 HiBound = AddOne(Prod);
896 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
898 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
899 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
902 } else if (DivRHS->isNegative()) { // Divisor is < 0.
904 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
905 if (CmpRHSV == 0) { // (X / neg) op 0
906 // e.g. X/-5 op 0 --> [-4, 5)
907 LoBound = AddOne(RangeSize);
908 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
909 if (HiBound == DivRHS) { // -INTMIN = INTMIN
910 HiOverflow = 1; // [INTMIN+1, overflow)
911 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
913 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
914 // e.g. X/-5 op 3 --> [-19, -14)
915 HiBound = AddOne(Prod);
916 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
918 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
919 } else { // (X / neg) op neg
920 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
921 LoOverflow = HiOverflow = ProdOV;
923 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
926 // Dividing by a negative swaps the condition. LT <-> GT
927 Pred = ICmpInst::getSwappedPredicate(Pred);
930 Value *X = DivI->getOperand(0);
932 default: llvm_unreachable("Unhandled icmp opcode!");
933 case ICmpInst::ICMP_EQ:
934 if (LoOverflow && HiOverflow)
935 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
937 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
938 ICmpInst::ICMP_UGE, X, LoBound);
940 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
941 ICmpInst::ICMP_ULT, X, HiBound);
942 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
944 case ICmpInst::ICMP_NE:
945 if (LoOverflow && HiOverflow)
946 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
948 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
949 ICmpInst::ICMP_ULT, X, LoBound);
951 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
952 ICmpInst::ICMP_UGE, X, HiBound);
953 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
954 DivIsSigned, false));
955 case ICmpInst::ICMP_ULT:
956 case ICmpInst::ICMP_SLT:
957 if (LoOverflow == +1) // Low bound is greater than input range.
958 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
959 if (LoOverflow == -1) // Low bound is less than input range.
960 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
961 return new ICmpInst(Pred, X, LoBound);
962 case ICmpInst::ICMP_UGT:
963 case ICmpInst::ICMP_SGT:
964 if (HiOverflow == +1) // High bound greater than input range.
965 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
966 if (HiOverflow == -1) // High bound less than input range.
967 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
968 if (Pred == ICmpInst::ICMP_UGT)
969 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
970 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
974 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
975 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
976 ConstantInt *ShAmt) {
977 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
979 // Check that the shift amount is in range. If not, don't perform
980 // undefined shifts. When the shift is visited it will be
982 uint32_t TypeBits = CmpRHSV.getBitWidth();
983 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
984 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
987 if (!ICI.isEquality()) {
988 // If we have an unsigned comparison and an ashr, we can't simplify this.
989 // Similarly for signed comparisons with lshr.
990 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
993 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
994 // by a power of 2. Since we already have logic to simplify these,
995 // transform to div and then simplify the resultant comparison.
996 if (Shr->getOpcode() == Instruction::AShr &&
997 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
1000 // Revisit the shift (to delete it).
1004 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
1007 Shr->getOpcode() == Instruction::AShr ?
1008 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
1009 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
1011 ICI.setOperand(0, Tmp);
1013 // If the builder folded the binop, just return it.
1014 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
1018 // Otherwise, fold this div/compare.
1019 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1020 TheDiv->getOpcode() == Instruction::UDiv);
1022 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1023 assert(Res && "This div/cst should have folded!");
1028 // If we are comparing against bits always shifted out, the
1029 // comparison cannot succeed.
1030 APInt Comp = CmpRHSV << ShAmtVal;
1031 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
1032 if (Shr->getOpcode() == Instruction::LShr)
1033 Comp = Comp.lshr(ShAmtVal);
1035 Comp = Comp.ashr(ShAmtVal);
1037 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1038 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1039 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1041 return ReplaceInstUsesWith(ICI, Cst);
1044 // Otherwise, check to see if the bits shifted out are known to be zero.
1045 // If so, we can compare against the unshifted value:
1046 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1047 if (Shr->hasOneUse() && Shr->isExact())
1048 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1050 if (Shr->hasOneUse()) {
1051 // Otherwise strength reduce the shift into an and.
1052 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1053 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
1055 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1056 Mask, Shr->getName()+".mask");
1057 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1063 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1065 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1068 const APInt &RHSV = RHS->getValue();
1070 switch (LHSI->getOpcode()) {
1071 case Instruction::Trunc:
1072 if (ICI.isEquality() && LHSI->hasOneUse()) {
1073 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1074 // of the high bits truncated out of x are known.
1075 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1076 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1077 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1078 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1080 // If all the high bits are known, we can do this xform.
1081 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1082 // Pull in the high bits from known-ones set.
1083 APInt NewRHS = RHS->getValue().zext(SrcBits);
1084 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1085 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1086 ConstantInt::get(ICI.getContext(), NewRHS));
1091 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1092 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1093 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1095 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1096 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1097 Value *CompareVal = LHSI->getOperand(0);
1099 // If the sign bit of the XorCST is not set, there is no change to
1100 // the operation, just stop using the Xor.
1101 if (!XorCST->isNegative()) {
1102 ICI.setOperand(0, CompareVal);
1107 // Was the old condition true if the operand is positive?
1108 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1110 // If so, the new one isn't.
1111 isTrueIfPositive ^= true;
1113 if (isTrueIfPositive)
1114 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1117 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1121 if (LHSI->hasOneUse()) {
1122 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1123 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1124 const APInt &SignBit = XorCST->getValue();
1125 ICmpInst::Predicate Pred = ICI.isSigned()
1126 ? ICI.getUnsignedPredicate()
1127 : ICI.getSignedPredicate();
1128 return new ICmpInst(Pred, LHSI->getOperand(0),
1129 ConstantInt::get(ICI.getContext(),
1133 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1134 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1135 const APInt &NotSignBit = XorCST->getValue();
1136 ICmpInst::Predicate Pred = ICI.isSigned()
1137 ? ICI.getUnsignedPredicate()
1138 : ICI.getSignedPredicate();
1139 Pred = ICI.getSwappedPredicate(Pred);
1140 return new ICmpInst(Pred, LHSI->getOperand(0),
1141 ConstantInt::get(ICI.getContext(),
1142 RHSV ^ NotSignBit));
1147 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1148 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1149 LHSI->getOperand(0)->hasOneUse()) {
1150 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1152 // If the LHS is an AND of a truncating cast, we can widen the
1153 // and/compare to be the input width without changing the value
1154 // produced, eliminating a cast.
1155 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1156 // We can do this transformation if either the AND constant does not
1157 // have its sign bit set or if it is an equality comparison.
1158 // Extending a relational comparison when we're checking the sign
1159 // bit would not work.
1160 if (ICI.isEquality() ||
1161 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1163 Builder->CreateAnd(Cast->getOperand(0),
1164 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1165 NewAnd->takeName(LHSI);
1166 return new ICmpInst(ICI.getPredicate(), NewAnd,
1167 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1171 // If the LHS is an AND of a zext, and we have an equality compare, we can
1172 // shrink the and/compare to the smaller type, eliminating the cast.
1173 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1174 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1175 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1176 // should fold the icmp to true/false in that case.
