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.setBit(Min.getBitWidth()-1);
164 Max.clearBit(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()));
717 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
718 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
721 // (X+1) >u X --> X <u (0-1) --> X != 255
722 // (X+2) >u X --> X <u (0-2) --> X <u 254
723 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
724 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
725 // If this is an NUW add, then this is always true.
727 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
728 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
731 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
732 ConstantInt *SMax = ConstantInt::get(X->getContext(),
733 APInt::getSignedMaxValue(BitWidth));
735 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
736 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
737 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
738 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
739 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
740 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
741 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
742 // If this is an NSW add, then we have two cases: if the constant is
743 // positive, then this is always false, if negative, this is always true.
745 bool isTrue = CI->getValue().isNegative();
746 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
749 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
752 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
753 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
754 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
755 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
756 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
757 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
759 // If this is an NSW add, then we have two cases: if the constant is
760 // positive, then this is always true, if negative, this is always false.
762 bool isTrue = !CI->getValue().isNegative();
763 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
766 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
767 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
768 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
771 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
772 /// and CmpRHS are both known to be integer constants.
773 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
774 ConstantInt *DivRHS) {
775 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
776 const APInt &CmpRHSV = CmpRHS->getValue();
778 // FIXME: If the operand types don't match the type of the divide
779 // then don't attempt this transform. The code below doesn't have the
780 // logic to deal with a signed divide and an unsigned compare (and
781 // vice versa). This is because (x /s C1) <s C2 produces different
782 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
783 // (x /u C1) <u C2. Simply casting the operands and result won't
784 // work. :( The if statement below tests that condition and bails
786 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
787 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
789 if (DivRHS->isZero())
790 return 0; // The ProdOV computation fails on divide by zero.
791 if (DivIsSigned && DivRHS->isAllOnesValue())
792 return 0; // The overflow computation also screws up here
794 return 0; // Not worth bothering, and eliminates some funny cases
797 // Compute Prod = CI * DivRHS. We are essentially solving an equation
798 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
799 // C2 (CI). By solving for X we can turn this into a range check
800 // instead of computing a divide.
801 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
803 // Determine if the product overflows by seeing if the product is
804 // not equal to the divide. Make sure we do the same kind of divide
805 // as in the LHS instruction that we're folding.
806 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
807 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
809 // Get the ICmp opcode
810 ICmpInst::Predicate Pred = ICI.getPredicate();
812 // Figure out the interval that is being checked. For example, a comparison
813 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
814 // Compute this interval based on the constants involved and the signedness of
815 // the compare/divide. This computes a half-open interval, keeping track of
816 // whether either value in the interval overflows. After analysis each
817 // overflow variable is set to 0 if it's corresponding bound variable is valid
818 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
819 int LoOverflow = 0, HiOverflow = 0;
820 Constant *LoBound = 0, *HiBound = 0;
822 if (!DivIsSigned) { // udiv
823 // e.g. X/5 op 3 --> [15, 20)
825 HiOverflow = LoOverflow = ProdOV;
827 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
828 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
829 if (CmpRHSV == 0) { // (X / pos) op 0
830 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
831 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
833 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
834 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
835 HiOverflow = LoOverflow = ProdOV;
837 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
838 } else { // (X / pos) op neg
839 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
840 HiBound = AddOne(Prod);
841 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
843 ConstantInt* DivNeg =
844 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
845 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
848 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
849 if (CmpRHSV == 0) { // (X / neg) op 0
850 // e.g. X/-5 op 0 --> [-4, 5)
851 LoBound = AddOne(DivRHS);
852 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
853 if (HiBound == DivRHS) { // -INTMIN = INTMIN
854 HiOverflow = 1; // [INTMIN+1, overflow)
855 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
857 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
858 // e.g. X/-5 op 3 --> [-19, -14)
859 HiBound = AddOne(Prod);
860 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
862 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
863 } else { // (X / neg) op neg
864 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
865 LoOverflow = HiOverflow = ProdOV;
867 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
870 // Dividing by a negative swaps the condition. LT <-> GT
871 Pred = ICmpInst::getSwappedPredicate(Pred);
874 Value *X = DivI->getOperand(0);
876 default: llvm_unreachable("Unhandled icmp opcode!");
877 case ICmpInst::ICMP_EQ:
878 if (LoOverflow && HiOverflow)
879 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
881 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
882 ICmpInst::ICMP_UGE, X, LoBound);
884 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
885 ICmpInst::ICMP_ULT, X, HiBound);
886 return ReplaceInstUsesWith(ICI,
887 InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
889 case ICmpInst::ICMP_NE:
890 if (LoOverflow && HiOverflow)
891 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
893 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
894 ICmpInst::ICMP_ULT, X, LoBound);
896 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
897 ICmpInst::ICMP_UGE, X, HiBound);
898 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
899 DivIsSigned, false));
900 case ICmpInst::ICMP_ULT:
901 case ICmpInst::ICMP_SLT:
902 if (LoOverflow == +1) // Low bound is greater than input range.
903 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
904 if (LoOverflow == -1) // Low bound is less than input range.
905 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
906 return new ICmpInst(Pred, X, LoBound);
907 case ICmpInst::ICMP_UGT:
908 case ICmpInst::ICMP_SGT:
909 if (HiOverflow == +1) // High bound greater than input range.
910 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
911 else if (HiOverflow == -1) // High bound less than input range.
912 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
913 if (Pred == ICmpInst::ICMP_UGT)
914 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
916 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
921 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
923 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
926 const APInt &RHSV = RHS->getValue();
928 switch (LHSI->getOpcode()) {
929 case Instruction::Trunc:
930 if (ICI.isEquality() && LHSI->hasOneUse()) {
931 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
932 // of the high bits truncated out of x are known.
933 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
934 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
935 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
936 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
937 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
939 // If all the high bits are known, we can do this xform.
940 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
941 // Pull in the high bits from known-ones set.
942 APInt NewRHS = RHS->getValue().zext(SrcBits);
944 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
945 ConstantInt::get(ICI.getContext(), NewRHS));
950 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
951 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
952 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
954 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
955 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
956 Value *CompareVal = LHSI->getOperand(0);
958 // If the sign bit of the XorCST is not set, there is no change to
959 // the operation, just stop using the Xor.
