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 static ConstantInt *getOne(Constant *C) {
26 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
29 /// AddOne - Add one to a ConstantInt
30 static Constant *AddOne(Constant *C) {
31 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
33 /// SubOne - Subtract one from a ConstantInt
34 static Constant *SubOne(Constant *C) {
35 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
38 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
39 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
42 static bool HasAddOverflow(ConstantInt *Result,
43 ConstantInt *In1, ConstantInt *In2,
46 if (In2->getValue().isNegative())
47 return Result->getValue().sgt(In1->getValue());
49 return Result->getValue().slt(In1->getValue());
51 return Result->getValue().ult(In1->getValue());
54 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
55 /// overflowed for this type.
56 static bool AddWithOverflow(Constant *&Result, Constant *In1,
57 Constant *In2, bool IsSigned = false) {
58 Result = ConstantExpr::getAdd(In1, In2);
60 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
61 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
62 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
63 if (HasAddOverflow(ExtractElement(Result, Idx),
64 ExtractElement(In1, Idx),
65 ExtractElement(In2, Idx),
72 return HasAddOverflow(cast<ConstantInt>(Result),
73 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
77 static bool HasSubOverflow(ConstantInt *Result,
78 ConstantInt *In1, ConstantInt *In2,
81 if (In2->getValue().isNegative())
82 return Result->getValue().slt(In1->getValue());
84 return Result->getValue().sgt(In1->getValue());
86 return Result->getValue().ugt(In1->getValue());
89 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
90 /// overflowed for this type.
91 static bool SubWithOverflow(Constant *&Result, Constant *In1,
92 Constant *In2, bool IsSigned = false) {
93 Result = ConstantExpr::getSub(In1, In2);
95 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
96 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
97 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
98 if (HasSubOverflow(ExtractElement(Result, Idx),
99 ExtractElement(In1, Idx),
100 ExtractElement(In2, Idx),
107 return HasSubOverflow(cast<ConstantInt>(Result),
108 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
112 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
113 /// comparison only checks the sign bit. If it only checks the sign bit, set
114 /// TrueIfSigned if the result of the comparison is true when the input value is
116 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
117 bool &TrueIfSigned) {
119 case ICmpInst::ICMP_SLT: // True if LHS s< 0
121 return RHS->isZero();
122 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
124 return RHS->isAllOnesValue();
125 case ICmpInst::ICMP_SGT: // True if LHS s> -1
126 TrueIfSigned = false;
127 return RHS->isAllOnesValue();
128 case ICmpInst::ICMP_UGT:
129 // True if LHS u> RHS and RHS == high-bit-mask - 1
131 return RHS->getValue() ==
132 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
133 case ICmpInst::ICMP_UGE:
134 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
136 return RHS->getValue().isSignBit();
142 // isHighOnes - Return true if the constant is of the form 1+0+.
143 // This is the same as lowones(~X).
144 static bool isHighOnes(const ConstantInt *CI) {
145 return (~CI->getValue() + 1).isPowerOf2();
148 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
149 /// set of known zero and one bits, compute the maximum and minimum values that
150 /// could have the specified known zero and known one bits, returning them in
152 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
153 const APInt& KnownOne,
154 APInt& Min, APInt& Max) {
155 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
156 KnownZero.getBitWidth() == Min.getBitWidth() &&
157 KnownZero.getBitWidth() == Max.getBitWidth() &&
158 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
159 APInt UnknownBits = ~(KnownZero|KnownOne);
161 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
162 // bit if it is unknown.
164 Max = KnownOne|UnknownBits;
166 if (UnknownBits.isNegative()) { // Sign bit is unknown
167 Min.setBit(Min.getBitWidth()-1);
168 Max.clearBit(Max.getBitWidth()-1);
172 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
173 // a set of known zero and one bits, compute the maximum and minimum values that
174 // could have the specified known zero and known one bits, returning them in
176 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
177 const APInt &KnownOne,
178 APInt &Min, APInt &Max) {
179 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
180 KnownZero.getBitWidth() == Min.getBitWidth() &&
181 KnownZero.getBitWidth() == Max.getBitWidth() &&
182 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
183 APInt UnknownBits = ~(KnownZero|KnownOne);
185 // The minimum value is when the unknown bits are all zeros.
187 // The maximum value is when the unknown bits are all ones.
188 Max = KnownOne|UnknownBits;
193 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
194 /// cmp pred (load (gep GV, ...)), cmpcst
195 /// where GV is a global variable with a constant initializer. Try to simplify
196 /// this into some simple computation that does not need the load. For example
197 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
199 /// If AndCst is non-null, then the loaded value is masked with that constant
200 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
201 Instruction *InstCombiner::
202 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
203 CmpInst &ICI, ConstantInt *AndCst) {
204 // We need TD information to know the pointer size unless this is inbounds.
205 if (!GEP->isInBounds() && TD == 0) return 0;
207 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer());
208 if (Init == 0 || Init->getNumOperands() > 1024) return 0;
210 // There are many forms of this optimization we can handle, for now, just do
211 // the simple index into a single-dimensional array.
213 // Require: GEP GV, 0, i {{, constant indices}}
214 if (GEP->getNumOperands() < 3 ||
215 !isa<ConstantInt>(GEP->getOperand(1)) ||
216 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
217 isa<Constant>(GEP->getOperand(2)))
220 // Check that indices after the variable are constants and in-range for the
221 // type they index. Collect the indices. This is typically for arrays of
223 SmallVector<unsigned, 4> LaterIndices;
225 const Type *EltTy = cast<ArrayType>(Init->getType())->getElementType();
226 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
227 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
228 if (Idx == 0) return 0; // Variable index.
230 uint64_t IdxVal = Idx->getZExtValue();
231 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
233 if (const StructType *STy = dyn_cast<StructType>(EltTy))
234 EltTy = STy->getElementType(IdxVal);
235 else if (const ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
236 if (IdxVal >= ATy->getNumElements()) return 0;
237 EltTy = ATy->getElementType();
239 return 0; // Unknown type.
242 LaterIndices.push_back(IdxVal);
245 enum { Overdefined = -3, Undefined = -2 };
247 // Variables for our state machines.
249 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
250 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
251 // and 87 is the second (and last) index. FirstTrueElement is -2 when
252 // undefined, otherwise set to the first true element. SecondTrueElement is
253 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
254 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
256 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
257 // form "i != 47 & i != 87". Same state transitions as for true elements.
258 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
260 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
261 /// define a state machine that triggers for ranges of values that the index
262 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
263 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
264 /// index in the range (inclusive). We use -2 for undefined here because we
265 /// use relative comparisons and don't want 0-1 to match -1.
266 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
268 // MagicBitvector - This is a magic bitvector where we set a bit if the
269 // comparison is true for element 'i'. If there are 64 elements or less in
270 // the array, this will fully represent all the comparison results.
271 uint64_t MagicBitvector = 0;
274 // Scan the array and see if one of our patterns matches.
275 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
276 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) {
277 Constant *Elt = Init->getOperand(i);
279 // If this is indexing an array of structures, get the structure element.
280 if (!LaterIndices.empty())
281 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices.data(),
282 LaterIndices.size());
284 // If the element is masked, handle it.
285 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
287 // Find out if the comparison would be true or false for the i'th element.
288 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
290 // If the result is undef for this element, ignore it.
291 if (isa<UndefValue>(C)) {
292 // Extend range state machines to cover this element in case there is an
293 // undef in the middle of the range.
294 if (TrueRangeEnd == (int)i-1)
296 if (FalseRangeEnd == (int)i-1)
301 // If we can't compute the result for any of the elements, we have to give
302 // up evaluating the entire conditional.
303 if (!isa<ConstantInt>(C)) return 0;
305 // Otherwise, we know if the comparison is true or false for this element,
306 // update our state machines.
307 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
309 // State machine for single/double/range index comparison.
311 // Update the TrueElement state machine.