1177 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1179 Builder->CreateAnd(Cast->getOperand(0),
1180 ConstantExpr::getTrunc(AndCST, Ty));
1181 NewAnd->takeName(LHSI);
1182 return new ICmpInst(ICI.getPredicate(), NewAnd,
1183 ConstantExpr::getTrunc(RHS, Ty));
1187 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1188 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1189 // happens a LOT in code produced by the C front-end, for bitfield
1191 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1192 if (Shift && !Shift->isShift())
1196 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1197 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1198 Type *AndTy = AndCST->getType(); // Type of the and.
1200 // We can fold this as long as we can't shift unknown bits
1201 // into the mask. This can only happen with signed shift
1202 // rights, as they sign-extend.
1204 bool CanFold = Shift->isLogicalShift();
1206 // To test for the bad case of the signed shr, see if any
1207 // of the bits shifted in could be tested after the mask.
1208 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1209 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1211 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1212 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1213 AndCST->getValue()) == 0)
1219 if (Shift->getOpcode() == Instruction::Shl)
1220 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1222 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1224 // Check to see if we are shifting out any of the bits being
1226 if (ConstantExpr::get(Shift->getOpcode(),
1227 NewCst, ShAmt) != RHS) {
1228 // If we shifted bits out, the fold is not going to work out.
1229 // As a special case, check to see if this means that the
1230 // result is always true or false now.
1231 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1232 return ReplaceInstUsesWith(ICI,
1233 ConstantInt::getFalse(ICI.getContext()));
1234 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1235 return ReplaceInstUsesWith(ICI,
1236 ConstantInt::getTrue(ICI.getContext()));
1238 ICI.setOperand(1, NewCst);
1239 Constant *NewAndCST;
1240 if (Shift->getOpcode() == Instruction::Shl)
1241 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1243 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1244 LHSI->setOperand(1, NewAndCST);
1245 LHSI->setOperand(0, Shift->getOperand(0));
1246 Worklist.Add(Shift); // Shift is dead.
1252 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1253 // preferable because it allows the C<<Y expression to be hoisted out
1254 // of a loop if Y is invariant and X is not.
1255 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1256 ICI.isEquality() && !Shift->isArithmeticShift() &&
1257 !isa<Constant>(Shift->getOperand(0))) {
1260 if (Shift->getOpcode() == Instruction::LShr) {
1261 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1263 // Insert a logical shift.
1264 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1267 // Compute X & (C << Y).
1269 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1271 ICI.setOperand(0, NewAnd);
1275 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
1276 // bit set in (X & AndCST) will produce a result greater than RHSV.
1277 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1278 unsigned NTZ = AndCST->getValue().countTrailingZeros();
1279 if ((NTZ < AndCST->getBitWidth()) &&
1280 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
1281 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1282 Constant::getNullValue(RHS->getType()));
1286 // Try to optimize things like "A[i]&42 == 0" to index computations.
1287 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1288 if (GetElementPtrInst *GEP =
1289 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1290 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1291 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1292 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1293 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1294 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1300 case Instruction::Or: {
1301 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1304 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1305 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1306 // -> and (icmp eq P, null), (icmp eq Q, null).
1307 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1308 Constant::getNullValue(P->getType()));
1309 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1310 Constant::getNullValue(Q->getType()));
1312 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1313 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1315 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1321 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1322 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1325 if (!ICI.isEquality()) {
1326 // If this is a signed comparison to 0 and the mul is sign preserving,
1327 // use the mul LHS operand instead.
1328 ICmpInst::Predicate pred = ICI.getPredicate();
1329 if (isSignTest(pred, RHS) && !Val->isZero() &&
1330 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1331 return new ICmpInst(Val->isNegative() ?
1332 ICmpInst::getSwappedPredicate(pred) : pred,
1333 LHSI->getOperand(0),
1334 Constant::getNullValue(RHS->getType()));
1340 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1341 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1344 uint32_t TypeBits = RHSV.getBitWidth();
1346 // Check that the shift amount is in range. If not, don't perform
1347 // undefined shifts. When the shift is visited it will be
1349 if (ShAmt->uge(TypeBits))
1352 if (ICI.isEquality()) {
1353 // If we are comparing against bits always shifted out, the
1354 // comparison cannot succeed.
1356 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1358 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1359 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1361 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1362 return ReplaceInstUsesWith(ICI, Cst);
1365 // If the shift is NUW, then it is just shifting out zeros, no need for an
1367 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1368 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1369 ConstantExpr::getLShr(RHS, ShAmt));
1371 // If the shift is NSW and we compare to 0, then it is just shifting out
1372 // sign bits, no need for an AND either.
1373 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1374 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1375 ConstantExpr::getLShr(RHS, ShAmt));
1377 if (LHSI->hasOneUse()) {
1378 // Otherwise strength reduce the shift into an and.
1379 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1381 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1382 TypeBits-ShAmtVal));
1385 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1386 return new ICmpInst(ICI.getPredicate(), And,
1387 ConstantExpr::getLShr(RHS, ShAmt));
1391 // If this is a signed comparison to 0 and the shift is sign preserving,
1392 // use the shift LHS operand instead.
1393 ICmpInst::Predicate pred = ICI.getPredicate();
1394 if (isSignTest(pred, RHS) &&
1395 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1396 return new ICmpInst(pred,
1397 LHSI->getOperand(0),
1398 Constant::getNullValue(RHS->getType()));
1400 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1401 bool TrueIfSigned = false;
1402 if (LHSI->hasOneUse() &&
1403 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1404 // (X << 31) <s 0 --> (X&1) != 0
1405 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1406 APInt::getOneBitSet(TypeBits,
1407 TypeBits-ShAmt->getZExtValue()-1));
1409 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1410 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1411 And, Constant::getNullValue(And->getType()));
1414 // Transform (icmp pred iM (shl iM %v, N), CI)
1415 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1416 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1417 // This enables to get rid of the shift in favor of a trunc which can be
1418 // free on the target. It has the additional benefit of comparing to a
1419 // smaller constant, which will be target friendly.
1420 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1421 if (LHSI->hasOneUse() &&
1422 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1423 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1424 Constant *NCI = ConstantExpr::getTrunc(
1425 ConstantExpr::getAShr(RHS,
1426 ConstantInt::get(RHS->getType(), Amt)),
1428 return new ICmpInst(ICI.getPredicate(),
1429 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1436 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1437 case Instruction::AShr: {
1438 // Handle equality comparisons of shift-by-constant.
1439 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1440 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1441 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1445 // Handle exact shr's.
1446 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1447 if (RHSV.isMinValue())
1448 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1453 case Instruction::SDiv:
1454 case Instruction::UDiv:
1455 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1456 // Fold this div into the comparison, producing a range check.
1457 // Determine, based on the divide type, what the range is being
1458 // checked. If there is an overflow on the low or high side, remember
1459 // it, otherwise compute the range [low, hi) bounding the new value.
1460 // See: InsertRangeTest above for the kinds of replacements possible.
