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/IntrinsicInst.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/Target/TargetData.h"
19 #include "llvm/Support/ConstantRange.h"
20 #include "llvm/Support/GetElementPtrTypeIterator.h"
21 #include "llvm/Support/PatternMatch.h"
23 using namespace PatternMatch;
25 /// AddOne - Add one to a ConstantInt
26 static Constant *AddOne(Constant *C) {
27 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
29 /// SubOne - Subtract one from a ConstantInt
30 static Constant *SubOne(ConstantInt *C) {
31 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
34 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
35 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
38 static bool HasAddOverflow(ConstantInt *Result,
39 ConstantInt *In1, ConstantInt *In2,
42 if (In2->getValue().isNegative())
43 return Result->getValue().sgt(In1->getValue());
45 return Result->getValue().slt(In1->getValue());
47 return Result->getValue().ult(In1->getValue());
50 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
51 /// overflowed for this type.
52 static bool AddWithOverflow(Constant *&Result, Constant *In1,
53 Constant *In2, bool IsSigned = false) {
54 Result = ConstantExpr::getAdd(In1, In2);
56 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
57 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
58 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
59 if (HasAddOverflow(ExtractElement(Result, Idx),
60 ExtractElement(In1, Idx),
61 ExtractElement(In2, Idx),
68 return HasAddOverflow(cast<ConstantInt>(Result),
69 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
73 static bool HasSubOverflow(ConstantInt *Result,
74 ConstantInt *In1, ConstantInt *In2,
77 if (In2->getValue().isNegative())
78 return Result->getValue().slt(In1->getValue());
80 return Result->getValue().sgt(In1->getValue());
82 return Result->getValue().ugt(In1->getValue());
85 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
86 /// overflowed for this type.
87 static bool SubWithOverflow(Constant *&Result, Constant *In1,
88 Constant *In2, bool IsSigned = false) {
89 Result = ConstantExpr::getSub(In1, In2);
91 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
92 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
93 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
94 if (HasSubOverflow(ExtractElement(Result, Idx),
95 ExtractElement(In1, Idx),
96 ExtractElement(In2, Idx),
103 return HasSubOverflow(cast<ConstantInt>(Result),
104 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
108 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
109 /// comparison only checks the sign bit. If it only checks the sign bit, set
110 /// TrueIfSigned if the result of the comparison is true when the input value is
112 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
113 bool &TrueIfSigned) {
115 case ICmpInst::ICMP_SLT: // True if LHS s< 0
117 return RHS->isZero();
118 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
120 return RHS->isAllOnesValue();
121 case ICmpInst::ICMP_SGT: // True if LHS s> -1
122 TrueIfSigned = false;
123 return RHS->isAllOnesValue();
124 case ICmpInst::ICMP_UGT:
125 // True if LHS u> RHS and RHS == high-bit-mask - 1
127 return RHS->getValue() ==
128 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
129 case ICmpInst::ICMP_UGE:
130 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
132 return RHS->getValue().isSignBit();
138 // isHighOnes - Return true if the constant is of the form 1+0+.
139 // This is the same as lowones(~X).
140 static bool isHighOnes(const ConstantInt *CI) {
141 return (~CI->getValue() + 1).isPowerOf2();
144 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
145 /// set of known zero and one bits, compute the maximum and minimum values that
146 /// could have the specified known zero and known one bits, returning them in
148 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
149 const APInt& KnownOne,
150 APInt& Min, APInt& Max) {
151 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
152 KnownZero.getBitWidth() == Min.getBitWidth() &&
153 KnownZero.getBitWidth() == Max.getBitWidth() &&
154 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
155 APInt UnknownBits = ~(KnownZero|KnownOne);
157 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
158 // bit if it is unknown.
160 Max = KnownOne|UnknownBits;
162 if (UnknownBits.isNegative()) { // Sign bit is unknown
163 Min.set(Min.getBitWidth()-1);
164 Max.clear(Max.getBitWidth()-1);
168 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
169 // a set of known zero and one bits, compute the maximum and minimum values that
170 // could have the specified known zero and known one bits, returning them in
172 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
173 const APInt &KnownOne,
174 APInt &Min, APInt &Max) {
175 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
176 KnownZero.getBitWidth() == Min.getBitWidth() &&
177 KnownZero.getBitWidth() == Max.getBitWidth() &&
178 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
179 APInt UnknownBits = ~(KnownZero|KnownOne);
181 // The minimum value is when the unknown bits are all zeros.
183 // The maximum value is when the unknown bits are all ones.
184 Max = KnownOne|UnknownBits;
189 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
190 /// cmp pred (load (gep GV, ...)), cmpcst
191 /// where GV is a global variable with a constant initializer. Try to simplify
192 /// this into some simple computation that does not need the load. For example
193 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
195 /// If AndCst is non-null, then the loaded value is masked with that constant
196 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
197 Instruction *InstCombiner::
198 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
199 CmpInst &ICI, ConstantInt *AndCst) {
200 // We need TD information to know the pointer size unless this is inbounds.
201 if (!GEP->isInBounds() && TD == 0) return 0;
203 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer());
204 if (Init == 0 || Init->getNumOperands() > 1024) return 0;
206 // There are many forms of this optimization we can handle, for now, just do
207 // the simple index into a single-dimensional array.
209 // Require: GEP GV, 0, i {{, constant indices}}
210 if (GEP->getNumOperands() < 3 ||
211 !isa<ConstantInt>(GEP->getOperand(1)) ||
212 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
213 isa<Constant>(GEP->getOperand(2)))
216 // Check that indices after the variable are constants and in-range for the
217 // type they index. Collect the indices. This is typically for arrays of
219 SmallVector<unsigned, 4> LaterIndices;
221 const Type *EltTy = cast<ArrayType>(Init->getType())->getElementType();
222 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
223 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
224 if (Idx == 0) return 0; // Variable index.
226 uint64_t IdxVal = Idx->getZExtValue();
227 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
229 if (const StructType *STy = dyn_cast<StructType>(EltTy))
230 EltTy = STy->getElementType(IdxVal);
231 else if (const ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
232 if (IdxVal >= ATy->getNumElements()) return 0;
233 EltTy = ATy->getElementType();
235 return 0; // Unknown type.
238 LaterIndices.push_back(IdxVal);
241 enum { Overdefined = -3, Undefined = -2 };
243 // Variables for our state machines.
245 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
246 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
247 // and 87 is the second (and last) index. FirstTrueElement is -2 when
248 // undefined, otherwise set to the first true element. SecondTrueElement is
249 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
250 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
252 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
253 // form "i != 47 & i != 87". Same state transitions as for true elements.
254 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
256 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
257 /// define a state machine that triggers for ranges of values that the index
258 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
259 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
260 /// index in the range (inclusive). We use -2 for undefined here because we
261 /// use relative comparisons and don't want 0-1 to match -1.
262 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
264 // MagicBitvector - This is a magic bitvector where we set a bit if the
265 // comparison is true for element 'i'. If there are 64 elements or less in
266 // the array, this will fully represent all the comparison results.
267 uint64_t MagicBitvector = 0;
270 // Scan the array and see if one of our patterns matches.
271 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
272 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) {
273 Constant *Elt = Init->getOperand(i);
275 // If this is indexing an array of structures, get the structure element.
276 if (!LaterIndices.empty())
277 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices.data(),
278 LaterIndices.size());
280 // If the element is masked, handle it.
281 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
283 // Find out if the comparison would be true or false for the i'th element.
284 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
286 // If the result is undef for this element, ignore it.
287 if (isa<UndefValue>(C)) {
288 // Extend range state machines to cover this element in case there is an
289 // undef in the middle of the range.
290 if (TrueRangeEnd == (int)i-1)
292 if (FalseRangeEnd == (int)i-1)
297 // If we can't compute the result for any of the elements, we have to give
298 // up evaluating the entire conditional.