312 if (FirstTrueElement == Undefined)
313 FirstTrueElement = TrueRangeEnd = i; // First true element.
315 // Update double-compare state machine.
316 if (SecondTrueElement == Undefined)
317 SecondTrueElement = i;
319 SecondTrueElement = Overdefined;
321 // Update range state machine.
322 if (TrueRangeEnd == (int)i-1)
325 TrueRangeEnd = Overdefined;
328 // Update the FalseElement state machine.
329 if (FirstFalseElement == Undefined)
330 FirstFalseElement = FalseRangeEnd = i; // First false element.
332 // Update double-compare state machine.
333 if (SecondFalseElement == Undefined)
334 SecondFalseElement = i;
336 SecondFalseElement = Overdefined;
338 // Update range state machine.
339 if (FalseRangeEnd == (int)i-1)
342 FalseRangeEnd = Overdefined;
347 // If this element is in range, update our magic bitvector.
348 if (i < 64 && IsTrueForElt)
349 MagicBitvector |= 1ULL << i;
351 // If all of our states become overdefined, bail out early. Since the
352 // predicate is expensive, only check it every 8 elements. This is only
353 // really useful for really huge arrays.
354 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
355 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
356 FalseRangeEnd == Overdefined)
360 // Now that we've scanned the entire array, emit our new comparison(s). We
361 // order the state machines in complexity of the generated code.
362 Value *Idx = GEP->getOperand(2);
364 // If the index is larger than the pointer size of the target, truncate the
365 // index down like the GEP would do implicitly. We don't have to do this for
366 // an inbounds GEP because the index can't be out of range.
367 if (!GEP->isInBounds() &&
368 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
369 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
371 // If the comparison is only true for one or two elements, emit direct
373 if (SecondTrueElement != Overdefined) {
374 // None true -> false.
375 if (FirstTrueElement == Undefined)
376 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
378 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
380 // True for one element -> 'i == 47'.
381 if (SecondTrueElement == Undefined)
382 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
384 // True for two elements -> 'i == 47 | i == 72'.
385 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
386 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
387 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
388 return BinaryOperator::CreateOr(C1, C2);
391 // If the comparison is only false for one or two elements, emit direct
393 if (SecondFalseElement != Overdefined) {
394 // None false -> true.
395 if (FirstFalseElement == Undefined)
396 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
398 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
400 // False for one element -> 'i != 47'.
401 if (SecondFalseElement == Undefined)
402 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
404 // False for two elements -> 'i != 47 & i != 72'.
405 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
406 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
407 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
408 return BinaryOperator::CreateAnd(C1, C2);
411 // If the comparison can be replaced with a range comparison for the elements
412 // where it is true, emit the range check.
413 if (TrueRangeEnd != Overdefined) {
414 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
416 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
417 if (FirstTrueElement) {
418 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
419 Idx = Builder->CreateAdd(Idx, Offs);
422 Value *End = ConstantInt::get(Idx->getType(),
423 TrueRangeEnd-FirstTrueElement+1);
424 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
427 // False range check.
428 if (FalseRangeEnd != Overdefined) {
429 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
430 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
431 if (FirstFalseElement) {
432 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
433 Idx = Builder->CreateAdd(Idx, Offs);
436 Value *End = ConstantInt::get(Idx->getType(),
437 FalseRangeEnd-FirstFalseElement);
438 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
442 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
443 // of this load, replace it with computation that does:
444 // ((magic_cst >> i) & 1) != 0
445 if (Init->getNumOperands() <= 32 ||
446 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) {
448 if (Init->getNumOperands() <= 32)
449 Ty = Type::getInt32Ty(Init->getContext());
451 Ty = Type::getInt64Ty(Init->getContext());
452 Value *V = Builder->CreateIntCast(Idx, Ty, false);
453 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
454 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
455 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
462 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
463 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
464 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
465 /// be complex, and scales are involved. The above expression would also be
466 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
467 /// This later form is less amenable to optimization though, and we are allowed
468 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
470 /// If we can't emit an optimized form for this expression, this returns null.
472 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
474 TargetData &TD = *IC.getTargetData();
475 gep_type_iterator GTI = gep_type_begin(GEP);
477 // Check to see if this gep only has a single variable index. If so, and if
478 // any constant indices are a multiple of its scale, then we can compute this
479 // in terms of the scale of the variable index. For example, if the GEP
480 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
481 // because the expression will cross zero at the same point.
482 unsigned i, e = GEP->getNumOperands();
484 for (i = 1; i != e; ++i, ++GTI) {
485 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
486 // Compute the aggregate offset of constant indices.
487 if (CI->isZero()) continue;
489 // Handle a struct index, which adds its field offset to the pointer.
490 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
491 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
493 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
494 Offset += Size*CI->getSExtValue();
497 // Found our variable index.
502 // If there are no variable indices, we must have a constant offset, just
503 // evaluate it the general way.
504 if (i == e) return 0;
506 Value *VariableIdx = GEP->getOperand(i);
507 // Determine the scale factor of the variable element. For example, this is
508 // 4 if the variable index is into an array of i32.
509 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
511 // Verify that there are no other variable indices. If so, emit the hard way.
512 for (++i, ++GTI; i != e; ++i, ++GTI) {
513 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
516 // Compute the aggregate offset of constant indices.
517 if (CI->isZero()) continue;
519 // Handle a struct index, which adds its field offset to the pointer.
520 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
521 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
523 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
524 Offset += Size*CI->getSExtValue();
528 // Okay, we know we have a single variable index, which must be a
529 // pointer/array/vector index. If there is no offset, life is simple, return
531 unsigned IntPtrWidth = TD.getPointerSizeInBits();
533 // Cast to intptrty in case a truncation occurs. If an extension is needed,
534 // we don't need to bother extending: the extension won't affect where the
535 // computation crosses zero.
536 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
537 VariableIdx = new TruncInst(VariableIdx,
538 TD.getIntPtrType(VariableIdx->getContext()),
539 VariableIdx->getName(), &I);
543 // Otherwise, there is an index. The computation we will do will be modulo
544 // the pointer size, so get it.
545 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
547 Offset &= PtrSizeMask;
548 VariableScale &= PtrSizeMask;
550 // To do this transformation, any constant index must be a multiple of the
551 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
552 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
553 // multiple of the variable scale.
554 int64_t NewOffs = Offset / (int64_t)VariableScale;
555 if (Offset != NewOffs*(int64_t)VariableScale)
558 // Okay, we can do this evaluation. Start by converting the index to intptr.
559 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
560 if (VariableIdx->getType() != IntPtrTy)
561 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
563 VariableIdx->getName(), &I);
564 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
565 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
568 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
569 /// else. At this point we know that the GEP is on the LHS of the comparison.
570 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
571 ICmpInst::Predicate Cond,
573 // Look through bitcasts.
574 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
575 RHS = BCI->getOperand(0);
577 Value *PtrBase = GEPLHS->getOperand(0);
578 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
579 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
580 // This transformation (ignoring the base and scales) is valid because we
581 // know pointers can't overflow since the gep is inbounds. See if we can
582 // output an optimized form.
583 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
585 // If not, synthesize the offset the hard way.
587 Offset = EmitGEPOffset(GEPLHS);
588 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
589 Constant::getNullValue(Offset->getType()));
590 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
591 // If the base pointers are different, but the indices are the same, just
592 // compare the base pointer.
593 if (PtrBase != GEPRHS->getOperand(0)) {
594 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
595 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
596 GEPRHS->getOperand(0)->getType();
598 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
599 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
600 IndicesTheSame = false;
604 // If all indices are the same, just compare the base pointers.
606 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
607 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
609 // Otherwise, the base pointers are different and the indices are
610 // different, bail out.
614 // If one of the GEPs has all zero indices, recurse.
615 bool AllZeros = true;
616 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
617 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
618 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
623 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
624 ICmpInst::getSwappedPredicate(Cond), I);
626 // If the other GEP has all zero indices, recurse.
628 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
629 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
630 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
635 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
637 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
638 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
639 // If the GEPs only differ by one index, compare it.