1461 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1462 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1467 case Instruction::Add:
1468 // Fold: icmp pred (add X, C1), C2
1469 if (!ICI.isEquality()) {
1470 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1472 const APInt &LHSV = LHSC->getValue();
1474 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1477 if (ICI.isSigned()) {
1478 if (CR.getLower().isSignBit()) {
1479 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1480 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1481 } else if (CR.getUpper().isSignBit()) {
1482 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1483 ConstantInt::get(ICI.getContext(),CR.getLower()));
1486 if (CR.getLower().isMinValue()) {
1487 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1488 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1489 } else if (CR.getUpper().isMinValue()) {
1490 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1491 ConstantInt::get(ICI.getContext(),CR.getLower()));
1498 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1499 if (ICI.isEquality()) {
1500 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1502 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1503 // the second operand is a constant, simplify a bit.
1504 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1505 switch (BO->getOpcode()) {
1506 case Instruction::SRem:
1507 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1508 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1509 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1510 if (V.sgt(1) && V.isPowerOf2()) {
1512 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1514 return new ICmpInst(ICI.getPredicate(), NewRem,
1515 Constant::getNullValue(BO->getType()));
1519 case Instruction::Add:
1520 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1521 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1522 if (BO->hasOneUse())
1523 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1524 ConstantExpr::getSub(RHS, BOp1C));
1525 } else if (RHSV == 0) {
1526 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1527 // efficiently invertible, or if the add has just this one use.
1528 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1530 if (Value *NegVal = dyn_castNegVal(BOp1))
1531 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1532 if (Value *NegVal = dyn_castNegVal(BOp0))
1533 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1534 if (BO->hasOneUse()) {
1535 Value *Neg = Builder->CreateNeg(BOp1);
1537 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1541 case Instruction::Xor:
1542 // For the xor case, we can xor two constants together, eliminating
1543 // the explicit xor.
1544 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1545 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1546 ConstantExpr::getXor(RHS, BOC));
1547 } else if (RHSV == 0) {
1548 // Replace ((xor A, B) != 0) with (A != B)
1549 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1553 case Instruction::Sub:
1554 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1555 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1556 if (BO->hasOneUse())
1557 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1558 ConstantExpr::getSub(BOp0C, RHS));
1559 } else if (RHSV == 0) {
1560 // Replace ((sub A, B) != 0) with (A != B)
1561 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1565 case Instruction::Or:
1566 // If bits are being or'd in that are not present in the constant we
1567 // are comparing against, then the comparison could never succeed!
1568 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1569 Constant *NotCI = ConstantExpr::getNot(RHS);
1570 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1571 return ReplaceInstUsesWith(ICI,
1572 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1577 case Instruction::And:
1578 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1579 // If bits are being compared against that are and'd out, then the
1580 // comparison can never succeed!
1581 if ((RHSV & ~BOC->getValue()) != 0)
1582 return ReplaceInstUsesWith(ICI,
1583 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1586 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1587 if (RHS == BOC && RHSV.isPowerOf2())
1588 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1589 ICmpInst::ICMP_NE, LHSI,
1590 Constant::getNullValue(RHS->getType()));
1592 // Don't perform the following transforms if the AND has multiple uses
1593 if (!BO->hasOneUse())
1596 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1597 if (BOC->getValue().isSignBit()) {
1598 Value *X = BO->getOperand(0);
1599 Constant *Zero = Constant::getNullValue(X->getType());
1600 ICmpInst::Predicate pred = isICMP_NE ?
1601 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1602 return new ICmpInst(pred, X, Zero);
1605 // ((X & ~7) == 0) --> X < 8
1606 if (RHSV == 0 && isHighOnes(BOC)) {
1607 Value *X = BO->getOperand(0);
1608 Constant *NegX = ConstantExpr::getNeg(BOC);
1609 ICmpInst::Predicate pred = isICMP_NE ?
1610 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1611 return new ICmpInst(pred, X, NegX);
1615 case Instruction::Mul:
1617 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1618 // The trivial case (mul X, 0) is handled by InstSimplify
1619 // General case : (mul X, C) != 0 iff X != 0
1620 // (mul X, C) == 0 iff X == 0
1622 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1623 Constant::getNullValue(RHS->getType()));
1629 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1630 // Handle icmp {eq|ne} <intrinsic>, intcst.
1631 switch (II->getIntrinsicID()) {
1632 case Intrinsic::bswap:
1634 ICI.setOperand(0, II->getArgOperand(0));
1635 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1637 case Intrinsic::ctlz:
1638 case Intrinsic::cttz:
1639 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1640 if (RHSV == RHS->getType()->getBitWidth()) {
1642 ICI.setOperand(0, II->getArgOperand(0));
1643 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1647 case Intrinsic::ctpop:
1648 // popcount(A) == 0 -> A == 0 and likewise for !=
1649 if (RHS->isZero()) {
1651 ICI.setOperand(0, II->getArgOperand(0));
1652 ICI.setOperand(1, RHS);
1664 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1665 /// We only handle extending casts so far.
1667 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1668 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1669 Value *LHSCIOp = LHSCI->getOperand(0);
1670 Type *SrcTy = LHSCIOp->getType();
1671 Type *DestTy = LHSCI->getType();
1674 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1675 // integer type is the same size as the pointer type.
1676 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1677 TD->getPointerSizeInBits() ==
1678 cast<IntegerType>(DestTy)->getBitWidth()) {
1680 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1681 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1682 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1683 RHSOp = RHSC->getOperand(0);
1684 // If the pointer types don't match, insert a bitcast.
1685 if (LHSCIOp->getType() != RHSOp->getType())
1686 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1690 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1693 // The code below only handles extension cast instructions, so far.
1695 if (LHSCI->getOpcode() != Instruction::ZExt &&
1696 LHSCI->getOpcode() != Instruction::SExt)
1699 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1700 bool isSignedCmp = ICI.isSigned();
1702 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1703 // Not an extension from the same type?
1704 RHSCIOp = CI->getOperand(0);
1705 if (RHSCIOp->getType() != LHSCIOp->getType())
1708 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1709 // and the other is a zext), then we can't handle this.
1710 if (CI->getOpcode() != LHSCI->getOpcode())
1713 // Deal with equality cases early.
1714 if (ICI.isEquality())
1715 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1717 // A signed comparison of sign extended values simplifies into a
1718 // signed comparison.
1719 if (isSignedCmp && isSignedExt)
1720 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1722 // The other three cases all fold into an unsigned comparison.
1723 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1726 // If we aren't dealing with a constant on the RHS, exit early
1727 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1731 // Compute the constant that would happen if we truncated to SrcTy then
1732 // reextended to DestTy.
1733 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1734 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1737 // If the re-extended constant didn't change...
1739 // Deal with equality cases early.
1740 if (ICI.isEquality())
1741 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1743 // A signed comparison of sign extended values simplifies into a
1744 // signed comparison.
1745 if (isSignedExt && isSignedCmp)
1746 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1748 // The other three cases all fold into an unsigned comparison.
1749 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1752 // The re-extended constant changed so the constant cannot be represented
1753 // in the shorter type. Consequently, we cannot emit a simple comparison.
1754 // All the cases that fold to true or false will have already been handled
1755 // by SimplifyICmpInst, so only deal with the tricky case.