299 if (!isa<ConstantInt>(C)) return 0;
301 // Otherwise, we know if the comparison is true or false for this element,
302 // update our state machines.
303 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
305 // State machine for single/double/range index comparison.
307 // Update the TrueElement state machine.
308 if (FirstTrueElement == Undefined)
309 FirstTrueElement = TrueRangeEnd = i; // First true element.
311 // Update double-compare state machine.
312 if (SecondTrueElement == Undefined)
313 SecondTrueElement = i;
315 SecondTrueElement = Overdefined;
317 // Update range state machine.
318 if (TrueRangeEnd == (int)i-1)
321 TrueRangeEnd = Overdefined;
324 // Update the FalseElement state machine.
325 if (FirstFalseElement == Undefined)
326 FirstFalseElement = FalseRangeEnd = i; // First false element.
328 // Update double-compare state machine.
329 if (SecondFalseElement == Undefined)
330 SecondFalseElement = i;
332 SecondFalseElement = Overdefined;
334 // Update range state machine.
335 if (FalseRangeEnd == (int)i-1)
338 FalseRangeEnd = Overdefined;
343 // If this element is in range, update our magic bitvector.
344 if (i < 64 && IsTrueForElt)
345 MagicBitvector |= 1ULL << i;
347 // If all of our states become overdefined, bail out early. Since the
348 // predicate is expensive, only check it every 8 elements. This is only
349 // really useful for really huge arrays.
350 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
351 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
352 FalseRangeEnd == Overdefined)
356 // Now that we've scanned the entire array, emit our new comparison(s). We
357 // order the state machines in complexity of the generated code.
358 Value *Idx = GEP->getOperand(2);
360 // If the index is larger than the pointer size of the target, truncate the
361 // index down like the GEP would do implicitly. We don't have to do this for
362 // an inbounds GEP because the index can't be out of range.
363 if (!GEP->isInBounds() &&
364 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
365 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
367 // If the comparison is only true for one or two elements, emit direct
369 if (SecondTrueElement != Overdefined) {
370 // None true -> false.
371 if (FirstTrueElement == Undefined)
372 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
374 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
376 // True for one element -> 'i == 47'.
377 if (SecondTrueElement == Undefined)
378 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
380 // True for two elements -> 'i == 47 | i == 72'.
381 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
382 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
383 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
384 return BinaryOperator::CreateOr(C1, C2);
387 // If the comparison is only false for one or two elements, emit direct
389 if (SecondFalseElement != Overdefined) {
390 // None false -> true.
391 if (FirstFalseElement == Undefined)
392 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
394 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
396 // False for one element -> 'i != 47'.
397 if (SecondFalseElement == Undefined)
398 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
400 // False for two elements -> 'i != 47 & i != 72'.
401 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
402 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
403 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
404 return BinaryOperator::CreateAnd(C1, C2);
407 // If the comparison can be replaced with a range comparison for the elements
408 // where it is true, emit the range check.
409 if (TrueRangeEnd != Overdefined) {
410 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
412 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
413 if (FirstTrueElement) {
414 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
415 Idx = Builder->CreateAdd(Idx, Offs);
418 Value *End = ConstantInt::get(Idx->getType(),
419 TrueRangeEnd-FirstTrueElement+1);
420 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
423 // False range check.
424 if (FalseRangeEnd != Overdefined) {
425 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
426 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
427 if (FirstFalseElement) {
428 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
429 Idx = Builder->CreateAdd(Idx, Offs);
432 Value *End = ConstantInt::get(Idx->getType(),
433 FalseRangeEnd-FirstFalseElement);
434 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
438 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
439 // of this load, replace it with computation that does:
440 // ((magic_cst >> i) & 1) != 0
441 if (Init->getNumOperands() <= 32 ||
442 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) {
444 if (Init->getNumOperands() <= 32)
445 Ty = Type::getInt32Ty(Init->getContext());
447 Ty = Type::getInt64Ty(Init->getContext());
448 Value *V = Builder->CreateIntCast(Idx, Ty, false);
449 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
450 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
451 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
458 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
459 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
460 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
461 /// be complex, and scales are involved. The above expression would also be
462 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
463 /// This later form is less amenable to optimization though, and we are allowed
464 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
466 /// If we can't emit an optimized form for this expression, this returns null.
468 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
470 TargetData &TD = *IC.getTargetData();
471 gep_type_iterator GTI = gep_type_begin(GEP);
473 // Check to see if this gep only has a single variable index. If so, and if
474 // any constant indices are a multiple of its scale, then we can compute this
475 // in terms of the scale of the variable index. For example, if the GEP
476 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
477 // because the expression will cross zero at the same point.
478 unsigned i, e = GEP->getNumOperands();
480 for (i = 1; i != e; ++i, ++GTI) {
481 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
482 // Compute the aggregate offset of constant indices.
483 if (CI->isZero()) continue;
485 // Handle a struct index, which adds its field offset to the pointer.
486 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
487 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
489 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
490 Offset += Size*CI->getSExtValue();
493 // Found our variable index.
498 // If there are no variable indices, we must have a constant offset, just
499 // evaluate it the general way.
500 if (i == e) return 0;
502 Value *VariableIdx = GEP->getOperand(i);
503 // Determine the scale factor of the variable element. For example, this is
504 // 4 if the variable index is into an array of i32.
505 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
507 // Verify that there are no other variable indices. If so, emit the hard way.
508 for (++i, ++GTI; i != e; ++i, ++GTI) {
509 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
512 // Compute the aggregate offset of constant indices.
513 if (CI->isZero()) continue;
515 // Handle a struct index, which adds its field offset to the pointer.
516 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
517 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
519 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
520 Offset += Size*CI->getSExtValue();
524 // Okay, we know we have a single variable index, which must be a
525 // pointer/array/vector index. If there is no offset, life is simple, return
527 unsigned IntPtrWidth = TD.getPointerSizeInBits();
529 // Cast to intptrty in case a truncation occurs. If an extension is needed,
530 // we don't need to bother extending: the extension won't affect where the
531 // computation crosses zero.
532 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
533 VariableIdx = new TruncInst(VariableIdx,
534 TD.getIntPtrType(VariableIdx->getContext()),
535 VariableIdx->getName(), &I);
539 // Otherwise, there is an index. The computation we will do will be modulo
540 // the pointer size, so get it.
541 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
543 Offset &= PtrSizeMask;
544 VariableScale &= PtrSizeMask;
546 // To do this transformation, any constant index must be a multiple of the
547 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
548 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
549 // multiple of the variable scale.
550 int64_t NewOffs = Offset / (int64_t)VariableScale;
551 if (Offset != NewOffs*(int64_t)VariableScale)
554 // Okay, we can do this evaluation. Start by converting the index to intptr.
555 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
556 if (VariableIdx->getType() != IntPtrTy)
557 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
559 VariableIdx->getName(), &I);
560 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
561 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
564 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
565 /// else. At this point we know that the GEP is on the LHS of the comparison.
566 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
567 ICmpInst::Predicate Cond,
569 // Look through bitcasts.
570 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
571 RHS = BCI->getOperand(0);
573 Value *PtrBase = GEPLHS->getOperand(0);
574 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
575 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
576 // This transformation (ignoring the base and scales) is valid because we
577 // know pointers can't overflow since the gep is inbounds. See if we can
578 // output an optimized form.
579 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
581 // If not, synthesize the offset the hard way.
583 Offset = EmitGEPOffset(GEPLHS);
584 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
585 Constant::getNullValue(Offset->getType()));
586 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
587 // If the base pointers are different, but the indices are the same, just
588 // compare the base pointer.
589 if (PtrBase != GEPRHS->getOperand(0)) {
590 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
591 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
592 GEPRHS->getOperand(0)->getType();
594 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
595 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
596 IndicesTheSame = false;
600 // If all indices are the same, just compare the base pointers.