960 if (!XorCST->getValue().isNegative()) {
961 ICI.setOperand(0, CompareVal);
966 // Was the old condition true if the operand is positive?
967 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
969 // If so, the new one isn't.
970 isTrueIfPositive ^= true;
972 if (isTrueIfPositive)
973 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
976 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
980 if (LHSI->hasOneUse()) {
981 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
982 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
983 const APInt &SignBit = XorCST->getValue();
984 ICmpInst::Predicate Pred = ICI.isSigned()
985 ? ICI.getUnsignedPredicate()
986 : ICI.getSignedPredicate();
987 return new ICmpInst(Pred, LHSI->getOperand(0),
988 ConstantInt::get(ICI.getContext(),
992 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
993 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
994 const APInt &NotSignBit = XorCST->getValue();
995 ICmpInst::Predicate Pred = ICI.isSigned()
996 ? ICI.getUnsignedPredicate()
997 : ICI.getSignedPredicate();
998 Pred = ICI.getSwappedPredicate(Pred);
999 return new ICmpInst(Pred, LHSI->getOperand(0),
1000 ConstantInt::get(ICI.getContext(),
1001 RHSV ^ NotSignBit));
1006 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1007 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1008 LHSI->getOperand(0)->hasOneUse()) {
1009 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1011 // If the LHS is an AND of a truncating cast, we can widen the
1012 // and/compare to be the input width without changing the value
1013 // produced, eliminating a cast.
1014 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1015 // We can do this transformation if either the AND constant does not
1016 // have its sign bit set or if it is an equality comparison.
1017 // Extending a relational comparison when we're checking the sign
1018 // bit would not work.
1019 if (Cast->hasOneUse() &&
1020 (ICI.isEquality() ||
1021 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
1023 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
1024 APInt NewCST = AndCST->getValue().zext(BitWidth);
1025 APInt NewCI = RHSV.zext(BitWidth);
1027 Builder->CreateAnd(Cast->getOperand(0),
1028 ConstantInt::get(ICI.getContext(), NewCST),
1030 return new ICmpInst(ICI.getPredicate(), NewAnd,
1031 ConstantInt::get(ICI.getContext(), NewCI));
1035 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1036 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1037 // happens a LOT in code produced by the C front-end, for bitfield
1039 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1040 if (Shift && !Shift->isShift())
1044 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1045 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1046 const Type *AndTy = AndCST->getType(); // Type of the and.
1048 // We can fold this as long as we can't shift unknown bits
1049 // into the mask. This can only happen with signed shift
1050 // rights, as they sign-extend.
1052 bool CanFold = Shift->isLogicalShift();
1054 // To test for the bad case of the signed shr, see if any
1055 // of the bits shifted in could be tested after the mask.
1056 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1057 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1059 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1060 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1061 AndCST->getValue()) == 0)
1067 if (Shift->getOpcode() == Instruction::Shl)
1068 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1070 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1072 // Check to see if we are shifting out any of the bits being
1074 if (ConstantExpr::get(Shift->getOpcode(),
1075 NewCst, ShAmt) != RHS) {
1076 // If we shifted bits out, the fold is not going to work out.
1077 // As a special case, check to see if this means that the
1078 // result is always true or false now.
1079 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1080 return ReplaceInstUsesWith(ICI,
1081 ConstantInt::getFalse(ICI.getContext()));
1082 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1083 return ReplaceInstUsesWith(ICI,
1084 ConstantInt::getTrue(ICI.getContext()));
1086 ICI.setOperand(1, NewCst);
1087 Constant *NewAndCST;
1088 if (Shift->getOpcode() == Instruction::Shl)
1089 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1091 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1092 LHSI->setOperand(1, NewAndCST);
1093 LHSI->setOperand(0, Shift->getOperand(0));
1094 Worklist.Add(Shift); // Shift is dead.
1100 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1101 // preferable because it allows the C<<Y expression to be hoisted out
1102 // of a loop if Y is invariant and X is not.
1103 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1104 ICI.isEquality() && !Shift->isArithmeticShift() &&
1105 !isa<Constant>(Shift->getOperand(0))) {
1108 if (Shift->getOpcode() == Instruction::LShr) {
1109 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
1111 // Insert a logical shift.
1112 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
1115 // Compute X & (C << Y).
1117 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1119 ICI.setOperand(0, NewAnd);
1124 // Try to optimize things like "A[i]&42 == 0" to index computations.
1125 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1126 if (GetElementPtrInst *GEP =
1127 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1128 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1129 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1130 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1131 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1132 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1138 case Instruction::Or: {
1139 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1142 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1143 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1144 // -> and (icmp eq P, null), (icmp eq Q, null).
1146 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1147 Constant::getNullValue(P->getType()));
1148 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1149 Constant::getNullValue(Q->getType()));
1151 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1152 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1154 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1160 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1161 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1164 uint32_t TypeBits = RHSV.getBitWidth();
1166 // Check that the shift amount is in range. If not, don't perform
1167 // undefined shifts. When the shift is visited it will be
1169 if (ShAmt->uge(TypeBits))
1172 if (ICI.isEquality()) {
1173 // If we are comparing against bits always shifted out, the
1174 // comparison cannot succeed.
1176 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1178 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1179 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1181 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1182 return ReplaceInstUsesWith(ICI, Cst);
1185 if (LHSI->hasOneUse()) {
1186 // Otherwise strength reduce the shift into an and.
1187 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1189 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1190 TypeBits-ShAmtVal));
1193 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1194 return new ICmpInst(ICI.getPredicate(), And,
1195 ConstantInt::get(ICI.getContext(),
1196 RHSV.lshr(ShAmtVal)));
1200 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1201 bool TrueIfSigned = false;
1202 if (LHSI->hasOneUse() &&
1203 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1204 // (X << 31) <s 0 --> (X&1) != 0
1205 Constant *Mask = ConstantInt::get(ICI.getContext(), APInt(TypeBits, 1) <<
1206 (TypeBits-ShAmt->getZExtValue()-1));
1208 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1209 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1210 And, Constant::getNullValue(And->getType()));
1215 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1216 case Instruction::AShr: {
1217 // Only handle equality comparisons of shift-by-constant.