640 unsigned NumDifferences = 0; // Keep track of # differences.
641 unsigned DiffOperand = 0; // The operand that differs.
642 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
643 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
644 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
645 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
646 // Irreconcilable differences.
650 if (NumDifferences++) break;
655 if (NumDifferences == 0) // SAME GEP?
656 return ReplaceInstUsesWith(I, // No comparison is needed here.
657 ConstantInt::get(Type::getInt1Ty(I.getContext()),
658 ICmpInst::isTrueWhenEqual(Cond)));
660 else if (NumDifferences == 1 && GEPsInBounds) {
661 Value *LHSV = GEPLHS->getOperand(DiffOperand);
662 Value *RHSV = GEPRHS->getOperand(DiffOperand);
663 // Make sure we do a signed comparison here.
664 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
668 // Only lower this if the icmp is the only user of the GEP or if we expect
669 // the result to fold to a constant!
672 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
673 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
674 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
675 Value *L = EmitGEPOffset(GEPLHS);
676 Value *R = EmitGEPOffset(GEPRHS);
677 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
683 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
684 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
685 Value *X, ConstantInt *CI,
686 ICmpInst::Predicate Pred,
688 // If we have X+0, exit early (simplifying logic below) and let it get folded
689 // elsewhere. icmp X+0, X -> icmp X, X
691 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
692 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
695 // (X+4) == X -> false.
696 if (Pred == ICmpInst::ICMP_EQ)
697 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
699 // (X+4) != X -> true.
700 if (Pred == ICmpInst::ICMP_NE)
701 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
703 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
704 // so the values can never be equal. Similarly for all other "or equals"
707 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
708 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
709 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
710 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
712 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
713 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
716 // (X+1) >u X --> X <u (0-1) --> X != 255
717 // (X+2) >u X --> X <u (0-2) --> X <u 254
718 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
719 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
720 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
722 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
723 ConstantInt *SMax = ConstantInt::get(X->getContext(),
724 APInt::getSignedMaxValue(BitWidth));
726 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
727 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
728 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
729 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
730 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
731 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
732 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
733 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
735 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
736 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
737 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
738 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
739 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
740 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
742 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
743 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
744 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
747 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
748 /// and CmpRHS are both known to be integer constants.
749 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
750 ConstantInt *DivRHS) {
751 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
752 const APInt &CmpRHSV = CmpRHS->getValue();
754 // FIXME: If the operand types don't match the type of the divide
755 // then don't attempt this transform. The code below doesn't have the
756 // logic to deal with a signed divide and an unsigned compare (and
757 // vice versa). This is because (x /s C1) <s C2 produces different
758 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
759 // (x /u C1) <u C2. Simply casting the operands and result won't
760 // work. :( The if statement below tests that condition and bails
762 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
763 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
765 if (DivRHS->isZero())
766 return 0; // The ProdOV computation fails on divide by zero.
767 if (DivIsSigned && DivRHS->isAllOnesValue())
768 return 0; // The overflow computation also screws up here
769 if (DivRHS->isOne()) {
770 // This eliminates some funny cases with INT_MIN.
771 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
775 // Compute Prod = CI * DivRHS. We are essentially solving an equation
776 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
777 // C2 (CI). By solving for X we can turn this into a range check
778 // instead of computing a divide.
779 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
781 // Determine if the product overflows by seeing if the product is
782 // not equal to the divide. Make sure we do the same kind of divide
783 // as in the LHS instruction that we're folding.
784 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
785 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
787 // Get the ICmp opcode
788 ICmpInst::Predicate Pred = ICI.getPredicate();
790 /// If the division is known to be exact, then there is no remainder from the
791 /// divide, so the covered range size is unit, otherwise it is the divisor.
792 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
794 // Figure out the interval that is being checked. For example, a comparison
795 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
796 // Compute this interval based on the constants involved and the signedness of
797 // the compare/divide. This computes a half-open interval, keeping track of
798 // whether either value in the interval overflows. After analysis each
799 // overflow variable is set to 0 if it's corresponding bound variable is valid
800 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
801 int LoOverflow = 0, HiOverflow = 0;
802 Constant *LoBound = 0, *HiBound = 0;
804 if (!DivIsSigned) { // udiv
805 // e.g. X/5 op 3 --> [15, 20)
807 HiOverflow = LoOverflow = ProdOV;
809 // If this is not an exact divide, then many values in the range collapse
810 // to the same result value.
811 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
814 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
815 if (CmpRHSV == 0) { // (X / pos) op 0
816 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
817 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
819 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
820 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
821 HiOverflow = LoOverflow = ProdOV;
823 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
824 } else { // (X / pos) op neg
825 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
826 HiBound = AddOne(Prod);
827 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
829 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
830 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
833 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
835 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
836 if (CmpRHSV == 0) { // (X / neg) op 0
837 // e.g. X/-5 op 0 --> [-4, 5)
838 LoBound = AddOne(RangeSize);
839 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
840 if (HiBound == DivRHS) { // -INTMIN = INTMIN
841 HiOverflow = 1; // [INTMIN+1, overflow)
842 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
844 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
845 // e.g. X/-5 op 3 --> [-19, -14)
846 HiBound = AddOne(Prod);
847 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
849 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
850 } else { // (X / neg) op neg
851 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
852 LoOverflow = HiOverflow = ProdOV;
854 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
857 // Dividing by a negative swaps the condition. LT <-> GT
858 Pred = ICmpInst::getSwappedPredicate(Pred);
861 Value *X = DivI->getOperand(0);
863 default: llvm_unreachable("Unhandled icmp opcode!");
864 case ICmpInst::ICMP_EQ:
865 if (LoOverflow && HiOverflow)
866 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
868 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
869 ICmpInst::ICMP_UGE, X, LoBound);
871 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
872 ICmpInst::ICMP_ULT, X, HiBound);
873 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
875 case ICmpInst::ICMP_NE:
876 if (LoOverflow && HiOverflow)
877 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
879 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
880 ICmpInst::ICMP_ULT, X, LoBound);
882 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
883 ICmpInst::ICMP_UGE, X, HiBound);
884 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
885 DivIsSigned, false));
886 case ICmpInst::ICMP_ULT:
887 case ICmpInst::ICMP_SLT:
888 if (LoOverflow == +1) // Low bound is greater than input range.
889 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
890 if (LoOverflow == -1) // Low bound is less than input range.
891 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
892 return new ICmpInst(Pred, X, LoBound);
893 case ICmpInst::ICMP_UGT:
894 case ICmpInst::ICMP_SGT:
895 if (HiOverflow == +1) // High bound greater than input range.
896 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
897 if (HiOverflow == -1) // High bound less than input range.
898 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
899 if (Pred == ICmpInst::ICMP_UGT)
900 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
901 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
905 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
906 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
907 ConstantInt *ShAmt) {
908 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
910 // Check that the shift amount is in range. If not, don't perform
911 // undefined shifts. When the shift is visited it will be
913 uint32_t TypeBits = CmpRHSV.getBitWidth();
914 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
915 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
918 if (!ICI.isEquality()) {
919 // If we have an unsigned comparison and an ashr, we can't simplify this.
920 // Similarly for signed comparisons with lshr.
921 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
924 // Otherwise, all lshr and all exact ashr's are equivalent to a udiv/sdiv by
925 // a power of 2. Since we already have logic to simplify these, transform
926 // to div and then simplify the resultant comparison.
927 if (Shr->getOpcode() == Instruction::AShr &&
931 // Revisit the shift (to delete it).
935 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
938 Shr->getOpcode() == Instruction::AShr ?
939 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
940 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
942 ICI.setOperand(0, Tmp);
944 // If the builder folded the binop, just return it.
945 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
949 // Otherwise, fold this div/compare.
950 assert(TheDiv->getOpcode() == Instruction::SDiv ||
951 TheDiv->getOpcode() == Instruction::UDiv);
953 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
954 assert(Res && "This div/cst should have folded!");
959 // If we are comparing against bits always shifted out, the
960 // comparison cannot succeed.