602 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
603 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
605 // Otherwise, the base pointers are different and the indices are
606 // different, bail out.
610 // If one of the GEPs has all zero indices, recurse.
611 bool AllZeros = true;
612 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
613 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
614 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
619 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
620 ICmpInst::getSwappedPredicate(Cond), I);
622 // If the other GEP has all zero indices, recurse.
624 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
625 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
626 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
631 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
633 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
634 // If the GEPs only differ by one index, compare it.
635 unsigned NumDifferences = 0; // Keep track of # differences.
636 unsigned DiffOperand = 0; // The operand that differs.
637 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
638 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
639 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
640 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
641 // Irreconcilable differences.
645 if (NumDifferences++) break;
650 if (NumDifferences == 0) // SAME GEP?
651 return ReplaceInstUsesWith(I, // No comparison is needed here.
652 ConstantInt::get(Type::getInt1Ty(I.getContext()),
653 ICmpInst::isTrueWhenEqual(Cond)));
655 else if (NumDifferences == 1) {
656 Value *LHSV = GEPLHS->getOperand(DiffOperand);
657 Value *RHSV = GEPRHS->getOperand(DiffOperand);
658 // Make sure we do a signed comparison here.
659 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
663 // Only lower this if the icmp is the only user of the GEP or if we expect
664 // the result to fold to a constant!
666 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
667 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
668 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
669 Value *L = EmitGEPOffset(GEPLHS);
670 Value *R = EmitGEPOffset(GEPRHS);
671 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
677 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
678 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
679 Value *X, ConstantInt *CI,
680 ICmpInst::Predicate Pred,
682 // If we have X+0, exit early (simplifying logic below) and let it get folded
683 // elsewhere. icmp X+0, X -> icmp X, X
685 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
686 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
689 // (X+4) == X -> false.
690 if (Pred == ICmpInst::ICMP_EQ)
691 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
693 // (X+4) != X -> true.
694 if (Pred == ICmpInst::ICMP_NE)
695 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
697 // If this is an instruction (as opposed to constantexpr) get NUW/NSW info.
698 bool isNUW = false, isNSW = false;
699 if (BinaryOperator *Add = dyn_cast<BinaryOperator>(TheAdd)) {
700 isNUW = Add->hasNoUnsignedWrap();
701 isNSW = Add->hasNoSignedWrap();
704 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
705 // so the values can never be equal. Similiarly for all other "or equals"
708 // (X+1) <u X --> X >u (MAXUINT-1) --> X != 255
709 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
710 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
711 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
712 // If this is an NUW add, then this is always false.
714 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
716 Value *R = ConstantExpr::getSub(ConstantInt::get(CI->getType(), -1ULL), CI);
717 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
720 // (X+1) >u X --> X <u (0-1) --> X != 255
721 // (X+2) >u X --> X <u (0-2) --> X <u 254
722 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
723 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
724 // If this is an NUW add, then this is always true.
726 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
727 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
730 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
731 ConstantInt *SMax = ConstantInt::get(X->getContext(),
732 APInt::getSignedMaxValue(BitWidth));
734 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
735 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
736 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
737 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
738 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
739 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
740 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
741 // If this is an NSW add, then we have two cases: if the constant is
742 // positive, then this is always false, if negative, this is always true.
744 bool isTrue = CI->getValue().isNegative();
745 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
748 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
751 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
752 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
753 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
754 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
755 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
756 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
758 // If this is an NSW add, then we have two cases: if the constant is
759 // positive, then this is always true, if negative, this is always false.
761 bool isTrue = !CI->getValue().isNegative();
762 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
765 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
766 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
767 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
770 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
771 /// and CmpRHS are both known to be integer constants.
772 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
773 ConstantInt *DivRHS) {
774 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
775 const APInt &CmpRHSV = CmpRHS->getValue();
777 // FIXME: If the operand types don't match the type of the divide
778 // then don't attempt this transform. The code below doesn't have the
779 // logic to deal with a signed divide and an unsigned compare (and
780 // vice versa). This is because (x /s C1) <s C2 produces different
781 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
782 // (x /u C1) <u C2. Simply casting the operands and result won't
783 // work. :( The if statement below tests that condition and bails
785 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
786 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
788 if (DivRHS->isZero())
789 return 0; // The ProdOV computation fails on divide by zero.
790 if (DivIsSigned && DivRHS->isAllOnesValue())
791 return 0; // The overflow computation also screws up here
793 return 0; // Not worth bothering, and eliminates some funny cases
796 // Compute Prod = CI * DivRHS. We are essentially solving an equation
797 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
798 // C2 (CI). By solving for X we can turn this into a range check
799 // instead of computing a divide.
800 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
802 // Determine if the product overflows by seeing if the product is
803 // not equal to the divide. Make sure we do the same kind of divide
804 // as in the LHS instruction that we're folding.
805 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
806 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
808 // Get the ICmp opcode
809 ICmpInst::Predicate Pred = ICI.getPredicate();
811 // Figure out the interval that is being checked. For example, a comparison
812 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
813 // Compute this interval based on the constants involved and the signedness of
814 // the compare/divide. This computes a half-open interval, keeping track of
815 // whether either value in the interval overflows. After analysis each
816 // overflow variable is set to 0 if it's corresponding bound variable is valid
817 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
818 int LoOverflow = 0, HiOverflow = 0;
819 Constant *LoBound = 0, *HiBound = 0;
821 if (!DivIsSigned) { // udiv
822 // e.g. X/5 op 3 --> [15, 20)
824 HiOverflow = LoOverflow = ProdOV;
826 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
827 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
828 if (CmpRHSV == 0) { // (X / pos) op 0
829 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
830 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
832 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
833 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
834 HiOverflow = LoOverflow = ProdOV;
836 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
837 } else { // (X / pos) op neg
838 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
839 HiBound = AddOne(Prod);
840 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
842 ConstantInt* DivNeg =
843 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
844 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
847 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
848 if (CmpRHSV == 0) { // (X / neg) op 0
849 // e.g. X/-5 op 0 --> [-4, 5)
850 LoBound = AddOne(DivRHS);
851 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
852 if (HiBound == DivRHS) { // -INTMIN = INTMIN
853 HiOverflow = 1; // [INTMIN+1, overflow)
854 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
856 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
857 // e.g. X/-5 op 3 --> [-19, -14)
858 HiBound = AddOne(Prod);
859 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
861 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
862 } else { // (X / neg) op neg
863 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
864 LoOverflow = HiOverflow = ProdOV;
866 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
869 // Dividing by a negative swaps the condition. LT <-> GT
870 Pred = ICmpInst::getSwappedPredicate(Pred);
873 Value *X = DivI->getOperand(0);
875 default: llvm_unreachable("Unhandled icmp opcode!");
876 case ICmpInst::ICMP_EQ:
877 if (LoOverflow && HiOverflow)
878 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
880 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
881 ICmpInst::ICMP_UGE, X, LoBound);
883 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
884 ICmpInst::ICMP_ULT, X, HiBound);
886 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
887 case ICmpInst::ICMP_NE:
888 if (LoOverflow && HiOverflow)
889 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
891 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
892 ICmpInst::ICMP_ULT, X, LoBound);
894 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
895 ICmpInst::ICMP_UGE, X, HiBound);
897 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
898 case ICmpInst::ICMP_ULT:
899 case ICmpInst::ICMP_SLT:
900 if (LoOverflow == +1) // Low bound is greater than input range.
901 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
902 if (LoOverflow == -1) // Low bound is less than input range.
903 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
904 return new ICmpInst(Pred, X, LoBound);
905 case ICmpInst::ICMP_UGT:
906 case ICmpInst::ICMP_SGT:
907 if (HiOverflow == +1) // High bound greater than input range.