1218 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1219 if (!ShAmt || !ICI.isEquality()) break;
1221 // Check that the shift amount is in range. If not, don't perform
1222 // undefined shifts. When the shift is visited it will be
1224 uint32_t TypeBits = RHSV.getBitWidth();
1225 if (ShAmt->uge(TypeBits))
1228 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1230 // If we are comparing against bits always shifted out, the
1231 // comparison cannot succeed.
1232 APInt Comp = RHSV << ShAmtVal;
1233 if (LHSI->getOpcode() == Instruction::LShr)
1234 Comp = Comp.lshr(ShAmtVal);
1236 Comp = Comp.ashr(ShAmtVal);
1238 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
1239 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1240 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1242 return ReplaceInstUsesWith(ICI, Cst);
1245 // Otherwise, check to see if the bits shifted out are known to be zero.
1246 // If so, we can compare against the unshifted value:
1247 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1248 if (LHSI->hasOneUse() &&
1249 MaskedValueIsZero(LHSI->getOperand(0),
1250 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
1251 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1252 ConstantExpr::getShl(RHS, ShAmt));
1255 if (LHSI->hasOneUse()) {
1256 // Otherwise strength reduce the shift into an and.
1257 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1258 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
1260 Value *And = Builder->CreateAnd(LHSI->getOperand(0),
1261 Mask, LHSI->getName()+".mask");
1262 return new ICmpInst(ICI.getPredicate(), And,
1263 ConstantExpr::getShl(RHS, ShAmt));
1268 case Instruction::SDiv:
1269 case Instruction::UDiv:
1270 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1271 // Fold this div into the comparison, producing a range check.
1272 // Determine, based on the divide type, what the range is being
1273 // checked. If there is an overflow on the low or high side, remember
1274 // it, otherwise compute the range [low, hi) bounding the new value.
1275 // See: InsertRangeTest above for the kinds of replacements possible.
1276 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1277 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1282 case Instruction::Add:
1283 // Fold: icmp pred (add X, C1), C2
1284 if (!ICI.isEquality()) {
1285 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1287 const APInt &LHSV = LHSC->getValue();
1289 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1292 if (ICI.isSigned()) {
1293 if (CR.getLower().isSignBit()) {
1294 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1295 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1296 } else if (CR.getUpper().isSignBit()) {
1297 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1298 ConstantInt::get(ICI.getContext(),CR.getLower()));
1301 if (CR.getLower().isMinValue()) {
1302 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1303 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1304 } else if (CR.getUpper().isMinValue()) {
1305 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1306 ConstantInt::get(ICI.getContext(),CR.getLower()));
1313 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1314 if (ICI.isEquality()) {
1315 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1317 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1318 // the second operand is a constant, simplify a bit.
1319 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1320 switch (BO->getOpcode()) {
1321 case Instruction::SRem:
1322 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1323 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1324 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1325 if (V.sgt(1) && V.isPowerOf2()) {
1327 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1329 return new ICmpInst(ICI.getPredicate(), NewRem,
1330 Constant::getNullValue(BO->getType()));
1334 case Instruction::Add:
1335 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1336 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1337 if (BO->hasOneUse())
1338 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1339 ConstantExpr::getSub(RHS, BOp1C));
1340 } else if (RHSV == 0) {
1341 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1342 // efficiently invertible, or if the add has just this one use.
1343 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1345 if (Value *NegVal = dyn_castNegVal(BOp1))
1346 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1347 else if (Value *NegVal = dyn_castNegVal(BOp0))
1348 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1349 else if (BO->hasOneUse()) {
1350 Value *Neg = Builder->CreateNeg(BOp1);
1352 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1356 case Instruction::Xor:
1357 // For the xor case, we can xor two constants together, eliminating
1358 // the explicit xor.
1359 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
1360 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1361 ConstantExpr::getXor(RHS, BOC));
1364 case Instruction::Sub:
1365 // Replace (([sub|xor] A, B) != 0) with (A != B)
1367 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1371 case Instruction::Or:
1372 // If bits are being or'd in that are not present in the constant we
1373 // are comparing against, then the comparison could never succeed!
1374 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1375 Constant *NotCI = ConstantExpr::getNot(RHS);
1376 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1377 return ReplaceInstUsesWith(ICI,
1378 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1383 case Instruction::And:
1384 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1385 // If bits are being compared against that are and'd out, then the
1386 // comparison can never succeed!
1387 if ((RHSV & ~BOC->getValue()) != 0)
1388 return ReplaceInstUsesWith(ICI,
1389 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1392 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1393 if (RHS == BOC && RHSV.isPowerOf2())
1394 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1395 ICmpInst::ICMP_NE, LHSI,
1396 Constant::getNullValue(RHS->getType()));
1398 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1399 if (BOC->getValue().isSignBit()) {
1400 Value *X = BO->getOperand(0);
1401 Constant *Zero = Constant::getNullValue(X->getType());
1402 ICmpInst::Predicate pred = isICMP_NE ?
1403 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1404 return new ICmpInst(pred, X, Zero);
1407 // ((X & ~7) == 0) --> X < 8
1408 if (RHSV == 0 && isHighOnes(BOC)) {
1409 Value *X = BO->getOperand(0);
1410 Constant *NegX = ConstantExpr::getNeg(BOC);
1411 ICmpInst::Predicate pred = isICMP_NE ?
1412 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1413 return new ICmpInst(pred, X, NegX);
1418 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1419 // Handle icmp {eq|ne} <intrinsic>, intcst.