961 APInt Comp = CmpRHSV << ShAmtVal;
962 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
963 if (Shr->getOpcode() == Instruction::LShr)
964 Comp = Comp.lshr(ShAmtVal);
966 Comp = Comp.ashr(ShAmtVal);
968 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
969 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
970 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
972 return ReplaceInstUsesWith(ICI, Cst);
975 // Otherwise, check to see if the bits shifted out are known to be zero.
976 // If so, we can compare against the unshifted value:
977 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
978 if (Shr->hasOneUse() && Shr->isExact())
979 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
981 if (Shr->hasOneUse()) {
982 // Otherwise strength reduce the shift into an and.
983 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
984 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
986 Value *And = Builder->CreateAnd(Shr->getOperand(0),
987 Mask, Shr->getName()+".mask");
988 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
994 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
996 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
999 const APInt &RHSV = RHS->getValue();
1001 switch (LHSI->getOpcode()) {
1002 case Instruction::Trunc:
1003 if (ICI.isEquality() && LHSI->hasOneUse()) {
1004 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1005 // of the high bits truncated out of x are known.
1006 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1007 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1008 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
1009 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1010 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
1012 // If all the high bits are known, we can do this xform.
1013 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1014 // Pull in the high bits from known-ones set.
1015 APInt NewRHS = RHS->getValue().zext(SrcBits);
1017 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1018 ConstantInt::get(ICI.getContext(), NewRHS));
1023 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1024 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1025 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1027 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1028 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1029 Value *CompareVal = LHSI->getOperand(0);
1031 // If the sign bit of the XorCST is not set, there is no change to
1032 // the operation, just stop using the Xor.
1033 if (!XorCST->getValue().isNegative()) {
1034 ICI.setOperand(0, CompareVal);
1039 // Was the old condition true if the operand is positive?
1040 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1042 // If so, the new one isn't.
1043 isTrueIfPositive ^= true;
1045 if (isTrueIfPositive)
1046 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1049 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1053 if (LHSI->hasOneUse()) {
1054 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1055 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1056 const APInt &SignBit = XorCST->getValue();
1057 ICmpInst::Predicate Pred = ICI.isSigned()
1058 ? ICI.getUnsignedPredicate()
1059 : ICI.getSignedPredicate();
1060 return new ICmpInst(Pred, LHSI->getOperand(0),
1061 ConstantInt::get(ICI.getContext(),
1065 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1066 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
1067 const APInt &NotSignBit = XorCST->getValue();
1068 ICmpInst::Predicate Pred = ICI.isSigned()
1069 ? ICI.getUnsignedPredicate()
1070 : ICI.getSignedPredicate();
1071 Pred = ICI.getSwappedPredicate(Pred);
1072 return new ICmpInst(Pred, LHSI->getOperand(0),
1073 ConstantInt::get(ICI.getContext(),
1074 RHSV ^ NotSignBit));
1079 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1080 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1081 LHSI->getOperand(0)->hasOneUse()) {
1082 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1084 // If the LHS is an AND of a truncating cast, we can widen the
1085 // and/compare to be the input width without changing the value
1086 // produced, eliminating a cast.
1087 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1088 // We can do this transformation if either the AND constant does not
1089 // have its sign bit set or if it is an equality comparison.
1090 // Extending a relational comparison when we're checking the sign
1091 // bit would not work.
1092 if (Cast->hasOneUse() &&
1093 (ICI.isEquality() ||
1094 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
1096 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
1097 APInt NewCST = AndCST->getValue().zext(BitWidth);
1098 APInt NewCI = RHSV.zext(BitWidth);
1100 Builder->CreateAnd(Cast->getOperand(0),
1101 ConstantInt::get(ICI.getContext(), NewCST),
1103 return new ICmpInst(ICI.getPredicate(), NewAnd,
1104 ConstantInt::get(ICI.getContext(), NewCI));
1108 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1109 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1110 // happens a LOT in code produced by the C front-end, for bitfield
1112 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1113 if (Shift && !Shift->isShift())
1117 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1118 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1119 const Type *AndTy = AndCST->getType(); // Type of the and.
1121 // We can fold this as long as we can't shift unknown bits
1122 // into the mask. This can only happen with signed shift
1123 // rights, as they sign-extend.
1125 bool CanFold = Shift->isLogicalShift();
1127 // To test for the bad case of the signed shr, see if any
1128 // of the bits shifted in could be tested after the mask.
1129 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1130 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1132 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1133 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1134 AndCST->getValue()) == 0)
1140 if (Shift->getOpcode() == Instruction::Shl)
1141 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1143 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1145 // Check to see if we are shifting out any of the bits being
1147 if (ConstantExpr::get(Shift->getOpcode(),
1148 NewCst, ShAmt) != RHS) {
1149 // If we shifted bits out, the fold is not going to work out.
1150 // As a special case, check to see if this means that the
1151 // result is always true or false now.
1152 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1153 return ReplaceInstUsesWith(ICI,
1154 ConstantInt::getFalse(ICI.getContext()));
1155 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1156 return ReplaceInstUsesWith(ICI,
1157 ConstantInt::getTrue(ICI.getContext()));
1159 ICI.setOperand(1, NewCst);
1160 Constant *NewAndCST;
1161 if (Shift->getOpcode() == Instruction::Shl)
1162 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1164 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1165 LHSI->setOperand(1, NewAndCST);
1166 LHSI->setOperand(0, Shift->getOperand(0));
1167 Worklist.Add(Shift); // Shift is dead.
1173 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1174 // preferable because it allows the C<<Y expression to be hoisted out
1175 // of a loop if Y is invariant and X is not.
1176 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1177 ICI.isEquality() && !Shift->isArithmeticShift() &&
1178 !isa<Constant>(Shift->getOperand(0))) {
1181 if (Shift->getOpcode() == Instruction::LShr) {
1182 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
1184 // Insert a logical shift.
1185 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
1188 // Compute X & (C << Y).
1190 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1192 ICI.setOperand(0, NewAnd);
1197 // Try to optimize things like "A[i]&42 == 0" to index computations.
1198 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1199 if (GetElementPtrInst *GEP =
1200 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1201 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1202 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1203 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1204 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1205 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1211 case Instruction::Or: {
1212 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1215 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1216 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1217 // -> and (icmp eq P, null), (icmp eq Q, null).
1218 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1219 Constant::getNullValue(P->getType()));
1220 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1221 Constant::getNullValue(Q->getType()));
1223 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1224 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1226 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1232 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1233 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1236 uint32_t TypeBits = RHSV.getBitWidth();
1238 // Check that the shift amount is in range. If not, don't perform
1239 // undefined shifts. When the shift is visited it will be
1241 if (ShAmt->uge(TypeBits))
1244 if (ICI.isEquality()) {
1245 // If we are comparing against bits always shifted out, the
1246 // comparison cannot succeed.
1248 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1250 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1251 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1253 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1254 return ReplaceInstUsesWith(ICI, Cst);
1257 // If the shift is NUW, then it is just shifting out zeros, no need for an
1259 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1260 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1261 ConstantExpr::getLShr(RHS, ShAmt));
1263 if (LHSI->hasOneUse()) {
1264 // Otherwise strength reduce the shift into an and.
1265 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1267 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1268 TypeBits-ShAmtVal));
1271 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1272 return new ICmpInst(ICI.getPredicate(), And,
1273 ConstantExpr::getLShr(RHS, ShAmt));
1277 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1278 bool TrueIfSigned = false;
1279 if (LHSI->hasOneUse() &&
1280 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1281 // (X << 31) <s 0 --> (X&1) != 0
1282 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1283 APInt::getOneBitSet(TypeBits,
1284 TypeBits-ShAmt->getZExtValue()-1));
1286 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1287 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1288 And, Constant::getNullValue(And->getType()));
1293 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1294 case Instruction::AShr: {
1295 // Handle equality comparisons of shift-by-constant.