908 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
909 else if (HiOverflow == -1) // High bound less than input range.
910 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
911 if (Pred == ICmpInst::ICMP_UGT)
912 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
914 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
919 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
921 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
924 const APInt &RHSV = RHS->getValue();
926 switch (LHSI->getOpcode()) {
927 case Instruction::Trunc:
928 if (ICI.isEquality() && LHSI->hasOneUse()) {
929 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
930 // of the high bits truncated out of x are known.
931 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
932 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
933 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
934 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
935 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
937 // If all the high bits are known, we can do this xform.
938 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
939 // Pull in the high bits from known-ones set.
940 APInt NewRHS(RHS->getValue());
941 NewRHS.zext(SrcBits);
943 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
944 ConstantInt::get(ICI.getContext(), NewRHS));
949 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
950 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
951 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
953 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
954 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
955 Value *CompareVal = LHSI->getOperand(0);
957 // If the sign bit of the XorCST is not set, there is no change to
958 // the operation, just stop using the Xor.
959 if (!XorCST->getValue().isNegative()) {
960 ICI.setOperand(0, CompareVal);
965 // Was the old condition true if the operand is positive?
966 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
968 // If so, the new one isn't.
969 isTrueIfPositive ^= true;
971 if (isTrueIfPositive)
972 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
975 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
979 if (LHSI->hasOneUse()) {
980 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
981 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
982 const APInt &SignBit = XorCST->getValue();
983 ICmpInst::Predicate Pred = ICI.isSigned()
984 ? ICI.getUnsignedPredicate()
985 : ICI.getSignedPredicate();
986 return new ICmpInst(Pred, LHSI->getOperand(0),
987 ConstantInt::get(ICI.getContext(),
991 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
992 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
993 const APInt &NotSignBit = XorCST->getValue();
994 ICmpInst::Predicate Pred = ICI.isSigned()
995 ? ICI.getUnsignedPredicate()
996 : ICI.getSignedPredicate();
997 Pred = ICI.getSwappedPredicate(Pred);
998 return new ICmpInst(Pred, LHSI->getOperand(0),
999 ConstantInt::get(ICI.getContext(),
1000 RHSV ^ NotSignBit));
1005 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1006 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1007 LHSI->getOperand(0)->hasOneUse()) {
1008 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1010 // If the LHS is an AND of a truncating cast, we can widen the
1011 // and/compare to be the input width without changing the value
1012 // produced, eliminating a cast.
1013 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1014 // We can do this transformation if either the AND constant does not
1015 // have its sign bit set or if it is an equality comparison.
1016 // Extending a relational comparison when we're checking the sign
1017 // bit would not work.
1018 if (Cast->hasOneUse() &&
1019 (ICI.isEquality() ||
1020 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
1022 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
1023 APInt NewCST = AndCST->getValue();
1024 NewCST.zext(BitWidth);
1026 NewCI.zext(BitWidth);
1028 Builder->CreateAnd(Cast->getOperand(0),
1029 ConstantInt::get(ICI.getContext(), NewCST),
1031 return new ICmpInst(ICI.getPredicate(), NewAnd,
1032 ConstantInt::get(ICI.getContext(), NewCI));
1036 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1037 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1038 // happens a LOT in code produced by the C front-end, for bitfield
1040 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1041 if (Shift && !Shift->isShift())
1045 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1046 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1047 const Type *AndTy = AndCST->getType(); // Type of the and.
1049 // We can fold this as long as we can't shift unknown bits
1050 // into the mask. This can only happen with signed shift
1051 // rights, as they sign-extend.
1053 bool CanFold = Shift->isLogicalShift();
1055 // To test for the bad case of the signed shr, see if any
1056 // of the bits shifted in could be tested after the mask.
1057 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1058 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1060 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1061 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1062 AndCST->getValue()) == 0)
1068 if (Shift->getOpcode() == Instruction::Shl)
1069 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1071 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1073 // Check to see if we are shifting out any of the bits being
1075 if (ConstantExpr::get(Shift->getOpcode(),
1076 NewCst, ShAmt) != RHS) {
1077 // If we shifted bits out, the fold is not going to work out.
1078 // As a special case, check to see if this means that the
1079 // result is always true or false now.
1080 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1081 return ReplaceInstUsesWith(ICI,
1082 ConstantInt::getFalse(ICI.getContext()));
1083 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1084 return ReplaceInstUsesWith(ICI,
1085 ConstantInt::getTrue(ICI.getContext()));
1087 ICI.setOperand(1, NewCst);
1088 Constant *NewAndCST;
1089 if (Shift->getOpcode() == Instruction::Shl)
1090 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1092 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1093 LHSI->setOperand(1, NewAndCST);
1094 LHSI->setOperand(0, Shift->getOperand(0));
1095 Worklist.Add(Shift); // Shift is dead.
1101 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1102 // preferable because it allows the C<<Y expression to be hoisted out
1103 // of a loop if Y is invariant and X is not.
1104 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1105 ICI.isEquality() && !Shift->isArithmeticShift() &&
1106 !isa<Constant>(Shift->getOperand(0))) {
1109 if (Shift->getOpcode() == Instruction::LShr) {
1110 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
1112 // Insert a logical shift.
1113 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
1116 // Compute X & (C << Y).
1118 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1120 ICI.setOperand(0, NewAnd);
1125 // Try to optimize things like "A[i]&42 == 0" to index computations.
1126 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1127 if (GetElementPtrInst *GEP =
1128 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1129 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1130 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1131 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1132 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1133 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1139 case Instruction::Or: {
1140 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1143 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1144 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1145 // -> and (icmp eq P, null), (icmp eq Q, null).
1147 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1148 Constant::getNullValue(P->getType()));
1149 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1150 Constant::getNullValue(Q->getType()));
1152 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1153 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1155 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1161 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1162 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1165 uint32_t TypeBits = RHSV.getBitWidth();
1167 // Check that the shift amount is in range. If not, don't perform
1168 // undefined shifts. When the shift is visited it will be
1170 if (ShAmt->uge(TypeBits))
1173 if (ICI.isEquality()) {
1174 // If we are comparing against bits always shifted out, the
1175 // comparison cannot succeed.
1177 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1179 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1180 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1182 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1183 return ReplaceInstUsesWith(ICI, Cst);
1186 if (LHSI->hasOneUse()) {
1187 // Otherwise strength reduce the shift into an and.
1188 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1190 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1191 TypeBits-ShAmtVal));
1194 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1195 return new ICmpInst(ICI.getPredicate(), And,
1196 ConstantInt::get(ICI.getContext(),
1197 RHSV.lshr(ShAmtVal)));
1201 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1202 bool TrueIfSigned = false;
1203 if (LHSI->hasOneUse() &&
1204 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1205 // (X << 31) <s 0 --> (X&1) != 0
1206 Constant *Mask = ConstantInt::get(ICI.getContext(), APInt(TypeBits, 1) <<
1207 (TypeBits-ShAmt->getZExtValue()-1));
1209 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1210 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1211 And, Constant::getNullValue(And->getType()));
1216 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1217 case Instruction::AShr: {
1218 // Only handle equality comparisons of shift-by-constant.
1219 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1220 if (!ShAmt || !ICI.isEquality()) break;
1222 // Check that the shift amount is in range. If not, don't perform
1223 // undefined shifts. When the shift is visited it will be
1225 uint32_t TypeBits = RHSV.getBitWidth();
1226 if (ShAmt->uge(TypeBits))
1229 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1231 // If we are comparing against bits always shifted out, the
1232 // comparison cannot succeed.
1233 APInt Comp = RHSV << ShAmtVal;
1234 if (LHSI->getOpcode() == Instruction::LShr)
1235 Comp = Comp.lshr(ShAmtVal);
1237 Comp = Comp.ashr(ShAmtVal);
1239 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
1240 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1241 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1243 return ReplaceInstUsesWith(ICI, Cst);
1246 // Otherwise, check to see if the bits shifted out are known to be zero.