1420 switch (II->getIntrinsicID()) {
1421 case Intrinsic::bswap:
1423 ICI.setOperand(0, II->getArgOperand(0));
1424 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1426 case Intrinsic::ctlz:
1427 case Intrinsic::cttz:
1428 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1429 if (RHSV == RHS->getType()->getBitWidth()) {
1431 ICI.setOperand(0, II->getArgOperand(0));
1432 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1436 case Intrinsic::ctpop:
1437 // popcount(A) == 0 -> A == 0 and likewise for !=
1438 if (RHS->isZero()) {
1440 ICI.setOperand(0, II->getArgOperand(0));
1441 ICI.setOperand(1, RHS);
1453 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1454 /// We only handle extending casts so far.
1456 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1457 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1458 Value *LHSCIOp = LHSCI->getOperand(0);
1459 const Type *SrcTy = LHSCIOp->getType();
1460 const Type *DestTy = LHSCI->getType();
1463 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1464 // integer type is the same size as the pointer type.
1465 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1466 TD->getPointerSizeInBits() ==
1467 cast<IntegerType>(DestTy)->getBitWidth()) {
1469 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1470 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1471 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1472 RHSOp = RHSC->getOperand(0);
1473 // If the pointer types don't match, insert a bitcast.
1474 if (LHSCIOp->getType() != RHSOp->getType())
1475 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1479 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1482 // The code below only handles extension cast instructions, so far.
1484 if (LHSCI->getOpcode() != Instruction::ZExt &&
1485 LHSCI->getOpcode() != Instruction::SExt)
1488 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1489 bool isSignedCmp = ICI.isSigned();
1491 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1492 // Not an extension from the same type?
1493 RHSCIOp = CI->getOperand(0);
1494 if (RHSCIOp->getType() != LHSCIOp->getType())
1497 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1498 // and the other is a zext), then we can't handle this.
1499 if (CI->getOpcode() != LHSCI->getOpcode())
1502 // Deal with equality cases early.
1503 if (ICI.isEquality())
1504 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1506 // A signed comparison of sign extended values simplifies into a
1507 // signed comparison.
1508 if (isSignedCmp && isSignedExt)
1509 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1511 // The other three cases all fold into an unsigned comparison.
1512 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1515 // If we aren't dealing with a constant on the RHS, exit early
1516 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1520 // Compute the constant that would happen if we truncated to SrcTy then
1521 // reextended to DestTy.
1522 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1523 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1526 // If the re-extended constant didn't change...
1528 // Deal with equality cases early.
1529 if (ICI.isEquality())
1530 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1532 // A signed comparison of sign extended values simplifies into a
1533 // signed comparison.
1534 if (isSignedExt && isSignedCmp)
1535 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1537 // The other three cases all fold into an unsigned comparison.
1538 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1541 // The re-extended constant changed so the constant cannot be represented
1542 // in the shorter type. Consequently, we cannot emit a simple comparison.
1544 // First, handle some easy cases. We know the result cannot be equal at this
1545 // point so handle the ICI.isEquality() cases
1546 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1547 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
1548 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1549 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
1551 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1552 // should have been folded away previously and not enter in here.
1555 // We're performing a signed comparison.
1556 if (cast<ConstantInt>(CI)->getValue().isNegative())
1557 Result = ConstantInt::getFalse(ICI.getContext()); // X < (small) --> false
1559 Result = ConstantInt::getTrue(ICI.getContext()); // X < (large) --> true
1561 // We're performing an unsigned comparison.
1563 // We're performing an unsigned comp with a sign extended value.
1564 // This is true if the input is >= 0. [aka >s -1]
1565 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1566 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1568 // Unsigned extend & unsigned compare -> always true.
1569 Result = ConstantInt::getTrue(ICI.getContext());
1573 // Finally, return the value computed.
1574 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
1575 ICI.getPredicate() == ICmpInst::ICMP_SLT)
1576 return ReplaceInstUsesWith(ICI, Result);
1578 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
1579 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
1580 "ICmp should be folded!");
1581 if (Constant *CI = dyn_cast<Constant>(Result))
1582 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
1583 return BinaryOperator::CreateNot(Result);
1588 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1589 bool Changed = false;
1590 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1592 /// Orders the operands of the compare so that they are listed from most
1593 /// complex to least complex. This puts constants before unary operators,
1594 /// before binary operators.
1595 if (getComplexity(Op0) < getComplexity(Op1)) {
1597 std::swap(Op0, Op1);
1601 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1602 return ReplaceInstUsesWith(I, V);
1604 const Type *Ty = Op0->getType();
1606 // icmp's with boolean values can always be turned into bitwise operations
1607 if (Ty->isIntegerTy(1)) {
1608 switch (I.getPredicate()) {
1609 default: llvm_unreachable("Invalid icmp instruction!");
1610 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1611 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1612 return BinaryOperator::CreateNot(Xor);
1614 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1615 return BinaryOperator::CreateXor(Op0, Op1);
1617 case ICmpInst::ICMP_UGT:
1618 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1620 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1621 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1622 return BinaryOperator::CreateAnd(Not, Op1);
1624 case ICmpInst::ICMP_SGT:
1625 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1627 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1628 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1629 return BinaryOperator::CreateAnd(Not, Op0);
1631 case ICmpInst::ICMP_UGE:
1632 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1634 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1635 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1636 return BinaryOperator::CreateOr(Not, Op1);
1638 case ICmpInst::ICMP_SGE:
1639 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1641 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1642 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1643 return BinaryOperator::CreateOr(Not, Op0);
1648 unsigned BitWidth = 0;
1650 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1651 else if (Ty->isIntOrIntVectorTy())
1652 BitWidth = Ty->getScalarSizeInBits();
1654 bool isSignBit = false;
1656 // See if we are doing a comparison with a constant.
1657 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1658 Value *A = 0, *B = 0;
1660 // Match the following pattern, which is a common idiom when writing
1661 // overflow-safe integer arithmetic function. The source performs an
1662 // addition in wider type, and explicitly checks for overflow using
1663 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1664 // sadd_with_overflow intrinsic.