1296 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1297 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1298 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1302 // Handle exact shr's.
1303 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1304 if (RHSV.isMinValue())
1305 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1310 case Instruction::SDiv:
1311 case Instruction::UDiv:
1312 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1313 // Fold this div into the comparison, producing a range check.
1314 // Determine, based on the divide type, what the range is being
1315 // checked. If there is an overflow on the low or high side, remember
1316 // it, otherwise compute the range [low, hi) bounding the new value.
1317 // See: InsertRangeTest above for the kinds of replacements possible.
1318 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1319 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1324 case Instruction::Add:
1325 // Fold: icmp pred (add X, C1), C2
1326 if (!ICI.isEquality()) {
1327 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1329 const APInt &LHSV = LHSC->getValue();
1331 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1334 if (ICI.isSigned()) {
1335 if (CR.getLower().isSignBit()) {
1336 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1337 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1338 } else if (CR.getUpper().isSignBit()) {
1339 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1340 ConstantInt::get(ICI.getContext(),CR.getLower()));
1343 if (CR.getLower().isMinValue()) {
1344 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1345 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1346 } else if (CR.getUpper().isMinValue()) {
1347 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1348 ConstantInt::get(ICI.getContext(),CR.getLower()));
1355 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1356 if (ICI.isEquality()) {
1357 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1359 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1360 // the second operand is a constant, simplify a bit.
1361 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1362 switch (BO->getOpcode()) {
1363 case Instruction::SRem:
1364 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1365 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1366 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1367 if (V.sgt(1) && V.isPowerOf2()) {
1369 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1371 return new ICmpInst(ICI.getPredicate(), NewRem,
1372 Constant::getNullValue(BO->getType()));
1376 case Instruction::Add:
1377 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1378 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1379 if (BO->hasOneUse())
1380 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1381 ConstantExpr::getSub(RHS, BOp1C));
1382 } else if (RHSV == 0) {
1383 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1384 // efficiently invertible, or if the add has just this one use.
1385 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1387 if (Value *NegVal = dyn_castNegVal(BOp1))
1388 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1389 if (Value *NegVal = dyn_castNegVal(BOp0))
1390 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1391 if (BO->hasOneUse()) {
1392 Value *Neg = Builder->CreateNeg(BOp1);
1394 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1398 case Instruction::Xor:
1399 // For the xor case, we can xor two constants together, eliminating
1400 // the explicit xor.
1401 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
1402 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1403 ConstantExpr::getXor(RHS, BOC));
1406 case Instruction::Sub:
1407 // Replace (([sub|xor] A, B) != 0) with (A != B)
1409 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1413 case Instruction::Or:
1414 // If bits are being or'd in that are not present in the constant we
1415 // are comparing against, then the comparison could never succeed!
1416 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1417 Constant *NotCI = ConstantExpr::getNot(RHS);
1418 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1419 return ReplaceInstUsesWith(ICI,
1420 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1425 case Instruction::And:
1426 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1427 // If bits are being compared against that are and'd out, then the
1428 // comparison can never succeed!
1429 if ((RHSV & ~BOC->getValue()) != 0)
1430 return ReplaceInstUsesWith(ICI,
1431 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1434 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1435 if (RHS == BOC && RHSV.isPowerOf2())
1436 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1437 ICmpInst::ICMP_NE, LHSI,
1438 Constant::getNullValue(RHS->getType()));
1440 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1441 if (BOC->getValue().isSignBit()) {
1442 Value *X = BO->getOperand(0);
1443 Constant *Zero = Constant::getNullValue(X->getType());
1444 ICmpInst::Predicate pred = isICMP_NE ?
1445 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1446 return new ICmpInst(pred, X, Zero);
1449 // ((X & ~7) == 0) --> X < 8
1450 if (RHSV == 0 && isHighOnes(BOC)) {
1451 Value *X = BO->getOperand(0);
1452 Constant *NegX = ConstantExpr::getNeg(BOC);
1453 ICmpInst::Predicate pred = isICMP_NE ?
1454 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1455 return new ICmpInst(pred, X, NegX);
1460 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1461 // Handle icmp {eq|ne} <intrinsic>, intcst.
1462 switch (II->getIntrinsicID()) {
1463 case Intrinsic::bswap:
1465 ICI.setOperand(0, II->getArgOperand(0));
1466 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1468 case Intrinsic::ctlz:
1469 case Intrinsic::cttz:
1470 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1471 if (RHSV == RHS->getType()->getBitWidth()) {
1473 ICI.setOperand(0, II->getArgOperand(0));
1474 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1478 case Intrinsic::ctpop:
1479 // popcount(A) == 0 -> A == 0 and likewise for !=
1480 if (RHS->isZero()) {
1482 ICI.setOperand(0, II->getArgOperand(0));
1483 ICI.setOperand(1, RHS);
1495 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1496 /// We only handle extending casts so far.
1498 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1499 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1500 Value *LHSCIOp = LHSCI->getOperand(0);
1501 const Type *SrcTy = LHSCIOp->getType();
1502 const Type *DestTy = LHSCI->getType();
1505 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1506 // integer type is the same size as the pointer type.
1507 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1508 TD->getPointerSizeInBits() ==
1509 cast<IntegerType>(DestTy)->getBitWidth()) {
1511 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1512 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1513 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1514 RHSOp = RHSC->getOperand(0);
1515 // If the pointer types don't match, insert a bitcast.
1516 if (LHSCIOp->getType() != RHSOp->getType())
1517 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1521 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1524 // The code below only handles extension cast instructions, so far.
1526 if (LHSCI->getOpcode() != Instruction::ZExt &&
1527 LHSCI->getOpcode() != Instruction::SExt)
1530 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1531 bool isSignedCmp = ICI.isSigned();
1533 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1534 // Not an extension from the same type?
1535 RHSCIOp = CI->getOperand(0);
1536 if (RHSCIOp->getType() != LHSCIOp->getType())
1539 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1540 // and the other is a zext), then we can't handle this.
1541 if (CI->getOpcode() != LHSCI->getOpcode())
1544 // Deal with equality cases early.
1545 if (ICI.isEquality())
1546 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1548 // A signed comparison of sign extended values simplifies into a
1549 // signed comparison.
1550 if (isSignedCmp && isSignedExt)
1551 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1553 // The other three cases all fold into an unsigned comparison.
1554 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1557 // If we aren't dealing with a constant on the RHS, exit early
1558 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1562 // Compute the constant that would happen if we truncated to SrcTy then
1563 // reextended to DestTy.
1564 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1565 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1568 // If the re-extended constant didn't change...
1570 // Deal with equality cases early.
1571 if (ICI.isEquality())
1572 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1574 // A signed comparison of sign extended values simplifies into a
1575 // signed comparison.
1576 if (isSignedExt && isSignedCmp)
1577 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1579 // The other three cases all fold into an unsigned comparison.
1580 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1583 // The re-extended constant changed so the constant cannot be represented
1584 // in the shorter type. Consequently, we cannot emit a simple comparison.
1585 // All the cases that fold to true or false will have already been handled
1586 // by SimplifyICmpInst, so only deal with the tricky case.
1588 if (isSignedCmp || !isSignedExt)
1591 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1592 // should have been folded away previously and not enter in here.
1594 // We're performing an unsigned comp with a sign extended value.
1595 // This is true if the input is >= 0. [aka >s -1]
1596 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1597 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1599 // Finally, return the value computed.
1600 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1601 return ReplaceInstUsesWith(ICI, Result);
1603 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1604 return BinaryOperator::CreateNot(Result);
1607 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1608 /// I = icmp ugt (add (add A, B), CI2), CI1
1609 /// If this is of the form:
1611 /// if (sum+128 >u 255)
1612 /// Then replace it with llvm.sadd.with.overflow.i8.