1247 // If so, we can compare against the unshifted value:
1248 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1249 if (LHSI->hasOneUse() &&
1250 MaskedValueIsZero(LHSI->getOperand(0),
1251 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
1252 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1253 ConstantExpr::getShl(RHS, ShAmt));
1256 if (LHSI->hasOneUse()) {
1257 // Otherwise strength reduce the shift into an and.
1258 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1259 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
1261 Value *And = Builder->CreateAnd(LHSI->getOperand(0),
1262 Mask, LHSI->getName()+".mask");
1263 return new ICmpInst(ICI.getPredicate(), And,
1264 ConstantExpr::getShl(RHS, ShAmt));
1269 case Instruction::SDiv:
1270 case Instruction::UDiv:
1271 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1272 // Fold this div into the comparison, producing a range check.
1273 // Determine, based on the divide type, what the range is being
1274 // checked. If there is an overflow on the low or high side, remember
1275 // it, otherwise compute the range [low, hi) bounding the new value.
1276 // See: InsertRangeTest above for the kinds of replacements possible.
1277 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1278 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1283 case Instruction::Add:
1284 // Fold: icmp pred (add X, C1), C2
1285 if (!ICI.isEquality()) {
1286 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1288 const APInt &LHSV = LHSC->getValue();
1290 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1293 if (ICI.isSigned()) {
1294 if (CR.getLower().isSignBit()) {
1295 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1296 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1297 } else if (CR.getUpper().isSignBit()) {
1298 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1299 ConstantInt::get(ICI.getContext(),CR.getLower()));
1302 if (CR.getLower().isMinValue()) {
1303 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1304 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1305 } else if (CR.getUpper().isMinValue()) {
1306 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1307 ConstantInt::get(ICI.getContext(),CR.getLower()));
1314 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1315 if (ICI.isEquality()) {
1316 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1318 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1319 // the second operand is a constant, simplify a bit.
1320 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1321 switch (BO->getOpcode()) {
1322 case Instruction::SRem:
1323 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1324 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1325 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1326 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
1328 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1330 return new ICmpInst(ICI.getPredicate(), NewRem,
1331 Constant::getNullValue(BO->getType()));
1335 case Instruction::Add:
1336 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1337 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1338 if (BO->hasOneUse())
1339 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1340 ConstantExpr::getSub(RHS, BOp1C));
1341 } else if (RHSV == 0) {
1342 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1343 // efficiently invertible, or if the add has just this one use.
1344 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1346 if (Value *NegVal = dyn_castNegVal(BOp1))
1347 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1348 else if (Value *NegVal = dyn_castNegVal(BOp0))
1349 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1350 else if (BO->hasOneUse()) {
1351 Value *Neg = Builder->CreateNeg(BOp1);
1353 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1357 case Instruction::Xor:
1358 // For the xor case, we can xor two constants together, eliminating
1359 // the explicit xor.
1360 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
1361 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1362 ConstantExpr::getXor(RHS, BOC));
1365 case Instruction::Sub:
1366 // Replace (([sub|xor] A, B) != 0) with (A != B)
1368 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1372 case Instruction::Or:
1373 // If bits are being or'd in that are not present in the constant we
1374 // are comparing against, then the comparison could never succeed!
1375 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1376 Constant *NotCI = ConstantExpr::getNot(RHS);
1377 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1378 return ReplaceInstUsesWith(ICI,
1379 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1384 case Instruction::And:
1385 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1386 // If bits are being compared against that are and'd out, then the
1387 // comparison can never succeed!
1388 if ((RHSV & ~BOC->getValue()) != 0)
1389 return ReplaceInstUsesWith(ICI,
1390 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1393 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1394 if (RHS == BOC && RHSV.isPowerOf2())
1395 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1396 ICmpInst::ICMP_NE, LHSI,
1397 Constant::getNullValue(RHS->getType()));
1399 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1400 if (BOC->getValue().isSignBit()) {
1401 Value *X = BO->getOperand(0);
1402 Constant *Zero = Constant::getNullValue(X->getType());
1403 ICmpInst::Predicate pred = isICMP_NE ?
1404 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1405 return new ICmpInst(pred, X, Zero);
1408 // ((X & ~7) == 0) --> X < 8
1409 if (RHSV == 0 && isHighOnes(BOC)) {
1410 Value *X = BO->getOperand(0);
1411 Constant *NegX = ConstantExpr::getNeg(BOC);
1412 ICmpInst::Predicate pred = isICMP_NE ?
1413 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1414 return new ICmpInst(pred, X, NegX);
1419 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1420 // Handle icmp {eq|ne} <intrinsic>, intcst.
1421 if (II->getIntrinsicID() == Intrinsic::bswap) {
1423 ICI.setOperand(0, II->getOperand(1));
1424 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1432 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1433 /// We only handle extending casts so far.
1435 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1436 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1437 Value *LHSCIOp = LHSCI->getOperand(0);
1438 const Type *SrcTy = LHSCIOp->getType();
1439 const Type *DestTy = LHSCI->getType();
1442 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1443 // integer type is the same size as the pointer type.
1444 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1445 TD->getPointerSizeInBits() ==
1446 cast<IntegerType>(DestTy)->getBitWidth()) {
1448 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1449 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1450 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1451 RHSOp = RHSC->getOperand(0);
1452 // If the pointer types don't match, insert a bitcast.
1453 if (LHSCIOp->getType() != RHSOp->getType())
1454 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1458 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1461 // The code below only handles extension cast instructions, so far.
1463 if (LHSCI->getOpcode() != Instruction::ZExt &&
1464 LHSCI->getOpcode() != Instruction::SExt)
1467 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1468 bool isSignedCmp = ICI.isSigned();
1470 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1471 // Not an extension from the same type?
1472 RHSCIOp = CI->getOperand(0);
1473 if (RHSCIOp->getType() != LHSCIOp->getType())
1476 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1477 // and the other is a zext), then we can't handle this.
1478 if (CI->getOpcode() != LHSCI->getOpcode())
1481 // Deal with equality cases early.
1482 if (ICI.isEquality())
1483 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1485 // A signed comparison of sign extended values simplifies into a
1486 // signed comparison.
1487 if (isSignedCmp && isSignedExt)
1488 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1490 // The other three cases all fold into an unsigned comparison.
1491 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1494 // If we aren't dealing with a constant on the RHS, exit early
1495 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1499 // Compute the constant that would happen if we truncated to SrcTy then
1500 // reextended to DestTy.
1501 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1502 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1505 // If the re-extended constant didn't change...
1507 // Deal with equality cases early.
1508 if (ICI.isEquality())
1509 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1511 // A signed comparison of sign extended values simplifies into a
1512 // signed comparison.
1513 if (isSignedExt && isSignedCmp)
1514 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1516 // The other three cases all fold into an unsigned comparison.
1517 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1520 // The re-extended constant changed so the constant cannot be represented
1521 // in the shorter type. Consequently, we cannot emit a simple comparison.
1523 // First, handle some easy cases. We know the result cannot be equal at this
1524 // point so handle the ICI.isEquality() cases
1525 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1526 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
1527 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1528 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
1530 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1531 // should have been folded away previously and not enter in here.
1534 // We're performing a signed comparison.
1535 if (cast<ConstantInt>(CI)->getValue().isNegative())
1536 Result = ConstantInt::getFalse(ICI.getContext()); // X < (small) --> false
1538 Result = ConstantInt::getTrue(ICI.getContext()); // X < (large) --> true
1540 // We're performing an unsigned comparison.
1542 // We're performing an unsigned comp with a sign extended value.
1543 // This is true if the input is >= 0. [aka >s -1]
1544 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1545 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1547 // Unsigned extend & unsigned compare -> always true.