1665 // FIXME: This could probably be generalized to handle other overflow-safe
1666 // operations if we worked out the formulas to compute the appropriate
1669 // INT64 : a, b, sum = a + b
1670 // if sum < INT32_MIN || sum > INT_MAX then
1677 // I = icmp ugt (add (add A B) CI2) CI
1678 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1679 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)),
1680 m_ConstantInt(CI2)))) {
1681 const IntegerType *WideType = cast<IntegerType>(CI->getType());
1682 unsigned WideWidth = WideType->getBitWidth();
1683 unsigned NarrowWidth = WideWidth / 2;
1684 const IntegerType *NarrowType =
1685 IntegerType::get(CI->getContext(), NarrowWidth);
1687 // NarrowAllOnes and NarrowSignBit are the magic constants used to
1688 // perform an overflow check in the wider type: 0x00..00FF..FF and
1689 // 0x00..0010..00 respectively, where the highest set bit in each is
1690 // what would be the sign bit in the narrower type.
1691 ConstantInt *NarrowAllOnes = cast<ConstantInt>(ConstantInt::get(WideType,
1692 APInt::getAllOnesValue(NarrowWidth).zext(WideWidth)));
1693 APInt SignBit(WideWidth, 0);
1694 SignBit.setBit(NarrowWidth-1);
1695 ConstantInt *NarrowSignBit =
1696 cast<ConstantInt>(ConstantInt::get(WideType, SignBit));
1698 if (CI == NarrowAllOnes && CI2 == NarrowSignBit) {
1699 Module *M = I.getParent()->getParent()->getParent();
1701 const Type *IntrinsicType = NarrowType;
1702 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1705 BasicBlock *InitialBlock = Builder->GetInsertBlock();
1706 BasicBlock::iterator InitialInsert = Builder->GetInsertPoint();
1708 // If the pattern matches, truncate the inputs to the narrower type and
1709 // use the sadd_with_overflow intrinsic to efficiently compute both the
1710 // result and the overflow bit.
1711 Instruction *OrigAdd =
1712 cast<Instruction>(cast<Instruction>(I.getOperand(0))->getOperand(0));
1713 Builder->SetInsertPoint(OrigAdd->getParent(),
1714 BasicBlock::iterator(OrigAdd));
1716 Value *TruncA = Builder->CreateTrunc(A, NarrowType, A->getName());
1717 Value *TruncB = Builder->CreateTrunc(B, NarrowType, B->getName());
1718 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB);
1719 Value *Add = Builder->CreateExtractValue(Call, 0);
1720 Value *ZExt = Builder->CreateZExt(Add, WideType);
1722 // The inner add was the result of the narrow add, zero extended to the
1723 // wider type. Replace it with the result computed by the intrinsic.
1724 OrigAdd->replaceAllUsesWith(ZExt);
1726 Builder->SetInsertPoint(InitialBlock, InitialInsert);
1728 return ExtractValueInst::Create(Call, 1);
1733 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1734 if (I.isEquality() && CI->isZero() &&
1735 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1736 // (icmp cond A B) if cond is equality
1737 return new ICmpInst(I.getPredicate(), A, B);
1740 // If we have an icmp le or icmp ge instruction, turn it into the
1741 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1742 // them being folded in the code below. The SimplifyICmpInst code has
1743 // already handled the edge cases for us, so we just assert on them.
1744 switch (I.getPredicate()) {
1746 case ICmpInst::ICMP_ULE:
1747 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1748 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1749 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1750 case ICmpInst::ICMP_SLE:
1751 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1752 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1753 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1754 case ICmpInst::ICMP_UGE:
1755 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1756 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1757 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1758 case ICmpInst::ICMP_SGE:
1759 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1760 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1761 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1764 // If this comparison is a normal comparison, it demands all
1765 // bits, if it is a sign bit comparison, it only demands the sign bit.
1767 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1770 // See if we can fold the comparison based on range information we can get
1771 // by checking whether bits are known to be zero or one in the input.
1772 if (BitWidth != 0) {
1773 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1774 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1776 if (SimplifyDemandedBits(I.getOperandUse(0),
1777 isSignBit ? APInt::getSignBit(BitWidth)
1778 : APInt::getAllOnesValue(BitWidth),
1779 Op0KnownZero, Op0KnownOne, 0))
1781 if (SimplifyDemandedBits(I.getOperandUse(1),
1782 APInt::getAllOnesValue(BitWidth),
1783 Op1KnownZero, Op1KnownOne, 0))
1786 // Given the known and unknown bits, compute a range that the LHS could be
1787 // in. Compute the Min, Max and RHS values based on the known bits. For the
1788 // EQ and NE we use unsigned values.
1789 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1790 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1792 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1794 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1797 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1799 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1803 // If Min and Max are known to be the same, then SimplifyDemandedBits
1804 // figured out that the LHS is a constant. Just constant fold this now so
1805 // that code below can assume that Min != Max.
1806 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1807 return new ICmpInst(I.getPredicate(),
1808 ConstantInt::get(I.getContext(), Op0Min), Op1);
1809 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1810 return new ICmpInst(I.getPredicate(), Op0,
1811 ConstantInt::get(I.getContext(), Op1Min));
1813 // Based on the range information we know about the LHS, see if we can
1814 // simplify this comparison. For example, (x&4) < 8 is always true.
1815 switch (I.getPredicate()) {
1816 default: llvm_unreachable("Unknown icmp opcode!");
1817 case ICmpInst::ICMP_EQ: {
1818 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1819 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1821 // If all bits are known zero except for one, then we know at most one
1822 // bit is set. If the comparison is against zero, then this is a check
1823 // to see if *that* bit is set.
1824 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1825 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1826 // If the LHS is an AND with the same constant, look through it.
1828 ConstantInt *LHSC = 0;
1829 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1830 LHSC->getValue() != Op0KnownZeroInverted)
1833 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1834 // then turn "((1 << x)&8) == 0" into "x != 3".
1836 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1837 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1838 return new ICmpInst(ICmpInst::ICMP_NE, X,
1839 ConstantInt::get(X->getType(), CmpVal));
1842 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1843 // then turn "((8 >>u x)&1) == 0" into "x != 3".