1757 if (isSignedCmp || !isSignedExt)
1760 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1761 // should have been folded away previously and not enter in here.
1763 // We're performing an unsigned comp with a sign extended value.
1764 // This is true if the input is >= 0. [aka >s -1]
1765 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1766 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1768 // Finally, return the value computed.
1769 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1770 return ReplaceInstUsesWith(ICI, Result);
1772 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1773 return BinaryOperator::CreateNot(Result);
1776 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1777 /// I = icmp ugt (add (add A, B), CI2), CI1
1778 /// If this is of the form:
1780 /// if (sum+128 >u 255)
1781 /// Then replace it with llvm.sadd.with.overflow.i8.
1783 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1784 ConstantInt *CI2, ConstantInt *CI1,
1786 // The transformation we're trying to do here is to transform this into an
1787 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1788 // with a narrower add, and discard the add-with-constant that is part of the
1789 // range check (if we can't eliminate it, this isn't profitable).
1791 // In order to eliminate the add-with-constant, the compare can be its only
1793 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1794 if (!AddWithCst->hasOneUse()) return 0;
1796 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1797 if (!CI2->getValue().isPowerOf2()) return 0;
1798 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1799 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1801 // The width of the new add formed is 1 more than the bias.
1804 // Check to see that CI1 is an all-ones value with NewWidth bits.
1805 if (CI1->getBitWidth() == NewWidth ||
1806 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1809 // This is only really a signed overflow check if the inputs have been
1810 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1811 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1812 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1813 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1814 IC.ComputeNumSignBits(B) < NeededSignBits)
1817 // In order to replace the original add with a narrower
1818 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1819 // and truncates that discard the high bits of the add. Verify that this is
1821 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1822 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1824 if (*UI == AddWithCst) continue;
1826 // Only accept truncates for now. We would really like a nice recursive
1827 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1828 // chain to see which bits of a value are actually demanded. If the
1829 // original add had another add which was then immediately truncated, we
1830 // could still do the transformation.
1831 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1833 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1836 // If the pattern matches, truncate the inputs to the narrower type and
1837 // use the sadd_with_overflow intrinsic to efficiently compute both the
1838 // result and the overflow bit.
1839 Module *M = I.getParent()->getParent()->getParent();
1841 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1842 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1845 InstCombiner::BuilderTy *Builder = IC.Builder;
1847 // Put the new code above the original add, in case there are any uses of the
1848 // add between the add and the compare.
1849 Builder->SetInsertPoint(OrigAdd);
1851 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1852 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1853 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1854 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1855 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1857 // The inner add was the result of the narrow add, zero extended to the
1858 // wider type. Replace it with the result computed by the intrinsic.
1859 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1861 // The original icmp gets replaced with the overflow value.
1862 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1865 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1867 // Don't bother doing this transformation for pointers, don't do it for
1869 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1871 // If the add is a constant expr, then we don't bother transforming it.
1872 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1873 if (OrigAdd == 0) return 0;
1875 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1877 // Put the new code above the original add, in case there are any uses of the
1878 // add between the add and the compare.
1879 InstCombiner::BuilderTy *Builder = IC.Builder;
1880 Builder->SetInsertPoint(OrigAdd);
1882 Module *M = I.getParent()->getParent()->getParent();
1883 Type *Ty = LHS->getType();
1884 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1885 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1886 Value *Add = Builder->CreateExtractValue(Call, 0);
1888 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1890 // The original icmp gets replaced with the overflow value.
1891 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1894 // DemandedBitsLHSMask - When performing a comparison against a constant,
1895 // it is possible that not all the bits in the LHS are demanded. This helper
1896 // method computes the mask that IS demanded.
1897 static APInt DemandedBitsLHSMask(ICmpInst &I,
1898 unsigned BitWidth, bool isSignCheck) {
1900 return APInt::getSignBit(BitWidth);
1902 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1903 if (!CI) return APInt::getAllOnesValue(BitWidth);
1904 const APInt &RHS = CI->getValue();
1906 switch (I.getPredicate()) {
1907 // For a UGT comparison, we don't care about any bits that
1908 // correspond to the trailing ones of the comparand. The value of these
1909 // bits doesn't impact the outcome of the comparison, because any value
1910 // greater than the RHS must differ in a bit higher than these due to carry.
1911 case ICmpInst::ICMP_UGT: {
1912 unsigned trailingOnes = RHS.countTrailingOnes();
1913 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1917 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1918 // Any value less than the RHS must differ in a higher bit because of carries.
1919 case ICmpInst::ICMP_ULT: {
1920 unsigned trailingZeros = RHS.countTrailingZeros();
1921 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1926 return APInt::getAllOnesValue(BitWidth);
1931 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1932 bool Changed = false;
1933 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1935 /// Orders the operands of the compare so that they are listed from most
1936 /// complex to least complex. This puts constants before unary operators,
1937 /// before binary operators.
1938 if (getComplexity(Op0) < getComplexity(Op1)) {
1940 std::swap(Op0, Op1);
1944 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1945 return ReplaceInstUsesWith(I, V);
1947 // comparing -val or val with non-zero is the same as just comparing val
1948 // ie, abs(val) != 0 -> val != 0
1949 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
1951 Value *Cond, *SelectTrue, *SelectFalse;
1952 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
1953 m_Value(SelectFalse)))) {
1954 if (Value *V = dyn_castNegVal(SelectTrue)) {
1955 if (V == SelectFalse)
1956 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1958 else if (Value *V = dyn_castNegVal(SelectFalse)) {
1959 if (V == SelectTrue)
1960 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1965 Type *Ty = Op0->getType();
1967 // icmp's with boolean values can always be turned into bitwise operations
1968 if (Ty->isIntegerTy(1)) {
1969 switch (I.getPredicate()) {
1970 default: llvm_unreachable("Invalid icmp instruction!");
1971 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1972 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1973 return BinaryOperator::CreateNot(Xor);
1975 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1976 return BinaryOperator::CreateXor(Op0, Op1);
1978 case ICmpInst::ICMP_UGT:
1979 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1981 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1982 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1983 return BinaryOperator::CreateAnd(Not, Op1);
1985 case ICmpInst::ICMP_SGT:
1986 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1988 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1989 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1990 return BinaryOperator::CreateAnd(Not, Op0);
1992 case ICmpInst::ICMP_UGE:
1993 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1995 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1996 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1997 return BinaryOperator::CreateOr(Not, Op1);
1999 case ICmpInst::ICMP_SGE:
2000 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2002 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2003 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2004 return BinaryOperator::CreateOr(Not, Op0);
2009 unsigned BitWidth = 0;
2010 if (Ty->isIntOrIntVectorTy())
2011 BitWidth = Ty->getScalarSizeInBits();
2012 else if (TD) // Pointers require TD info to get their size.
2013 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
2015 bool isSignBit = false;
2017 // See if we are doing a comparison with a constant.
2018 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2019 Value *A = 0, *B = 0;
2021 // Match the following pattern, which is a common idiom when writing
2022 // overflow-safe integer arithmetic function. The source performs an
2023 // addition in wider type, and explicitly checks for overflow using
2024 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2025 // sadd_with_overflow intrinsic.