1548 Result = ConstantInt::getTrue(ICI.getContext());
1552 // Finally, return the value computed.
1553 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
1554 ICI.getPredicate() == ICmpInst::ICMP_SLT)
1555 return ReplaceInstUsesWith(ICI, Result);
1557 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
1558 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
1559 "ICmp should be folded!");
1560 if (Constant *CI = dyn_cast<Constant>(Result))
1561 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
1562 return BinaryOperator::CreateNot(Result);
1567 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1568 bool Changed = false;
1570 /// Orders the operands of the compare so that they are listed from most
1571 /// complex to least complex. This puts constants before unary operators,
1572 /// before binary operators.
1573 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
1578 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1580 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1581 return ReplaceInstUsesWith(I, V);
1583 const Type *Ty = Op0->getType();
1585 // icmp's with boolean values can always be turned into bitwise operations
1586 if (Ty == Type::getInt1Ty(I.getContext())) {
1587 switch (I.getPredicate()) {
1588 default: llvm_unreachable("Invalid icmp instruction!");
1589 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1590 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1591 return BinaryOperator::CreateNot(Xor);
1593 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1594 return BinaryOperator::CreateXor(Op0, Op1);
1596 case ICmpInst::ICMP_UGT:
1597 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1599 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1600 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1601 return BinaryOperator::CreateAnd(Not, Op1);
1603 case ICmpInst::ICMP_SGT:
1604 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1606 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1607 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1608 return BinaryOperator::CreateAnd(Not, Op0);
1610 case ICmpInst::ICMP_UGE:
1611 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1613 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1614 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1615 return BinaryOperator::CreateOr(Not, Op1);
1617 case ICmpInst::ICMP_SGE:
1618 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1620 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1621 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1622 return BinaryOperator::CreateOr(Not, Op0);
1627 unsigned BitWidth = 0;
1629 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1630 else if (Ty->isIntOrIntVector())
1631 BitWidth = Ty->getScalarSizeInBits();
1633 bool isSignBit = false;
1635 // See if we are doing a comparison with a constant.
1636 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1637 Value *A = 0, *B = 0;
1639 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1640 if (I.isEquality() && CI->isZero() &&
1641 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1642 // (icmp cond A B) if cond is equality
1643 return new ICmpInst(I.getPredicate(), A, B);
1646 // If we have an icmp le or icmp ge instruction, turn it into the
1647 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1648 // them being folded in the code below. The SimplifyICmpInst code has
1649 // already handled the edge cases for us, so we just assert on them.
1650 switch (I.getPredicate()) {
1652 case ICmpInst::ICMP_ULE:
1653 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1654 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1655 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1656 case ICmpInst::ICMP_SLE:
1657 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1658 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1659 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1660 case ICmpInst::ICMP_UGE:
1661 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1662 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1663 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1664 case ICmpInst::ICMP_SGE:
1665 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1666 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1667 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1670 // If this comparison is a normal comparison, it demands all
1671 // bits, if it is a sign bit comparison, it only demands the sign bit.
1673 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1676 // See if we can fold the comparison based on range information we can get
1677 // by checking whether bits are known to be zero or one in the input.
1678 if (BitWidth != 0) {
1679 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1680 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1682 if (SimplifyDemandedBits(I.getOperandUse(0),
1683 isSignBit ? APInt::getSignBit(BitWidth)
1684 : APInt::getAllOnesValue(BitWidth),
1685 Op0KnownZero, Op0KnownOne, 0))
1687 if (SimplifyDemandedBits(I.getOperandUse(1),
1688 APInt::getAllOnesValue(BitWidth),
1689 Op1KnownZero, Op1KnownOne, 0))
1692 // Given the known and unknown bits, compute a range that the LHS could be
1693 // in. Compute the Min, Max and RHS values based on the known bits. For the
1694 // EQ and NE we use unsigned values.
1695 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1696 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1698 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1700 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1703 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1705 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1709 // If Min and Max are known to be the same, then SimplifyDemandedBits
1710 // figured out that the LHS is a constant. Just constant fold this now so
1711 // that code below can assume that Min != Max.
1712 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1713 return new ICmpInst(I.getPredicate(),
1714 ConstantInt::get(I.getContext(), Op0Min), Op1);
1715 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1716 return new ICmpInst(I.getPredicate(), Op0,
1717 ConstantInt::get(I.getContext(), Op1Min));
1719 // Based on the range information we know about the LHS, see if we can
1720 // simplify this comparison. For example, (x&4) < 8 is always true.
1721 switch (I.getPredicate()) {
1722 default: llvm_unreachable("Unknown icmp opcode!");
1723 case ICmpInst::ICMP_EQ:
1724 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1725 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1727 case ICmpInst::ICMP_NE:
1728 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1729 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1731 case ICmpInst::ICMP_ULT:
1732 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
1733 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1734 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
1735 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1736 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
1737 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1738 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1739 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
1740 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1741 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1743 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
1744 if (CI->isMinValue(true))
1745 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1746 Constant::getAllOnesValue(Op0->getType()));
1749 case ICmpInst::ICMP_UGT:
1750 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
1751 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1752 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
1753 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1755 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
1756 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1757 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1758 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
1759 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1760 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1762 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
1763 if (CI->isMaxValue(true))
1764 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1765 Constant::getNullValue(Op0->getType()));
1768 case ICmpInst::ICMP_SLT:
1769 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
1770 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1771 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
1772 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1773 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
1774 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1775 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1776 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
1777 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1778 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1781 case ICmpInst::ICMP_SGT:
1782 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
1783 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1784 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
1785 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1787 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
1788 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1789 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1790 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
1791 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1792 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1795 case ICmpInst::ICMP_SGE:
1796 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
1797 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
1798 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1799 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
1800 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1802 case ICmpInst::ICMP_SLE:
1803 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
1804 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
1805 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1806 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
1807 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1809 case ICmpInst::ICMP_UGE:
1810 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
1811 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
1812 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1813 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
1814 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1816 case ICmpInst::ICMP_ULE:
1817 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
1818 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
1819 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1820 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
1821 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1825 // Turn a signed comparison into an unsigned one if both operands
1826 // are known to have the same sign.
1828 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
1829 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
1830 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
1833 // Test if the ICmpInst instruction is used exclusively by a select as
1834 // part of a minimum or maximum operation. If so, refrain from doing
1835 // any other folding. This helps out other analyses which understand
1836 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
1837 // and CodeGen. And in this case, at least one of the comparison
1838 // operands has at least one user besides the compare (the select),
1839 // which would often largely negate the benefit of folding anyway.
1841 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
1842 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
1843 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
1846 // See if we are doing a comparison between a constant and an instruction that
1847 // can be folded into the comparison.
1848 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1849 // Since the RHS is a ConstantInt (CI), if the left hand side is an
1850 // instruction, see if that instruction also has constants so that the
1851 // instruction can be folded into the icmp
1852 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
1853 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
1857 // Handle icmp with constant (but not simple integer constant) RHS
1858 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
1859 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
1860 switch (LHSI->getOpcode()) {
1861 case Instruction::GetElementPtr:
1862 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
1863 if (RHSC->isNullValue() &&
1864 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
1865 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
1866 Constant::getNullValue(LHSI->getOperand(0)->getType()));
1868 case Instruction::PHI:
1869 // Only fold icmp into the PHI if the phi and icmp are in the same
1870 // block. If in the same block, we're encouraging jump threading. If
1871 // not, we are just pessimizing the code by making an i1 phi.
1872 if (LHSI->getParent() == I.getParent())
1873 if (Instruction *NV = FoldOpIntoPhi(I, true))
1876 case Instruction::Select: {
1877 // If either operand of the select is a constant, we can fold the
1878 // comparison into the select arms, which will cause one to be
1879 // constant folded and the select turned into a bitwise or.