1614 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1615 ConstantInt *CI2, ConstantInt *CI1,
1617 // The transformation we're trying to do here is to transform this into an
1618 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1619 // with a narrower add, and discard the add-with-constant that is part of the
1620 // range check (if we can't eliminate it, this isn't profitable).
1622 // In order to eliminate the add-with-constant, the compare can be its only
1624 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1625 if (!AddWithCst->hasOneUse()) return 0;
1627 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1628 if (!CI2->getValue().isPowerOf2()) return 0;
1629 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1630 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1632 // The width of the new add formed is 1 more than the bias.
1635 // Check to see that CI1 is an all-ones value with NewWidth bits.
1636 if (CI1->getBitWidth() == NewWidth ||
1637 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1640 // In order to replace the original add with a narrower
1641 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1642 // and truncates that discard the high bits of the add. Verify that this is
1644 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1645 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1647 if (*UI == AddWithCst) continue;
1649 // Only accept truncates for now. We would really like a nice recursive
1650 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1651 // chain to see which bits of a value are actually demanded. If the
1652 // original add had another add which was then immediately truncated, we
1653 // could still do the transformation.
1654 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1656 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1659 // If the pattern matches, truncate the inputs to the narrower type and
1660 // use the sadd_with_overflow intrinsic to efficiently compute both the
1661 // result and the overflow bit.
1662 Module *M = I.getParent()->getParent()->getParent();
1664 const Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1665 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1668 InstCombiner::BuilderTy *Builder = IC.Builder;
1670 // Put the new code above the original add, in case there are any uses of the
1671 // add between the add and the compare.
1672 Builder->SetInsertPoint(OrigAdd);
1674 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1675 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1676 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1677 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1678 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1680 // The inner add was the result of the narrow add, zero extended to the
1681 // wider type. Replace it with the result computed by the intrinsic.
1682 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1684 // The original icmp gets replaced with the overflow value.
1685 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1688 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1690 // Don't bother doing this transformation for pointers, don't do it for
1692 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1694 // If the add is a constant expr, then we don't bother transforming it.
1695 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1696 if (OrigAdd == 0) return 0;
1698 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1700 // Put the new code above the original add, in case there are any uses of the
1701 // add between the add and the compare.
1702 InstCombiner::BuilderTy *Builder = IC.Builder;
1703 Builder->SetInsertPoint(OrigAdd);
1705 Module *M = I.getParent()->getParent()->getParent();
1706 const Type *Ty = LHS->getType();
1707 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, &Ty,1);
1708 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1709 Value *Add = Builder->CreateExtractValue(Call, 0);
1711 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1713 // The original icmp gets replaced with the overflow value.
1714 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1717 // DemandedBitsLHSMask - When performing a comparison against a constant,
1718 // it is possible that not all the bits in the LHS are demanded. This helper
1719 // method computes the mask that IS demanded.
1720 static APInt DemandedBitsLHSMask(ICmpInst &I,
1721 unsigned BitWidth, bool isSignCheck) {
1723 return APInt::getSignBit(BitWidth);
1725 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1726 if (!CI) return APInt::getAllOnesValue(BitWidth);
1727 const APInt &RHS = CI->getValue();
1729 switch (I.getPredicate()) {
1730 // For a UGT comparison, we don't care about any bits that
1731 // correspond to the trailing ones of the comparand. The value of these
1732 // bits doesn't impact the outcome of the comparison, because any value
1733 // greater than the RHS must differ in a bit higher than these due to carry.
1734 case ICmpInst::ICMP_UGT: {
1735 unsigned trailingOnes = RHS.countTrailingOnes();
1736 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1740 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1741 // Any value less than the RHS must differ in a higher bit because of carries.
1742 case ICmpInst::ICMP_ULT: {
1743 unsigned trailingZeros = RHS.countTrailingZeros();
1744 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1749 return APInt::getAllOnesValue(BitWidth);
1754 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1755 bool Changed = false;
1756 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1758 /// Orders the operands of the compare so that they are listed from most
1759 /// complex to least complex. This puts constants before unary operators,
1760 /// before binary operators.
1761 if (getComplexity(Op0) < getComplexity(Op1)) {
1763 std::swap(Op0, Op1);
1767 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1768 return ReplaceInstUsesWith(I, V);
1770 const Type *Ty = Op0->getType();
1772 // icmp's with boolean values can always be turned into bitwise operations
1773 if (Ty->isIntegerTy(1)) {
1774 switch (I.getPredicate()) {
1775 default: llvm_unreachable("Invalid icmp instruction!");
1776 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1777 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1778 return BinaryOperator::CreateNot(Xor);
1780 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1781 return BinaryOperator::CreateXor(Op0, Op1);
1783 case ICmpInst::ICMP_UGT:
1784 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1786 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1787 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1788 return BinaryOperator::CreateAnd(Not, Op1);
1790 case ICmpInst::ICMP_SGT:
1791 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1793 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1794 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1795 return BinaryOperator::CreateAnd(Not, Op0);
1797 case ICmpInst::ICMP_UGE:
1798 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1800 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1801 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1802 return BinaryOperator::CreateOr(Not, Op1);
1804 case ICmpInst::ICMP_SGE:
1805 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1807 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1808 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1809 return BinaryOperator::CreateOr(Not, Op0);
1814 unsigned BitWidth = 0;
1815 if (Ty->isIntOrIntVectorTy())
1816 BitWidth = Ty->getScalarSizeInBits();
1817 else if (TD) // Pointers require TD info to get their size.
1818 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1820 bool isSignBit = false;
1822 // See if we are doing a comparison with a constant.
1823 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1824 Value *A = 0, *B = 0;
1826 // Match the following pattern, which is a common idiom when writing
1827 // overflow-safe integer arithmetic function. The source performs an
1828 // addition in wider type, and explicitly checks for overflow using
1829 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1830 // sadd_with_overflow intrinsic.
1832 // TODO: This could probably be generalized to handle other overflow-safe
1833 // operations if we worked out the formulas to compute the appropriate
1837 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1839 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1840 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1841 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1842 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1846 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1847 if (I.isEquality() && CI->isZero() &&
1848 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1849 // (icmp cond A B) if cond is equality
1850 return new ICmpInst(I.getPredicate(), A, B);
1853 // If we have an icmp le or icmp ge instruction, turn it into the
1854 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1855 // them being folded in the code below. The SimplifyICmpInst code has
1856 // already handled the edge cases for us, so we just assert on them.
1857 switch (I.getPredicate()) {
1859 case ICmpInst::ICMP_ULE:
1860 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1861 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1862 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1863 case ICmpInst::ICMP_SLE:
1864 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1865 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1866 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1867 case ICmpInst::ICMP_UGE:
1868 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1869 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1870 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1871 case ICmpInst::ICMP_SGE:
1872 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1873 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1874 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1877 // If this comparison is a normal comparison, it demands all
1878 // bits, if it is a sign bit comparison, it only demands the sign bit.
1880 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1883 // See if we can fold the comparison based on range information we can get
1884 // by checking whether bits are known to be zero or one in the input.
1885 if (BitWidth != 0) {
1886 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1887 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1889 if (SimplifyDemandedBits(I.getOperandUse(0),
1890 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1891 Op0KnownZero, Op0KnownOne, 0))
1893 if (SimplifyDemandedBits(I.getOperandUse(1),
1894 APInt::getAllOnesValue(BitWidth),
1895 Op1KnownZero, Op1KnownOne, 0))
1898 // Given the known and unknown bits, compute a range that the LHS could be
1899 // in. Compute the Min, Max and RHS values based on the known bits. For the
1900 // EQ and NE we use unsigned values.
1901 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1902 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1904 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1906 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1909 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1911 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1915 // If Min and Max are known to be the same, then SimplifyDemandedBits
1916 // figured out that the LHS is a constant. Just constant fold this now so
1917 // that code below can assume that Min != Max.
1918 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1919 return new ICmpInst(I.getPredicate(),
1920 ConstantInt::get(Op0->getType(), Op0Min), Op1);
1921 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1922 return new ICmpInst(I.getPredicate(), Op0,
1923 ConstantInt::get(Op1->getType(), Op1Min));
1925 // Based on the range information we know about the LHS, see if we can
1926 // simplify this comparison. For example, (x&4) < 8 is always true.