2027 // TODO: This could probably be generalized to handle other overflow-safe
2028 // operations if we worked out the formulas to compute the appropriate
2032 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2034 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2035 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2036 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2037 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2041 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2042 if (I.isEquality() && CI->isZero() &&
2043 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2044 // (icmp cond A B) if cond is equality
2045 return new ICmpInst(I.getPredicate(), A, B);
2048 // If we have an icmp le or icmp ge instruction, turn it into the
2049 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2050 // them being folded in the code below. The SimplifyICmpInst code has
2051 // already handled the edge cases for us, so we just assert on them.
2052 switch (I.getPredicate()) {
2054 case ICmpInst::ICMP_ULE:
2055 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2056 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2057 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2058 case ICmpInst::ICMP_SLE:
2059 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2060 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2061 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2062 case ICmpInst::ICMP_UGE:
2063 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2064 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2065 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2066 case ICmpInst::ICMP_SGE:
2067 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2068 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2069 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2072 // If this comparison is a normal comparison, it demands all
2073 // bits, if it is a sign bit comparison, it only demands the sign bit.
2075 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2078 // See if we can fold the comparison based on range information we can get
2079 // by checking whether bits are known to be zero or one in the input.
2080 if (BitWidth != 0) {
2081 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2082 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2084 if (SimplifyDemandedBits(I.getOperandUse(0),
2085 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2086 Op0KnownZero, Op0KnownOne, 0))
2088 if (SimplifyDemandedBits(I.getOperandUse(1),
2089 APInt::getAllOnesValue(BitWidth),
2090 Op1KnownZero, Op1KnownOne, 0))
2093 // Given the known and unknown bits, compute a range that the LHS could be
2094 // in. Compute the Min, Max and RHS values based on the known bits. For the
2095 // EQ and NE we use unsigned values.
2096 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2097 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2099 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2101 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2104 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2106 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2110 // If Min and Max are known to be the same, then SimplifyDemandedBits
2111 // figured out that the LHS is a constant. Just constant fold this now so
2112 // that code below can assume that Min != Max.
2113 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2114 return new ICmpInst(I.getPredicate(),
2115 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2116 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2117 return new ICmpInst(I.getPredicate(), Op0,
2118 ConstantInt::get(Op1->getType(), Op1Min));
2120 // Based on the range information we know about the LHS, see if we can
2121 // simplify this comparison. For example, (x&4) < 8 is always true.
2122 switch (I.getPredicate()) {
2123 default: llvm_unreachable("Unknown icmp opcode!");
2124 case ICmpInst::ICMP_EQ: {
2125 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2126 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2128 // If all bits are known zero except for one, then we know at most one
2129 // bit is set. If the comparison is against zero, then this is a check
2130 // to see if *that* bit is set.
2131 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2132 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2133 // If the LHS is an AND with the same constant, look through it.
2135 ConstantInt *LHSC = 0;
2136 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2137 LHSC->getValue() != Op0KnownZeroInverted)
2140 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2141 // then turn "((1 << x)&8) == 0" into "x != 3".
2143 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2144 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2145 return new ICmpInst(ICmpInst::ICMP_NE, X,
2146 ConstantInt::get(X->getType(), CmpVal));
2149 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2150 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2152 if (Op0KnownZeroInverted == 1 &&
2153 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2154 return new ICmpInst(ICmpInst::ICMP_NE, X,
2155 ConstantInt::get(X->getType(),
2156 CI->countTrailingZeros()));
2161 case ICmpInst::ICMP_NE: {
2162 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2163 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2165 // If all bits are known zero except for one, then we know at most one
2166 // bit is set. If the comparison is against zero, then this is a check
2167 // to see if *that* bit is set.
2168 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2169 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2170 // If the LHS is an AND with the same constant, look through it.
2172 ConstantInt *LHSC = 0;
2173 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2174 LHSC->getValue() != Op0KnownZeroInverted)
2177 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2178 // then turn "((1 << x)&8) != 0" into "x == 3".
2180 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2181 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2182 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2183 ConstantInt::get(X->getType(), CmpVal));
2186 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2187 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2189 if (Op0KnownZeroInverted == 1 &&
2190 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2191 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2192 ConstantInt::get(X->getType(),
2193 CI->countTrailingZeros()));
2198 case ICmpInst::ICMP_ULT:
2199 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2200 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2201 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2202 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2203 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2204 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2205 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2206 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2207 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2208 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2210 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2211 if (CI->isMinValue(true))
2212 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2213 Constant::getAllOnesValue(Op0->getType()));
2216 case ICmpInst::ICMP_UGT:
2217 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2218 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2219 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2220 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2222 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2223 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2224 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2225 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2226 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2227 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2229 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2230 if (CI->isMaxValue(true))
2231 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2232 Constant::getNullValue(Op0->getType()));
2235 case ICmpInst::ICMP_SLT:
2236 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2237 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2238 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2239 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2240 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2241 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2242 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2243 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2244 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2245 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2248 case ICmpInst::ICMP_SGT:
2249 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2250 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2251 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2252 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2254 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2255 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2256 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2257 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2258 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2259 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2262 case ICmpInst::ICMP_SGE:
2263 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2264 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2265 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2266 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2267 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2269 case ICmpInst::ICMP_SLE:
2270 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2271 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2272 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2273 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2274 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2276 case ICmpInst::ICMP_UGE:
2277 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2278 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2279 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2280 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2281 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2283 case ICmpInst::ICMP_ULE:
2284 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2285 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2286 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2287 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2288 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2292 // Turn a signed comparison into an unsigned one if both operands
2293 // are known to have the same sign.
2295 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2296 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2297 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2300 // Test if the ICmpInst instruction is used exclusively by a select as
2301 // part of a minimum or maximum operation. If so, refrain from doing
2302 // any other folding. This helps out other analyses which understand
2303 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2304 // and CodeGen. And in this case, at least one of the comparison
2305 // operands has at least one user besides the compare (the select),
2306 // which would often largely negate the benefit of folding anyway.
2308 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2309 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2310 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2313 // See if we are doing a comparison between a constant and an instruction that
2314 // can be folded into the comparison.
2315 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2316 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2317 // instruction, see if that instruction also has constants so that the
2318 // instruction can be folded into the icmp
2319 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2320 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2324 // Handle icmp with constant (but not simple integer constant) RHS
2325 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2326 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2327 switch (LHSI->getOpcode()) {
2328 case Instruction::GetElementPtr:
2329 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2330 if (RHSC->isNullValue() &&
2331 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2332 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2333 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2335 case Instruction::PHI:
2336 // Only fold icmp into the PHI if the phi and icmp are in the same
2337 // block. If in the same block, we're encouraging jump threading. If
2338 // not, we are just pessimizing the code by making an i1 phi.
2339 if (LHSI->getParent() == I.getParent())
2340 if (Instruction *NV = FoldOpIntoPhi(I))
2343 case Instruction::Select: {
2344 // If either operand of the select is a constant, we can fold the
2345 // comparison into the select arms, which will cause one to be
2346 // constant folded and the select turned into a bitwise or.