1880 Value *Op1 = 0, *Op2 = 0;
1881 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
1882 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
1883 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
1884 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
1886 // We only want to perform this transformation if it will not lead to
1887 // additional code. This is true if either both sides of the select
1888 // fold to a constant (in which case the icmp is replaced with a select
1889 // which will usually simplify) or this is the only user of the
1890 // select (in which case we are trading a select+icmp for a simpler
1892 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
1894 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
1897 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
1899 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
1903 case Instruction::Call:
1904 // If we have (malloc != null), and if the malloc has a single use, we
1905 // can assume it is successful and remove the malloc.
1906 if (isMalloc(LHSI) && LHSI->hasOneUse() &&
1907 isa<ConstantPointerNull>(RHSC)) {
1908 // Need to explicitly erase malloc call here, instead of adding it to
1909 // Worklist, because it won't get DCE'd from the Worklist since
1910 // isInstructionTriviallyDead() returns false for function calls.
1911 // It is OK to replace LHSI/MallocCall with Undef because the
1912 // instruction that uses it will be erased via Worklist.
1913 if (extractMallocCall(LHSI)) {
1914 LHSI->replaceAllUsesWith(UndefValue::get(LHSI->getType()));
1915 EraseInstFromFunction(*LHSI);
1916 return ReplaceInstUsesWith(I,
1917 ConstantInt::get(Type::getInt1Ty(I.getContext()),
1918 !I.isTrueWhenEqual()));
1920 if (CallInst* MallocCall = extractMallocCallFromBitCast(LHSI))
1921 if (MallocCall->hasOneUse()) {
1922 MallocCall->replaceAllUsesWith(
1923 UndefValue::get(MallocCall->getType()));
1924 EraseInstFromFunction(*MallocCall);
1925 Worklist.Add(LHSI); // The malloc's bitcast use.
1926 return ReplaceInstUsesWith(I,
1927 ConstantInt::get(Type::getInt1Ty(I.getContext()),
1928 !I.isTrueWhenEqual()));
1932 case Instruction::IntToPtr:
1933 // icmp pred inttoptr(X), null -> icmp pred X, 0
1934 if (RHSC->isNullValue() && TD &&
1935 TD->getIntPtrType(RHSC->getContext()) ==
1936 LHSI->getOperand(0)->getType())
1937 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
1938 Constant::getNullValue(LHSI->getOperand(0)->getType()));
1941 case Instruction::Load:
1942 // Try to optimize things like "A[i] > 4" to index computations.
1943 if (GetElementPtrInst *GEP =
1944 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
1945 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1946 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1947 !cast<LoadInst>(LHSI)->isVolatile())
1948 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
1955 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
1956 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
1957 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
1959 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
1960 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
1961 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
1964 // Test to see if the operands of the icmp are casted versions of other
1965 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
1967 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
1968 if (isa<PointerType>(Op0->getType()) &&
1969 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
1970 // We keep moving the cast from the left operand over to the right
1971 // operand, where it can often be eliminated completely.
1972 Op0 = CI->getOperand(0);
1974 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
1975 // so eliminate it as well.
1976 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
1977 Op1 = CI2->getOperand(0);
1979 // If Op1 is a constant, we can fold the cast into the constant.
1980 if (Op0->getType() != Op1->getType()) {
1981 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1982 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
1984 // Otherwise, cast the RHS right before the icmp
1985 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
1988 return new ICmpInst(I.getPredicate(), Op0, Op1);
1992 if (isa<CastInst>(Op0)) {
1993 // Handle the special case of: icmp (cast bool to X), <cst>
1994 // This comes up when you have code like
1997 // For generality, we handle any zero-extension of any operand comparison
1998 // with a constant or another cast from the same type.
1999 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2000 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2004 // See if it's the same type of instruction on the left and right.
2005 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2006 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2007 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
2008 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
2009 switch (Op0I->getOpcode()) {
2011 case Instruction::Add:
2012 case Instruction::Sub:
2013 case Instruction::Xor:
2014 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2015 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
2016 Op1I->getOperand(0));
2017 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2018 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2019 if (CI->getValue().isSignBit()) {
2020 ICmpInst::Predicate Pred = I.isSigned()
2021 ? I.getUnsignedPredicate()
2022 : I.getSignedPredicate();
2023 return new ICmpInst(Pred, Op0I->getOperand(0),
2024 Op1I->getOperand(0));
2027 if (CI->getValue().isMaxSignedValue()) {
2028 ICmpInst::Predicate Pred = I.isSigned()
2029 ? I.getUnsignedPredicate()
2030 : I.getSignedPredicate();
2031 Pred = I.getSwappedPredicate(Pred);
2032 return new ICmpInst(Pred, Op0I->getOperand(0),
2033 Op1I->getOperand(0));
2037 case Instruction::Mul:
2038 if (!I.isEquality())
2041 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2042 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2043 // Mask = -1 >> count-trailing-zeros(Cst).
2044 if (!CI->isZero() && !CI->isOne()) {
2045 const APInt &AP = CI->getValue();
2046 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2047 APInt::getLowBitsSet(AP.getBitWidth(),
2049 AP.countTrailingZeros()));
2050 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
2051 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
2052 return new ICmpInst(I.getPredicate(), And1, And2);
2061 // ~x < ~y --> y < x
2063 if (match(Op0, m_Not(m_Value(A))) &&
2064 match(Op1, m_Not(m_Value(B))))
2065 return new ICmpInst(I.getPredicate(), B, A);
2068 if (I.isEquality()) {
2069 Value *A, *B, *C, *D;
2071 // -x == -y --> x == y
2072 if (match(Op0, m_Neg(m_Value(A))) &&
2073 match(Op1, m_Neg(m_Value(B))))
2074 return new ICmpInst(I.getPredicate(), A, B);
2076 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2077 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2078 Value *OtherVal = A == Op1 ? B : A;
2079 return new ICmpInst(I.getPredicate(), OtherVal,
2080 Constant::getNullValue(A->getType()));
2083 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2084 // A^c1 == C^c2 --> A == C^(c1^c2)
2085 ConstantInt *C1, *C2;
2086 if (match(B, m_ConstantInt(C1)) &&
2087 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2088 Constant *NC = ConstantInt::get(I.getContext(),
2089 C1->getValue() ^ C2->getValue());
2090 Value *Xor = Builder->CreateXor(C, NC, "tmp");
2091 return new ICmpInst(I.getPredicate(), A, Xor);
2094 // A^B == A^D -> B == D
2095 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2096 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2097 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2098 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2102 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2103 (A == Op0 || B == Op0)) {
2104 // A == (A^B) -> B == 0
2105 Value *OtherVal = A == Op0 ? B : A;
2106 return new ICmpInst(I.getPredicate(), OtherVal,
2107 Constant::getNullValue(A->getType()));
2110 // (A-B) == A -> B == 0
2111 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
2112 return new ICmpInst(I.getPredicate(), B,
2113 Constant::getNullValue(B->getType()));
2115 // A == (A-B) -> B == 0
2116 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
2117 return new ICmpInst(I.getPredicate(), B,
2118 Constant::getNullValue(B->getType()));
2120 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2121 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2122 match(Op0, m_And(m_Value(A), m_Value(B))) &&
2123 match(Op1, m_And(m_Value(C), m_Value(D)))) {
2124 Value *X = 0, *Y = 0, *Z = 0;
2127 X = B; Y = D; Z = A;
2128 } else if (A == D) {
2129 X = B; Y = C; Z = A;
2130 } else if (B == C) {
2131 X = A; Y = D; Z = B;
2132 } else if (B == D) {
2133 X = A; Y = C; Z = B;
2136 if (X) { // Build (X^Y) & Z
2137 Op1 = Builder->CreateXor(X, Y, "tmp");
2138 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
2139 I.setOperand(0, Op1);
2140 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2147 Value *X; ConstantInt *Cst;
2149 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2150 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2153 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2154 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2156 return Changed ? &I : 0;
2164 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2166 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2169 if (!isa<ConstantFP>(RHSC)) return 0;
2170 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2172 // Get the width of the mantissa. We don't want to hack on conversions that
2173 // might lose information from the integer, e.g. "i64 -> float"
2174 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2175 if (MantissaWidth == -1) return 0; // Unknown.