1844 ConstantInt *CI = 0;
1845 if (Op0KnownZeroInverted == 1 &&
1846 match(LHS, m_LShr(m_ConstantInt(CI), m_Value(X))) &&
1847 CI->getValue().isPowerOf2()) {
1848 unsigned CmpVal = CI->getValue().countTrailingZeros();
1849 return new ICmpInst(ICmpInst::ICMP_NE, X,
1850 ConstantInt::get(X->getType(), CmpVal));
1856 case ICmpInst::ICMP_NE: {
1857 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1858 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1860 // If all bits are known zero except for one, then we know at most one
1861 // bit is set. If the comparison is against zero, then this is a check
1862 // to see if *that* bit is set.
1863 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1864 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1865 // If the LHS is an AND with the same constant, look through it.
1867 ConstantInt *LHSC = 0;
1868 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1869 LHSC->getValue() != Op0KnownZeroInverted)
1872 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1873 // then turn "((1 << x)&8) != 0" into "x == 3".
1875 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1876 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1877 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1878 ConstantInt::get(X->getType(), CmpVal));
1881 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1882 // then turn "((8 >>u x)&1) != 0" into "x == 3".
1883 ConstantInt *CI = 0;
1884 if (Op0KnownZeroInverted == 1 &&
1885 match(LHS, m_LShr(m_ConstantInt(CI), m_Value(X))) &&
1886 CI->getValue().isPowerOf2()) {
1887 unsigned CmpVal = CI->getValue().countTrailingZeros();
1888 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1889 ConstantInt::get(X->getType(), CmpVal));
1895 case ICmpInst::ICMP_ULT:
1896 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
1897 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1898 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
1899 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1900 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
1901 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1902 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1903 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
1904 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1905 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1907 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
1908 if (CI->isMinValue(true))
1909 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1910 Constant::getAllOnesValue(Op0->getType()));
1913 case ICmpInst::ICMP_UGT:
1914 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
1915 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1916 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
1917 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1919 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
1920 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1921 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1922 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
1923 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1924 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1926 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
1927 if (CI->isMaxValue(true))
1928 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1929 Constant::getNullValue(Op0->getType()));
1932 case ICmpInst::ICMP_SLT:
1933 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
1934 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1935 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
1936 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1937 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
1938 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1939 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1940 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
1941 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1942 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1945 case ICmpInst::ICMP_SGT:
1946 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
1947 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1948 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
1949 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1951 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
1952 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1953 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1954 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
1955 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1956 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1959 case ICmpInst::ICMP_SGE:
1960 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
1961 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
1962 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1963 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
1964 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1966 case ICmpInst::ICMP_SLE:
1967 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
1968 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
1969 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1970 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
1971 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1973 case ICmpInst::ICMP_UGE:
1974 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
1975 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
1976 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1977 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
1978 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1980 case ICmpInst::ICMP_ULE:
1981 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
1982 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
1983 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1984 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
1985 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1989 // Turn a signed comparison into an unsigned one if both operands
1990 // are known to have the same sign.
1992 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
1993 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
1994 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
1997 // Test if the ICmpInst instruction is used exclusively by a select as
1998 // part of a minimum or maximum operation. If so, refrain from doing
1999 // any other folding. This helps out other analyses which understand
2000 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2001 // and CodeGen. And in this case, at least one of the comparison
2002 // operands has at least one user besides the compare (the select),
2003 // which would often largely negate the benefit of folding anyway.
2005 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2006 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2007 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2010 // See if we are doing a comparison between a constant and an instruction that
2011 // can be folded into the comparison.
2012 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2013 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2014 // instruction, see if that instruction also has constants so that the
2015 // instruction can be folded into the icmp
2016 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2017 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2021 // Handle icmp with constant (but not simple integer constant) RHS
2022 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2023 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2024 switch (LHSI->getOpcode()) {
2025 case Instruction::GetElementPtr:
2026 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2027 if (RHSC->isNullValue() &&
2028 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2029 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2030 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2032 case Instruction::PHI:
2033 // Only fold icmp into the PHI if the phi and icmp are in the same
2034 // block. If in the same block, we're encouraging jump threading. If
2035 // not, we are just pessimizing the code by making an i1 phi.
2036 if (LHSI->getParent() == I.getParent())
2037 if (Instruction *NV = FoldOpIntoPhi(I, true))
2040 case Instruction::Select: {
2041 // If either operand of the select is a constant, we can fold the
2042 // comparison into the select arms, which will cause one to be
2043 // constant folded and the select turned into a bitwise or.
2044 Value *Op1 = 0, *Op2 = 0;
2045 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2046 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2047 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2048 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2050 // We only want to perform this transformation if it will not lead to
2051 // additional code. This is true if either both sides of the select
2052 // fold to a constant (in which case the icmp is replaced with a select
2053 // which will usually simplify) or this is the only user of the
2054 // select (in which case we are trading a select+icmp for a simpler
2056 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2058 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2061 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2063 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2067 case Instruction::IntToPtr:
2068 // icmp pred inttoptr(X), null -> icmp pred X, 0
2069 if (RHSC->isNullValue() && TD &&
2070 TD->getIntPtrType(RHSC->getContext()) ==
2071 LHSI->getOperand(0)->getType())
2072 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2073 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2076 case Instruction::Load:
2077 // Try to optimize things like "A[i] > 4" to index computations.
2078 if (GetElementPtrInst *GEP =
2079 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2080 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2081 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2082 !cast<LoadInst>(LHSI)->isVolatile())
2083 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2090 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2091 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2092 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2094 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2095 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2096 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2099 // Test to see if the operands of the icmp are casted versions of other
2100 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2102 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2103 if (Op0->getType()->isPointerTy() &&
2104 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2105 // We keep moving the cast from the left operand over to the right
2106 // operand, where it can often be eliminated completely.
2107 Op0 = CI->getOperand(0);
2109 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2110 // so eliminate it as well.
2111 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2112 Op1 = CI2->getOperand(0);
2114 // If Op1 is a constant, we can fold the cast into the constant.
2115 if (Op0->getType() != Op1->getType()) {
2116 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2117 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2119 // Otherwise, cast the RHS right before the icmp
2120 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2123 return new ICmpInst(I.getPredicate(), Op0, Op1);
2127 if (isa<CastInst>(Op0)) {
2128 // Handle the special case of: icmp (cast bool to X), <cst>
2129 // This comes up when you have code like
2132 // For generality, we handle any zero-extension of any operand comparison
2133 // with a constant or another cast from the same type.