2347 Value *Op1 = 0, *Op2 = 0;
2348 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2349 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2350 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2351 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2353 // We only want to perform this transformation if it will not lead to
2354 // additional code. This is true if either both sides of the select
2355 // fold to a constant (in which case the icmp is replaced with a select
2356 // which will usually simplify) or this is the only user of the
2357 // select (in which case we are trading a select+icmp for a simpler
2359 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2361 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2364 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2366 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2370 case Instruction::IntToPtr:
2371 // icmp pred inttoptr(X), null -> icmp pred X, 0
2372 if (RHSC->isNullValue() && TD &&
2373 TD->getIntPtrType(RHSC->getContext()) ==
2374 LHSI->getOperand(0)->getType())
2375 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2376 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2379 case Instruction::Load:
2380 // Try to optimize things like "A[i] > 4" to index computations.
2381 if (GetElementPtrInst *GEP =
2382 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2383 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2384 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2385 !cast<LoadInst>(LHSI)->isVolatile())
2386 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2393 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2394 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2395 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2397 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2398 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2399 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2402 // Test to see if the operands of the icmp are casted versions of other
2403 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2405 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2406 if (Op0->getType()->isPointerTy() &&
2407 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2408 // We keep moving the cast from the left operand over to the right
2409 // operand, where it can often be eliminated completely.
2410 Op0 = CI->getOperand(0);
2412 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2413 // so eliminate it as well.
2414 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2415 Op1 = CI2->getOperand(0);
2417 // If Op1 is a constant, we can fold the cast into the constant.
2418 if (Op0->getType() != Op1->getType()) {
2419 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2420 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2422 // Otherwise, cast the RHS right before the icmp
2423 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2426 return new ICmpInst(I.getPredicate(), Op0, Op1);
2430 if (isa<CastInst>(Op0)) {
2431 // Handle the special case of: icmp (cast bool to X), <cst>
2432 // This comes up when you have code like
2435 // For generality, we handle any zero-extension of any operand comparison
2436 // with a constant or another cast from the same type.
2437 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2438 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2442 // Special logic for binary operators.
2443 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2444 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2446 CmpInst::Predicate Pred = I.getPredicate();
2447 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2448 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2449 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2450 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2451 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2452 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2453 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2454 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2455 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2457 // Analyze the case when either Op0 or Op1 is an add instruction.
2458 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2459 Value *A = 0, *B = 0, *C = 0, *D = 0;
2460 if (BO0 && BO0->getOpcode() == Instruction::Add)
2461 A = BO0->getOperand(0), B = BO0->getOperand(1);
2462 if (BO1 && BO1->getOpcode() == Instruction::Add)
2463 C = BO1->getOperand(0), D = BO1->getOperand(1);
2465 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2466 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2467 return new ICmpInst(Pred, A == Op1 ? B : A,
2468 Constant::getNullValue(Op1->getType()));
2470 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2471 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2472 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2475 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2476 if (A && C && (A == C || A == D || B == C || B == D) &&
2477 NoOp0WrapProblem && NoOp1WrapProblem &&
2478 // Try not to increase register pressure.
2479 BO0->hasOneUse() && BO1->hasOneUse()) {
2480 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2483 // C + B == C + D -> B == D
2486 } else if (A == D) {
2487 // D + B == C + D -> B == C
2490 } else if (B == C) {
2491 // A + C == C + D -> A == D
2496 // A + D == C + D -> A == C
2500 return new ICmpInst(Pred, Y, Z);
2503 // Analyze the case when either Op0 or Op1 is a sub instruction.
2504 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2505 A = 0; B = 0; C = 0; D = 0;
2506 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2507 A = BO0->getOperand(0), B = BO0->getOperand(1);
2508 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2509 C = BO1->getOperand(0), D = BO1->getOperand(1);
2511 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2512 if (A == Op1 && NoOp0WrapProblem)
2513 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2515 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2516 if (C == Op0 && NoOp1WrapProblem)
2517 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2519 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2520 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2521 // Try not to increase register pressure.
2522 BO0->hasOneUse() && BO1->hasOneUse())
2523 return new ICmpInst(Pred, A, C);
2525 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2526 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2527 // Try not to increase register pressure.
2528 BO0->hasOneUse() && BO1->hasOneUse())
2529 return new ICmpInst(Pred, D, B);
2531 BinaryOperator *SRem = NULL;
2532 // icmp (srem X, Y), Y
2533 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2534 Op1 == BO0->getOperand(1))
2536 // icmp Y, (srem X, Y)
2537 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2538 Op0 == BO1->getOperand(1))
2541 // We don't check hasOneUse to avoid increasing register pressure because
2542 // the value we use is the same value this instruction was already using.
2543 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2545 case ICmpInst::ICMP_EQ:
2546 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2547 case ICmpInst::ICMP_NE:
2548 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2549 case ICmpInst::ICMP_SGT:
2550 case ICmpInst::ICMP_SGE:
2551 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2552 Constant::getAllOnesValue(SRem->getType()));
2553 case ICmpInst::ICMP_SLT:
2554 case ICmpInst::ICMP_SLE:
2555 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2556 Constant::getNullValue(SRem->getType()));
2560 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2561 BO0->hasOneUse() && BO1->hasOneUse() &&
2562 BO0->getOperand(1) == BO1->getOperand(1)) {
2563 switch (BO0->getOpcode()) {
2565 case Instruction::Add:
2566 case Instruction::Sub:
2567 case Instruction::Xor:
2568 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2569 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2570 BO1->getOperand(0));
2571 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2572 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2573 if (CI->getValue().isSignBit()) {
2574 ICmpInst::Predicate Pred = I.isSigned()
2575 ? I.getUnsignedPredicate()
2576 : I.getSignedPredicate();
2577 return new ICmpInst(Pred, BO0->getOperand(0),
2578 BO1->getOperand(0));
2581 if (CI->isMaxValue(true)) {
2582 ICmpInst::Predicate Pred = I.isSigned()
2583 ? I.getUnsignedPredicate()
2584 : I.getSignedPredicate();
2585 Pred = I.getSwappedPredicate(Pred);
2586 return new ICmpInst(Pred, BO0->getOperand(0),
2587 BO1->getOperand(0));
2591 case Instruction::Mul:
2592 if (!I.isEquality())
2595 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2596 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2597 // Mask = -1 >> count-trailing-zeros(Cst).
2598 if (!CI->isZero() && !CI->isOne()) {
2599 const APInt &AP = CI->getValue();
2600 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2601 APInt::getLowBitsSet(AP.getBitWidth(),
2603 AP.countTrailingZeros()));
2604 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2605 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2606 return new ICmpInst(I.getPredicate(), And1, And2);
2610 case Instruction::UDiv:
2611 case Instruction::LShr:
2615 case Instruction::SDiv:
2616 case Instruction::AShr:
2617 if (!BO0->isExact() || !BO1->isExact())
2619 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2620 BO1->getOperand(0));
2621 case Instruction::Shl: {
2622 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2623 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2626 if (!NSW && I.isSigned())
2628 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2629 BO1->getOperand(0));
2636 // ~x < ~y --> y < x
2637 // ~x < cst --> ~cst < x
2638 if (match(Op0, m_Not(m_Value(A)))) {
2639 if (match(Op1, m_Not(m_Value(B))))
2640 return new ICmpInst(I.getPredicate(), B, A);
2641 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2642 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2645 // (a+b) <u a --> llvm.uadd.with.overflow.