2177 // Check to see that the input is converted from an integer type that is small
2178 // enough that preserves all bits. TODO: check here for "known" sign bits.
2179 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2180 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2182 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2183 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2187 // If the conversion would lose info, don't hack on this.
2188 if ((int)InputSize > MantissaWidth)
2191 // Otherwise, we can potentially simplify the comparison. We know that it
2192 // will always come through as an integer value and we know the constant is
2193 // not a NAN (it would have been previously simplified).
2194 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2196 ICmpInst::Predicate Pred;
2197 switch (I.getPredicate()) {
2198 default: llvm_unreachable("Unexpected predicate!");
2199 case FCmpInst::FCMP_UEQ:
2200 case FCmpInst::FCMP_OEQ:
2201 Pred = ICmpInst::ICMP_EQ;
2203 case FCmpInst::FCMP_UGT:
2204 case FCmpInst::FCMP_OGT:
2205 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2207 case FCmpInst::FCMP_UGE:
2208 case FCmpInst::FCMP_OGE:
2209 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2211 case FCmpInst::FCMP_ULT:
2212 case FCmpInst::FCMP_OLT:
2213 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2215 case FCmpInst::FCMP_ULE:
2216 case FCmpInst::FCMP_OLE:
2217 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2219 case FCmpInst::FCMP_UNE:
2220 case FCmpInst::FCMP_ONE:
2221 Pred = ICmpInst::ICMP_NE;
2223 case FCmpInst::FCMP_ORD:
2224 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2225 case FCmpInst::FCMP_UNO:
2226 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2229 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2231 // Now we know that the APFloat is a normal number, zero or inf.
2233 // See if the FP constant is too large for the integer. For example,
2234 // comparing an i8 to 300.0.
2235 unsigned IntWidth = IntTy->getScalarSizeInBits();
2238 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2239 // and large values.
2240 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2241 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2242 APFloat::rmNearestTiesToEven);
2243 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2244 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2245 Pred == ICmpInst::ICMP_SLE)
2246 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2247 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2250 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2251 // +INF and large values.
2252 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2253 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2254 APFloat::rmNearestTiesToEven);
2255 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2256 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2257 Pred == ICmpInst::ICMP_ULE)
2258 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2259 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2264 // See if the RHS value is < SignedMin.
2265 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2266 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2267 APFloat::rmNearestTiesToEven);
2268 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2269 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2270 Pred == ICmpInst::ICMP_SGE)
2271 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2272 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2276 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2277 // [0, UMAX], but it may still be fractional. See if it is fractional by
2278 // casting the FP value to the integer value and back, checking for equality.
2279 // Don't do this for zero, because -0.0 is not fractional.
2280 Constant *RHSInt = LHSUnsigned
2281 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2282 : ConstantExpr::getFPToSI(RHSC, IntTy);
2283 if (!RHS.isZero()) {
2284 bool Equal = LHSUnsigned
2285 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2286 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2288 // If we had a comparison against a fractional value, we have to adjust
2289 // the compare predicate and sometimes the value. RHSC is rounded towards
2290 // zero at this point.
2292 default: llvm_unreachable("Unexpected integer comparison!");
2293 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2294 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2295 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2296 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2297 case ICmpInst::ICMP_ULE:
2298 // (float)int <= 4.4 --> int <= 4
2299 // (float)int <= -4.4 --> false
2300 if (RHS.isNegative())
2301 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2303 case ICmpInst::ICMP_SLE:
2304 // (float)int <= 4.4 --> int <= 4
2305 // (float)int <= -4.4 --> int < -4
2306 if (RHS.isNegative())
2307 Pred = ICmpInst::ICMP_SLT;
2309 case ICmpInst::ICMP_ULT:
2310 // (float)int < -4.4 --> false
2311 // (float)int < 4.4 --> int <= 4
2312 if (RHS.isNegative())
2313 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2314 Pred = ICmpInst::ICMP_ULE;
2316 case ICmpInst::ICMP_SLT:
2317 // (float)int < -4.4 --> int < -4
2318 // (float)int < 4.4 --> int <= 4
2319 if (!RHS.isNegative())
2320 Pred = ICmpInst::ICMP_SLE;
2322 case ICmpInst::ICMP_UGT:
2323 // (float)int > 4.4 --> int > 4
2324 // (float)int > -4.4 --> true
2325 if (RHS.isNegative())
2326 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2328 case ICmpInst::ICMP_SGT:
2329 // (float)int > 4.4 --> int > 4
2330 // (float)int > -4.4 --> int >= -4
2331 if (RHS.isNegative())
2332 Pred = ICmpInst::ICMP_SGE;
2334 case ICmpInst::ICMP_UGE:
2335 // (float)int >= -4.4 --> true
2336 // (float)int >= 4.4 --> int > 4
2337 if (!RHS.isNegative())
2338 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2339 Pred = ICmpInst::ICMP_UGT;
2341 case ICmpInst::ICMP_SGE:
2342 // (float)int >= -4.4 --> int >= -4
2343 // (float)int >= 4.4 --> int > 4
2344 if (!RHS.isNegative())
2345 Pred = ICmpInst::ICMP_SGT;
2351 // Lower this FP comparison into an appropriate integer version of the
2353 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2356 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2357 bool Changed = false;
2359 /// Orders the operands of the compare so that they are listed from most
2360 /// complex to least complex. This puts constants before unary operators,
2361 /// before binary operators.
2362 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2367 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2369 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2370 return ReplaceInstUsesWith(I, V);
2372 // Simplify 'fcmp pred X, X'
2374 switch (I.getPredicate()) {
2375 default: llvm_unreachable("Unknown predicate!");
2376 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2377 case FCmpInst::FCMP_ULT: // True if unordered or less than
2378 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2379 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2380 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2381 I.setPredicate(FCmpInst::FCMP_UNO);
2382 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2385 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2386 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2387 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2388 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2389 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2390 I.setPredicate(FCmpInst::FCMP_ORD);
2391 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2396 // Handle fcmp with constant RHS
2397 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2398 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2399 switch (LHSI->getOpcode()) {
2400 case Instruction::PHI:
2401 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2402 // block. If in the same block, we're encouraging jump threading. If
2403 // not, we are just pessimizing the code by making an i1 phi.
2404 if (LHSI->getParent() == I.getParent())
2405 if (Instruction *NV = FoldOpIntoPhi(I, true))
2408 case Instruction::SIToFP:
2409 case Instruction::UIToFP:
2410 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2413 case Instruction::Select: {
2414 // If either operand of the select is a constant, we can fold the
2415 // comparison into the select arms, which will cause one to be
2416 // constant folded and the select turned into a bitwise or.
2417 Value *Op1 = 0, *Op2 = 0;
2418 if (LHSI->hasOneUse()) {
2419 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2420 // Fold the known value into the constant operand.
2421 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2422 // Insert a new FCmp of the other select operand.
2423 Op2 = Builder->CreateFCmp(I.getPredicate(),
2424 LHSI->getOperand(2), RHSC, I.getName());
2425 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2426 // Fold the known value into the constant operand.
2427 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2428 // Insert a new FCmp of the other select operand.
2429 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2435 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2438 case Instruction::Load:
2439 if (GetElementPtrInst *GEP =
2440 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2441 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2442 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2443 !cast<LoadInst>(LHSI)->isVolatile())
2444 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2451 return Changed ? &I : 0;