2134 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2135 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2139 // See if it's the same type of instruction on the left and right.
2140 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2141 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2142 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
2143 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
2144 switch (Op0I->getOpcode()) {
2146 case Instruction::Add:
2147 case Instruction::Sub:
2148 case Instruction::Xor:
2149 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2150 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
2151 Op1I->getOperand(0));
2152 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2153 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2154 if (CI->getValue().isSignBit()) {
2155 ICmpInst::Predicate Pred = I.isSigned()
2156 ? I.getUnsignedPredicate()
2157 : I.getSignedPredicate();
2158 return new ICmpInst(Pred, Op0I->getOperand(0),
2159 Op1I->getOperand(0));
2162 if (CI->getValue().isMaxSignedValue()) {
2163 ICmpInst::Predicate Pred = I.isSigned()
2164 ? I.getUnsignedPredicate()
2165 : I.getSignedPredicate();
2166 Pred = I.getSwappedPredicate(Pred);
2167 return new ICmpInst(Pred, Op0I->getOperand(0),
2168 Op1I->getOperand(0));
2172 case Instruction::Mul:
2173 if (!I.isEquality())
2176 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2177 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2178 // Mask = -1 >> count-trailing-zeros(Cst).
2179 if (!CI->isZero() && !CI->isOne()) {
2180 const APInt &AP = CI->getValue();
2181 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2182 APInt::getLowBitsSet(AP.getBitWidth(),
2184 AP.countTrailingZeros()));
2185 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
2186 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
2187 return new ICmpInst(I.getPredicate(), And1, And2);
2196 // ~x < ~y --> y < x
2198 if (match(Op0, m_Not(m_Value(A))) &&
2199 match(Op1, m_Not(m_Value(B))))
2200 return new ICmpInst(I.getPredicate(), B, A);
2203 if (I.isEquality()) {
2204 Value *A, *B, *C, *D;
2206 // -x == -y --> x == y
2207 if (match(Op0, m_Neg(m_Value(A))) &&
2208 match(Op1, m_Neg(m_Value(B))))
2209 return new ICmpInst(I.getPredicate(), A, B);
2211 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2212 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2213 Value *OtherVal = A == Op1 ? B : A;
2214 return new ICmpInst(I.getPredicate(), OtherVal,
2215 Constant::getNullValue(A->getType()));
2218 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2219 // A^c1 == C^c2 --> A == C^(c1^c2)
2220 ConstantInt *C1, *C2;
2221 if (match(B, m_ConstantInt(C1)) &&
2222 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2223 Constant *NC = ConstantInt::get(I.getContext(),
2224 C1->getValue() ^ C2->getValue());
2225 Value *Xor = Builder->CreateXor(C, NC, "tmp");
2226 return new ICmpInst(I.getPredicate(), A, Xor);
2229 // A^B == A^D -> B == D
2230 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2231 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2232 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2233 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2237 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2238 (A == Op0 || B == Op0)) {
2239 // A == (A^B) -> B == 0
2240 Value *OtherVal = A == Op0 ? B : A;
2241 return new ICmpInst(I.getPredicate(), OtherVal,
2242 Constant::getNullValue(A->getType()));
2245 // (A-B) == A -> B == 0
2246 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
2247 return new ICmpInst(I.getPredicate(), B,
2248 Constant::getNullValue(B->getType()));
2250 // A == (A-B) -> B == 0
2251 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
2252 return new ICmpInst(I.getPredicate(), B,
2253 Constant::getNullValue(B->getType()));
2255 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2256 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2257 match(Op0, m_And(m_Value(A), m_Value(B))) &&
2258 match(Op1, m_And(m_Value(C), m_Value(D)))) {
2259 Value *X = 0, *Y = 0, *Z = 0;
2262 X = B; Y = D; Z = A;
2263 } else if (A == D) {
2264 X = B; Y = C; Z = A;
2265 } else if (B == C) {
2266 X = A; Y = D; Z = B;
2267 } else if (B == D) {
2268 X = A; Y = C; Z = B;
2271 if (X) { // Build (X^Y) & Z
2272 Op1 = Builder->CreateXor(X, Y, "tmp");
2273 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
2274 I.setOperand(0, Op1);
2275 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2282 Value *X; ConstantInt *Cst;
2284 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2285 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2288 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2289 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2291 return Changed ? &I : 0;
2299 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2301 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2304 if (!isa<ConstantFP>(RHSC)) return 0;
2305 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2307 // Get the width of the mantissa. We don't want to hack on conversions that
2308 // might lose information from the integer, e.g. "i64 -> float"
2309 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2310 if (MantissaWidth == -1) return 0; // Unknown.
2312 // Check to see that the input is converted from an integer type that is small
2313 // enough that preserves all bits. TODO: check here for "known" sign bits.
2314 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2315 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2317 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2318 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2322 // If the conversion would lose info, don't hack on this.
2323 if ((int)InputSize > MantissaWidth)
2326 // Otherwise, we can potentially simplify the comparison. We know that it
2327 // will always come through as an integer value and we know the constant is
2328 // not a NAN (it would have been previously simplified).
2329 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2331 ICmpInst::Predicate Pred;
2332 switch (I.getPredicate()) {
2333 default: llvm_unreachable("Unexpected predicate!");
2334 case FCmpInst::FCMP_UEQ:
2335 case FCmpInst::FCMP_OEQ:
2336 Pred = ICmpInst::ICMP_EQ;
2338 case FCmpInst::FCMP_UGT:
2339 case FCmpInst::FCMP_OGT:
2340 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2342 case FCmpInst::FCMP_UGE:
2343 case FCmpInst::FCMP_OGE:
2344 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2346 case FCmpInst::FCMP_ULT:
2347 case FCmpInst::FCMP_OLT:
2348 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2350 case FCmpInst::FCMP_ULE:
2351 case FCmpInst::FCMP_OLE:
2352 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2354 case FCmpInst::FCMP_UNE:
2355 case FCmpInst::FCMP_ONE:
2356 Pred = ICmpInst::ICMP_NE;
2358 case FCmpInst::FCMP_ORD:
2359 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2360 case FCmpInst::FCMP_UNO:
2361 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2364 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2366 // Now we know that the APFloat is a normal number, zero or inf.