2646 // (a+b) <u b --> llvm.uadd.with.overflow.
2647 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2648 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2649 (Op1 == A || Op1 == B))
2650 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2653 // a >u (a+b) --> llvm.uadd.with.overflow.
2654 // b >u (a+b) --> llvm.uadd.with.overflow.
2655 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2656 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2657 (Op0 == A || Op0 == B))
2658 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2662 if (I.isEquality()) {
2663 Value *A, *B, *C, *D;
2665 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2666 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2667 Value *OtherVal = A == Op1 ? B : A;
2668 return new ICmpInst(I.getPredicate(), OtherVal,
2669 Constant::getNullValue(A->getType()));
2672 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2673 // A^c1 == C^c2 --> A == C^(c1^c2)
2674 ConstantInt *C1, *C2;
2675 if (match(B, m_ConstantInt(C1)) &&
2676 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2677 Constant *NC = ConstantInt::get(I.getContext(),
2678 C1->getValue() ^ C2->getValue());
2679 Value *Xor = Builder->CreateXor(C, NC);
2680 return new ICmpInst(I.getPredicate(), A, Xor);
2683 // A^B == A^D -> B == D
2684 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2685 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2686 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2687 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2691 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2692 (A == Op0 || B == Op0)) {
2693 // A == (A^B) -> B == 0
2694 Value *OtherVal = A == Op0 ? B : A;
2695 return new ICmpInst(I.getPredicate(), OtherVal,
2696 Constant::getNullValue(A->getType()));
2699 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2700 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2701 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2702 Value *X = 0, *Y = 0, *Z = 0;
2705 X = B; Y = D; Z = A;
2706 } else if (A == D) {
2707 X = B; Y = C; Z = A;
2708 } else if (B == C) {
2709 X = A; Y = D; Z = B;
2710 } else if (B == D) {
2711 X = A; Y = C; Z = B;
2714 if (X) { // Build (X^Y) & Z
2715 Op1 = Builder->CreateXor(X, Y);
2716 Op1 = Builder->CreateAnd(Op1, Z);
2717 I.setOperand(0, Op1);
2718 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2723 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2724 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2726 if ((Op0->hasOneUse() &&
2727 match(Op0, m_ZExt(m_Value(A))) &&
2728 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2729 (Op1->hasOneUse() &&
2730 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2731 match(Op1, m_ZExt(m_Value(A))))) {
2732 APInt Pow2 = Cst1->getValue() + 1;
2733 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2734 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2735 return new ICmpInst(I.getPredicate(), A,
2736 Builder->CreateTrunc(B, A->getType()));
2739 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2740 // "icmp (and X, mask), cst"
2742 if (Op0->hasOneUse() &&
2743 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2744 m_ConstantInt(ShAmt))))) &&
2745 match(Op1, m_ConstantInt(Cst1)) &&
2746 // Only do this when A has multiple uses. This is most important to do
2747 // when it exposes other optimizations.
2749 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2751 if (ShAmt < ASize) {
2753 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2756 APInt CmpV = Cst1->getValue().zext(ASize);
2759 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2760 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2766 Value *X; ConstantInt *Cst;
2768 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2769 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2772 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2773 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2775 return Changed ? &I : 0;
2783 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2785 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2788 if (!isa<ConstantFP>(RHSC)) return 0;
2789 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2791 // Get the width of the mantissa. We don't want to hack on conversions that
2792 // might lose information from the integer, e.g. "i64 -> float"
2793 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2794 if (MantissaWidth == -1) return 0; // Unknown.
2796 // Check to see that the input is converted from an integer type that is small
2797 // enough that preserves all bits. TODO: check here for "known" sign bits.
2798 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2799 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2801 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2802 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2806 // If the conversion would lose info, don't hack on this.
2807 if ((int)InputSize > MantissaWidth)
2810 // Otherwise, we can potentially simplify the comparison. We know that it
2811 // will always come through as an integer value and we know the constant is
2812 // not a NAN (it would have been previously simplified).
2813 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2815 ICmpInst::Predicate Pred;
2816 switch (I.getPredicate()) {
2817 default: llvm_unreachable("Unexpected predicate!");
2818 case FCmpInst::FCMP_UEQ:
2819 case FCmpInst::FCMP_OEQ:
2820 Pred = ICmpInst::ICMP_EQ;
2822 case FCmpInst::FCMP_UGT:
2823 case FCmpInst::FCMP_OGT:
2824 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2826 case FCmpInst::FCMP_UGE:
2827 case FCmpInst::FCMP_OGE:
2828 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2830 case FCmpInst::FCMP_ULT:
2831 case FCmpInst::FCMP_OLT:
2832 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2834 case FCmpInst::FCMP_ULE:
2835 case FCmpInst::FCMP_OLE:
2836 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2838 case FCmpInst::FCMP_UNE:
2839 case FCmpInst::FCMP_ONE:
2840 Pred = ICmpInst::ICMP_NE;
2842 case FCmpInst::FCMP_ORD:
2843 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2844 case FCmpInst::FCMP_UNO:
2845 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2848 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2850 // Now we know that the APFloat is a normal number, zero or inf.
2852 // See if the FP constant is too large for the integer. For example,
2853 // comparing an i8 to 300.0.
2854 unsigned IntWidth = IntTy->getScalarSizeInBits();
2857 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2858 // and large values.
2859 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2860 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2861 APFloat::rmNearestTiesToEven);
2862 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2863 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2864 Pred == ICmpInst::ICMP_SLE)
2865 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2866 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2869 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2870 // +INF and large values.
2871 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2872 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2873 APFloat::rmNearestTiesToEven);
2874 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2875 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2876 Pred == ICmpInst::ICMP_ULE)
2877 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2878 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2883 // See if the RHS value is < SignedMin.
2884 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2885 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2886 APFloat::rmNearestTiesToEven);
2887 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2888 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2889 Pred == ICmpInst::ICMP_SGE)
2890 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2891 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2894 // See if the RHS value is < UnsignedMin.
2895 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2896 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
2897 APFloat::rmNearestTiesToEven);
2898 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
2899 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
2900 Pred == ICmpInst::ICMP_UGE)
2901 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2902 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2906 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2907 // [0, UMAX], but it may still be fractional. See if it is fractional by
2908 // casting the FP value to the integer value and back, checking for equality.
2909 // Don't do this for zero, because -0.0 is not fractional.
2910 Constant *RHSInt = LHSUnsigned
2911 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2912 : ConstantExpr::getFPToSI(RHSC, IntTy);
2913 if (!RHS.isZero()) {
2914 bool Equal = LHSUnsigned
2915 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2916 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2918 // If we had a comparison against a fractional value, we have to adjust
2919 // the compare predicate and sometimes the value. RHSC is rounded towards
2920 // zero at this point.