1927 switch (I.getPredicate()) {
1928 default: llvm_unreachable("Unknown icmp opcode!");
1929 case ICmpInst::ICMP_EQ: {
1930 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1931 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
1933 // If all bits are known zero except for one, then we know at most one
1934 // bit is set. If the comparison is against zero, then this is a check
1935 // to see if *that* bit is set.
1936 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1937 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1938 // If the LHS is an AND with the same constant, look through it.
1940 ConstantInt *LHSC = 0;
1941 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1942 LHSC->getValue() != Op0KnownZeroInverted)
1945 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1946 // then turn "((1 << x)&8) == 0" into "x != 3".
1948 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1949 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1950 return new ICmpInst(ICmpInst::ICMP_NE, X,
1951 ConstantInt::get(X->getType(), CmpVal));
1954 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1955 // then turn "((8 >>u x)&1) == 0" into "x != 3".
1957 if (Op0KnownZeroInverted == 1 &&
1958 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1959 return new ICmpInst(ICmpInst::ICMP_NE, X,
1960 ConstantInt::get(X->getType(),
1961 CI->countTrailingZeros()));
1966 case ICmpInst::ICMP_NE: {
1967 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1968 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
1970 // If all bits are known zero except for one, then we know at most one
1971 // bit is set. If the comparison is against zero, then this is a check
1972 // to see if *that* bit is set.
1973 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1974 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1975 // If the LHS is an AND with the same constant, look through it.
1977 ConstantInt *LHSC = 0;
1978 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1979 LHSC->getValue() != Op0KnownZeroInverted)
1982 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1983 // then turn "((1 << x)&8) != 0" into "x == 3".
1985 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1986 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1987 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1988 ConstantInt::get(X->getType(), CmpVal));
1991 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1992 // then turn "((8 >>u x)&1) != 0" into "x == 3".
1994 if (Op0KnownZeroInverted == 1 &&
1995 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1996 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1997 ConstantInt::get(X->getType(),
1998 CI->countTrailingZeros()));
2003 case ICmpInst::ICMP_ULT:
2004 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2005 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2006 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2007 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2008 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2009 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2010 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2011 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2012 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2013 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2015 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2016 if (CI->isMinValue(true))
2017 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2018 Constant::getAllOnesValue(Op0->getType()));
2021 case ICmpInst::ICMP_UGT:
2022 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2023 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2024 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2025 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2027 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2028 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2029 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2030 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2031 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2032 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2034 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2035 if (CI->isMaxValue(true))
2036 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2037 Constant::getNullValue(Op0->getType()));
2040 case ICmpInst::ICMP_SLT:
2041 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2042 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2043 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2044 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2045 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2046 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2047 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2048 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2049 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2050 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2053 case ICmpInst::ICMP_SGT:
2054 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2055 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2056 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2057 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2059 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2060 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2061 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2062 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2063 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2064 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2067 case ICmpInst::ICMP_SGE:
2068 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2069 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2070 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2071 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2072 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2074 case ICmpInst::ICMP_SLE:
2075 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2076 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2077 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2078 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2079 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2081 case ICmpInst::ICMP_UGE:
2082 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2083 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2084 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2085 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2086 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2088 case ICmpInst::ICMP_ULE:
2089 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2090 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2091 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2092 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2093 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2097 // Turn a signed comparison into an unsigned one if both operands
2098 // are known to have the same sign.
2100 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2101 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2102 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2105 // Test if the ICmpInst instruction is used exclusively by a select as
2106 // part of a minimum or maximum operation. If so, refrain from doing
2107 // any other folding. This helps out other analyses which understand
2108 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2109 // and CodeGen. And in this case, at least one of the comparison
2110 // operands has at least one user besides the compare (the select),
2111 // which would often largely negate the benefit of folding anyway.
2113 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2114 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2115 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2118 // See if we are doing a comparison between a constant and an instruction that
2119 // can be folded into the comparison.
2120 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2121 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2122 // instruction, see if that instruction also has constants so that the
2123 // instruction can be folded into the icmp
2124 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2125 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2129 // Handle icmp with constant (but not simple integer constant) RHS
2130 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2131 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2132 switch (LHSI->getOpcode()) {
2133 case Instruction::GetElementPtr:
2134 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2135 if (RHSC->isNullValue() &&
2136 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2137 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2138 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2140 case Instruction::PHI:
2141 // Only fold icmp into the PHI if the phi and icmp are in the same
2142 // block. If in the same block, we're encouraging jump threading. If
2143 // not, we are just pessimizing the code by making an i1 phi.
2144 if (LHSI->getParent() == I.getParent())
2145 if (Instruction *NV = FoldOpIntoPhi(I))
2148 case Instruction::Select: {
2149 // If either operand of the select is a constant, we can fold the
2150 // comparison into the select arms, which will cause one to be
2151 // constant folded and the select turned into a bitwise or.
2152 Value *Op1 = 0, *Op2 = 0;
2153 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2154 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2155 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2156 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2158 // We only want to perform this transformation if it will not lead to
2159 // additional code. This is true if either both sides of the select
2160 // fold to a constant (in which case the icmp is replaced with a select
2161 // which will usually simplify) or this is the only user of the
2162 // select (in which case we are trading a select+icmp for a simpler
2164 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2166 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2169 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2171 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2175 case Instruction::IntToPtr:
2176 // icmp pred inttoptr(X), null -> icmp pred X, 0
2177 if (RHSC->isNullValue() && TD &&
2178 TD->getIntPtrType(RHSC->getContext()) ==
2179 LHSI->getOperand(0)->getType())
2180 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2181 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2184 case Instruction::Load:
2185 // Try to optimize things like "A[i] > 4" to index computations.
2186 if (GetElementPtrInst *GEP =
2187 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2188 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2189 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2190 !cast<LoadInst>(LHSI)->isVolatile())
2191 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2198 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2199 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2200 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2202 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2203 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2204 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2207 // Test to see if the operands of the icmp are casted versions of other
2208 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2210 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2211 if (Op0->getType()->isPointerTy() &&
2212 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2213 // We keep moving the cast from the left operand over to the right
2214 // operand, where it can often be eliminated completely.
2215 Op0 = CI->getOperand(0);
2217 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2218 // so eliminate it as well.
2219 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2220 Op1 = CI2->getOperand(0);
2222 // If Op1 is a constant, we can fold the cast into the constant.
2223 if (Op0->getType() != Op1->getType()) {
2224 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2225 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2227 // Otherwise, cast the RHS right before the icmp
2228 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2231 return new ICmpInst(I.getPredicate(), Op0, Op1);
2235 if (isa<CastInst>(Op0)) {
2236 // Handle the special case of: icmp (cast bool to X), <cst>
2237 // This comes up when you have code like
2240 // For generality, we handle any zero-extension of any operand comparison
2241 // with a constant or another cast from the same type.
2242 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2243 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2247 // Special logic for binary operators.
2248 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2249 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2251 CmpInst::Predicate Pred = I.getPredicate();
2252 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2253 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2254 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2255 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2256 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2257 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2258 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2259 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2260 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2262 // Analyze the case when either Op0 or Op1 is an add instruction.
2263 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2264 Value *A = 0, *B = 0, *C = 0, *D = 0;
2265 if (BO0 && BO0->getOpcode() == Instruction::Add)
2266 A = BO0->getOperand(0), B = BO0->getOperand(1);
2267 if (BO1 && BO1->getOpcode() == Instruction::Add)
2268 C = BO1->getOperand(0), D = BO1->getOperand(1);
2270 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2271 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2272 return new ICmpInst(Pred, A == Op1 ? B : A,
2273 Constant::getNullValue(Op1->getType()));
2275 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2276 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2277 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2280 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2281 if (A && C && (A == C || A == D || B == C || B == D) &&
2282 NoOp0WrapProblem && NoOp1WrapProblem &&
2283 // Try not to increase register pressure.