2368 // See if the FP constant is too large for the integer. For example,
2369 // comparing an i8 to 300.0.
2370 unsigned IntWidth = IntTy->getScalarSizeInBits();
2373 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2374 // and large values.
2375 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2376 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2377 APFloat::rmNearestTiesToEven);
2378 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2379 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2380 Pred == ICmpInst::ICMP_SLE)
2381 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2382 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2385 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2386 // +INF and large values.
2387 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2388 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2389 APFloat::rmNearestTiesToEven);
2390 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2391 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2392 Pred == ICmpInst::ICMP_ULE)
2393 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2394 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2399 // See if the RHS value is < SignedMin.
2400 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2401 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2402 APFloat::rmNearestTiesToEven);
2403 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2404 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2405 Pred == ICmpInst::ICMP_SGE)
2406 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2407 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2411 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2412 // [0, UMAX], but it may still be fractional. See if it is fractional by
2413 // casting the FP value to the integer value and back, checking for equality.
2414 // Don't do this for zero, because -0.0 is not fractional.
2415 Constant *RHSInt = LHSUnsigned
2416 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2417 : ConstantExpr::getFPToSI(RHSC, IntTy);
2418 if (!RHS.isZero()) {
2419 bool Equal = LHSUnsigned
2420 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2421 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2423 // If we had a comparison against a fractional value, we have to adjust
2424 // the compare predicate and sometimes the value. RHSC is rounded towards
2425 // zero at this point.
2427 default: llvm_unreachable("Unexpected integer comparison!");
2428 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2429 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2430 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2431 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2432 case ICmpInst::ICMP_ULE:
2433 // (float)int <= 4.4 --> int <= 4
2434 // (float)int <= -4.4 --> false
2435 if (RHS.isNegative())
2436 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2438 case ICmpInst::ICMP_SLE:
2439 // (float)int <= 4.4 --> int <= 4
2440 // (float)int <= -4.4 --> int < -4
2441 if (RHS.isNegative())
2442 Pred = ICmpInst::ICMP_SLT;
2444 case ICmpInst::ICMP_ULT:
2445 // (float)int < -4.4 --> false
2446 // (float)int < 4.4 --> int <= 4
2447 if (RHS.isNegative())
2448 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2449 Pred = ICmpInst::ICMP_ULE;
2451 case ICmpInst::ICMP_SLT:
2452 // (float)int < -4.4 --> int < -4
2453 // (float)int < 4.4 --> int <= 4
2454 if (!RHS.isNegative())
2455 Pred = ICmpInst::ICMP_SLE;
2457 case ICmpInst::ICMP_UGT:
2458 // (float)int > 4.4 --> int > 4
2459 // (float)int > -4.4 --> true
2460 if (RHS.isNegative())
2461 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2463 case ICmpInst::ICMP_SGT:
2464 // (float)int > 4.4 --> int > 4
2465 // (float)int > -4.4 --> int >= -4
2466 if (RHS.isNegative())
2467 Pred = ICmpInst::ICMP_SGE;
2469 case ICmpInst::ICMP_UGE:
2470 // (float)int >= -4.4 --> true
2471 // (float)int >= 4.4 --> int > 4
2472 if (!RHS.isNegative())
2473 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2474 Pred = ICmpInst::ICMP_UGT;
2476 case ICmpInst::ICMP_SGE:
2477 // (float)int >= -4.4 --> int >= -4
2478 // (float)int >= 4.4 --> int > 4
2479 if (!RHS.isNegative())
2480 Pred = ICmpInst::ICMP_SGT;
2486 // Lower this FP comparison into an appropriate integer version of the
2488 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2491 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2492 bool Changed = false;
2494 /// Orders the operands of the compare so that they are listed from most
2495 /// complex to least complex. This puts constants before unary operators,
2496 /// before binary operators.
2497 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2502 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2504 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2505 return ReplaceInstUsesWith(I, V);
2507 // Simplify 'fcmp pred X, X'
2509 switch (I.getPredicate()) {
2510 default: llvm_unreachable("Unknown predicate!");
2511 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2512 case FCmpInst::FCMP_ULT: // True if unordered or less than
2513 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2514 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2515 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2516 I.setPredicate(FCmpInst::FCMP_UNO);
2517 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2520 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2521 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2522 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2523 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2524 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2525 I.setPredicate(FCmpInst::FCMP_ORD);
2526 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2531 // Handle fcmp with constant RHS
2532 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2533 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2534 switch (LHSI->getOpcode()) {
2535 case Instruction::PHI:
2536 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2537 // block. If in the same block, we're encouraging jump threading. If
2538 // not, we are just pessimizing the code by making an i1 phi.
2539 if (LHSI->getParent() == I.getParent())
2540 if (Instruction *NV = FoldOpIntoPhi(I, true))
2543 case Instruction::SIToFP:
2544 case Instruction::UIToFP:
2545 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2548 case Instruction::Select: {
2549 // If either operand of the select is a constant, we can fold the
2550 // comparison into the select arms, which will cause one to be
2551 // constant folded and the select turned into a bitwise or.
2552 Value *Op1 = 0, *Op2 = 0;
2553 if (LHSI->hasOneUse()) {
2554 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2555 // Fold the known value into the constant operand.
2556 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2557 // Insert a new FCmp of the other select operand.
2558 Op2 = Builder->CreateFCmp(I.getPredicate(),
2559 LHSI->getOperand(2), RHSC, I.getName());
2560 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2561 // Fold the known value into the constant operand.
2562 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2563 // Insert a new FCmp of the other select operand.
2564 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2570 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2573 case Instruction::Load:
2574 if (GetElementPtrInst *GEP =
2575 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2576 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2577 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2578 !cast<LoadInst>(LHSI)->isVolatile())
2579 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2586 return Changed ? &I : 0;