2922 default: llvm_unreachable("Unexpected integer comparison!");
2923 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2924 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2925 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2926 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2927 case ICmpInst::ICMP_ULE:
2928 // (float)int <= 4.4 --> int <= 4
2929 // (float)int <= -4.4 --> false
2930 if (RHS.isNegative())
2931 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2933 case ICmpInst::ICMP_SLE:
2934 // (float)int <= 4.4 --> int <= 4
2935 // (float)int <= -4.4 --> int < -4
2936 if (RHS.isNegative())
2937 Pred = ICmpInst::ICMP_SLT;
2939 case ICmpInst::ICMP_ULT:
2940 // (float)int < -4.4 --> false
2941 // (float)int < 4.4 --> int <= 4
2942 if (RHS.isNegative())
2943 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2944 Pred = ICmpInst::ICMP_ULE;
2946 case ICmpInst::ICMP_SLT:
2947 // (float)int < -4.4 --> int < -4
2948 // (float)int < 4.4 --> int <= 4
2949 if (!RHS.isNegative())
2950 Pred = ICmpInst::ICMP_SLE;
2952 case ICmpInst::ICMP_UGT:
2953 // (float)int > 4.4 --> int > 4
2954 // (float)int > -4.4 --> true
2955 if (RHS.isNegative())
2956 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2958 case ICmpInst::ICMP_SGT:
2959 // (float)int > 4.4 --> int > 4
2960 // (float)int > -4.4 --> int >= -4
2961 if (RHS.isNegative())
2962 Pred = ICmpInst::ICMP_SGE;
2964 case ICmpInst::ICMP_UGE:
2965 // (float)int >= -4.4 --> true
2966 // (float)int >= 4.4 --> int > 4
2967 if (RHS.isNegative())
2968 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2969 Pred = ICmpInst::ICMP_UGT;
2971 case ICmpInst::ICMP_SGE:
2972 // (float)int >= -4.4 --> int >= -4
2973 // (float)int >= 4.4 --> int > 4
2974 if (!RHS.isNegative())
2975 Pred = ICmpInst::ICMP_SGT;
2981 // Lower this FP comparison into an appropriate integer version of the
2983 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2986 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2987 bool Changed = false;
2989 /// Orders the operands of the compare so that they are listed from most
2990 /// complex to least complex. This puts constants before unary operators,
2991 /// before binary operators.
2992 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2997 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2999 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
3000 return ReplaceInstUsesWith(I, V);
3002 // Simplify 'fcmp pred X, X'
3004 switch (I.getPredicate()) {
3005 default: llvm_unreachable("Unknown predicate!");
3006 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3007 case FCmpInst::FCMP_ULT: // True if unordered or less than
3008 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3009 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3010 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3011 I.setPredicate(FCmpInst::FCMP_UNO);
3012 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3015 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3016 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3017 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3018 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3019 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3020 I.setPredicate(FCmpInst::FCMP_ORD);
3021 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3026 // Handle fcmp with constant RHS
3027 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3028 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3029 switch (LHSI->getOpcode()) {
3030 case Instruction::FPExt: {
3031 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3032 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3033 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3037 const fltSemantics *Sem;
3038 // FIXME: This shouldn't be here.
3039 if (LHSExt->getSrcTy()->isHalfTy())
3040 Sem = &APFloat::IEEEhalf;
3041 else if (LHSExt->getSrcTy()->isFloatTy())
3042 Sem = &APFloat::IEEEsingle;
3043 else if (LHSExt->getSrcTy()->isDoubleTy())
3044 Sem = &APFloat::IEEEdouble;
3045 else if (LHSExt->getSrcTy()->isFP128Ty())
3046 Sem = &APFloat::IEEEquad;
3047 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3048 Sem = &APFloat::x87DoubleExtended;
3049 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3050 Sem = &APFloat::PPCDoubleDouble;
3055 APFloat F = RHSF->getValueAPF();
3056 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3058 // Avoid lossy conversions and denormals. Zero is a special case
3059 // that's OK to convert.
3063 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3064 APFloat::cmpLessThan) || Fabs.isZero()))
3066 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3067 ConstantFP::get(RHSC->getContext(), F));
3070 case Instruction::PHI:
3071 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3072 // block. If in the same block, we're encouraging jump threading. If
3073 // not, we are just pessimizing the code by making an i1 phi.
3074 if (LHSI->getParent() == I.getParent())
3075 if (Instruction *NV = FoldOpIntoPhi(I))
3078 case Instruction::SIToFP:
3079 case Instruction::UIToFP:
3080 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3083 case Instruction::Select: {
3084 // If either operand of the select is a constant, we can fold the
3085 // comparison into the select arms, which will cause one to be
3086 // constant folded and the select turned into a bitwise or.
3087 Value *Op1 = 0, *Op2 = 0;
3088 if (LHSI->hasOneUse()) {
3089 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3090 // Fold the known value into the constant operand.
3091 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3092 // Insert a new FCmp of the other select operand.
3093 Op2 = Builder->CreateFCmp(I.getPredicate(),
3094 LHSI->getOperand(2), RHSC, I.getName());
3095 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3096 // Fold the known value into the constant operand.
3097 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3098 // Insert a new FCmp of the other select operand.
3099 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
3105 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3108 case Instruction::FSub: {
3109 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3111 if (match(LHSI, m_FNeg(m_Value(Op))))
3112 return new FCmpInst(I.getSwappedPredicate(), Op,
3113 ConstantExpr::getFNeg(RHSC));
3116 case Instruction::Load:
3117 if (GetElementPtrInst *GEP =
3118 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3119 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3120 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3121 !cast<LoadInst>(LHSI)->isVolatile())
3122 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3126 case Instruction::Call: {
3127 CallInst *CI = cast<CallInst>(LHSI);
3129 // Various optimization for fabs compared with zero.
3130 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3131 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3133 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3134 Func == LibFunc::fabsl) {
3135 switch (I.getPredicate()) {
3137 // fabs(x) < 0 --> false
3138 case FCmpInst::FCMP_OLT:
3139 return ReplaceInstUsesWith(I, Builder->getFalse());
3140 // fabs(x) > 0 --> x != 0
3141 case FCmpInst::FCMP_OGT:
3142 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3144 // fabs(x) <= 0 --> x == 0
3145 case FCmpInst::FCMP_OLE:
3146 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3148 // fabs(x) >= 0 --> !isnan(x)
3149 case FCmpInst::FCMP_OGE:
3150 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3152 // fabs(x) == 0 --> x == 0
3153 // fabs(x) != 0 --> x != 0
3154 case FCmpInst::FCMP_OEQ:
3155 case FCmpInst::FCMP_UEQ:
3156 case FCmpInst::FCMP_ONE:
3157 case FCmpInst::FCMP_UNE:
3158 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3167 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3169 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3170 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3172 // fcmp (fpext x), (fpext y) -> fcmp x, y
3173 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3174 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3175 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3176 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3177 RHSExt->getOperand(0));
3179 return Changed ? &I : 0;