2284 BO0->hasOneUse() && BO1->hasOneUse()) {
2285 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2286 Value *Y = (A == C || A == D) ? B : A;
2287 Value *Z = (C == A || C == B) ? D : C;
2288 return new ICmpInst(Pred, Y, Z);
2291 // Analyze the case when either Op0 or Op1 is a sub instruction.
2292 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2293 A = 0; B = 0; C = 0; D = 0;
2294 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2295 A = BO0->getOperand(0), B = BO0->getOperand(1);
2296 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2297 C = BO1->getOperand(0), D = BO1->getOperand(1);
2299 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2300 if (A == Op1 && NoOp0WrapProblem)
2301 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2303 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2304 if (C == Op0 && NoOp1WrapProblem)
2305 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2307 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2308 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2309 // Try not to increase register pressure.
2310 BO0->hasOneUse() && BO1->hasOneUse())
2311 return new ICmpInst(Pred, A, C);
2313 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2314 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2315 // Try not to increase register pressure.
2316 BO0->hasOneUse() && BO1->hasOneUse())
2317 return new ICmpInst(Pred, D, B);
2319 BinaryOperator *SRem = NULL;
2320 // icmp (srem X, Y), Y
2321 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2322 Op1 == BO0->getOperand(1))
2324 // icmp Y, (srem X, Y)
2325 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2326 Op0 == BO1->getOperand(1))
2329 // We don't check hasOneUse to avoid increasing register pressure because
2330 // the value we use is the same value this instruction was already using.
2331 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2333 case ICmpInst::ICMP_EQ:
2334 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2335 case ICmpInst::ICMP_NE:
2336 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2337 case ICmpInst::ICMP_SGT:
2338 case ICmpInst::ICMP_SGE:
2339 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2340 Constant::getAllOnesValue(SRem->getType()));
2341 case ICmpInst::ICMP_SLT:
2342 case ICmpInst::ICMP_SLE:
2343 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2344 Constant::getNullValue(SRem->getType()));
2348 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2349 BO0->hasOneUse() && BO1->hasOneUse() &&
2350 BO0->getOperand(1) == BO1->getOperand(1)) {
2351 switch (BO0->getOpcode()) {
2353 case Instruction::Add:
2354 case Instruction::Sub:
2355 case Instruction::Xor:
2356 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2357 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2358 BO1->getOperand(0));
2359 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2360 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2361 if (CI->getValue().isSignBit()) {
2362 ICmpInst::Predicate Pred = I.isSigned()
2363 ? I.getUnsignedPredicate()
2364 : I.getSignedPredicate();
2365 return new ICmpInst(Pred, BO0->getOperand(0),
2366 BO1->getOperand(0));
2369 if (CI->getValue().isMaxSignedValue()) {
2370 ICmpInst::Predicate Pred = I.isSigned()
2371 ? I.getUnsignedPredicate()
2372 : I.getSignedPredicate();
2373 Pred = I.getSwappedPredicate(Pred);
2374 return new ICmpInst(Pred, BO0->getOperand(0),
2375 BO1->getOperand(0));
2379 case Instruction::Mul:
2380 if (!I.isEquality())
2383 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2384 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2385 // Mask = -1 >> count-trailing-zeros(Cst).
2386 if (!CI->isZero() && !CI->isOne()) {
2387 const APInt &AP = CI->getValue();
2388 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2389 APInt::getLowBitsSet(AP.getBitWidth(),
2391 AP.countTrailingZeros()));
2392 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2393 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2394 return new ICmpInst(I.getPredicate(), And1, And2);
2398 case Instruction::UDiv:
2399 case Instruction::LShr:
2403 case Instruction::SDiv:
2404 case Instruction::AShr:
2405 if (!BO0->isExact() || !BO1->isExact())
2407 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2408 BO1->getOperand(0));
2409 case Instruction::Shl: {
2410 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2411 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2414 if (!NSW && I.isSigned())
2416 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2417 BO1->getOperand(0));
2424 // ~x < ~y --> y < x
2425 // ~x < cst --> ~cst < x
2426 if (match(Op0, m_Not(m_Value(A)))) {
2427 if (match(Op1, m_Not(m_Value(B))))
2428 return new ICmpInst(I.getPredicate(), B, A);
2429 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2430 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2433 // (a+b) <u a --> llvm.uadd.with.overflow.
2434 // (a+b) <u b --> llvm.uadd.with.overflow.
2435 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2436 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2437 (Op1 == A || Op1 == B))
2438 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2441 // a >u (a+b) --> llvm.uadd.with.overflow.
2442 // b >u (a+b) --> llvm.uadd.with.overflow.
2443 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2444 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2445 (Op0 == A || Op0 == B))
2446 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2450 if (I.isEquality()) {
2451 Value *A, *B, *C, *D;
2453 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2454 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2455 Value *OtherVal = A == Op1 ? B : A;
2456 return new ICmpInst(I.getPredicate(), OtherVal,
2457 Constant::getNullValue(A->getType()));
2460 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2461 // A^c1 == C^c2 --> A == C^(c1^c2)
2462 ConstantInt *C1, *C2;
2463 if (match(B, m_ConstantInt(C1)) &&
2464 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2465 Constant *NC = ConstantInt::get(I.getContext(),
2466 C1->getValue() ^ C2->getValue());
2467 Value *Xor = Builder->CreateXor(C, NC, "tmp");
2468 return new ICmpInst(I.getPredicate(), A, Xor);
2471 // A^B == A^D -> B == D
2472 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2473 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2474 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2475 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2479 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2480 (A == Op0 || B == Op0)) {
2481 // A == (A^B) -> B == 0
2482 Value *OtherVal = A == Op0 ? B : A;
2483 return new ICmpInst(I.getPredicate(), OtherVal,
2484 Constant::getNullValue(A->getType()));
2487 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2488 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2489 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2490 Value *X = 0, *Y = 0, *Z = 0;
2493 X = B; Y = D; Z = A;
2494 } else if (A == D) {
2495 X = B; Y = C; Z = A;
2496 } else if (B == C) {
2497 X = A; Y = D; Z = B;
2498 } else if (B == D) {
2499 X = A; Y = C; Z = B;
2502 if (X) { // Build (X^Y) & Z
2503 Op1 = Builder->CreateXor(X, Y, "tmp");
2504 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
2505 I.setOperand(0, Op1);
2506 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2511 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2512 // "icmp (and X, mask), cst"
2515 if (Op0->hasOneUse() &&
2516 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2517 m_ConstantInt(ShAmt))))) &&
2518 match(Op1, m_ConstantInt(Cst1)) &&
2519 // Only do this when A has multiple uses. This is most important to do
2520 // when it exposes other optimizations.
2522 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2524 if (ShAmt < ASize) {
2526 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2529 APInt CmpV = Cst1->getValue().zext(ASize);
2532 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2533 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2539 Value *X; ConstantInt *Cst;
2541 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2542 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2545 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2546 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2548 return Changed ? &I : 0;
2556 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2558 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2561 if (!isa<ConstantFP>(RHSC)) return 0;
2562 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2564 // Get the width of the mantissa. We don't want to hack on conversions that
2565 // might lose information from the integer, e.g. "i64 -> float"
2566 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2567 if (MantissaWidth == -1) return 0; // Unknown.
2569 // Check to see that the input is converted from an integer type that is small
2570 // enough that preserves all bits. TODO: check here for "known" sign bits.
2571 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2572 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2574 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2575 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2579 // If the conversion would lose info, don't hack on this.
2580 if ((int)InputSize > MantissaWidth)
2583 // Otherwise, we can potentially simplify the comparison. We know that it
2584 // will always come through as an integer value and we know the constant is
2585 // not a NAN (it would have been previously simplified).
2586 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2588 ICmpInst::Predicate Pred;
2589 switch (I.getPredicate()) {
2590 default: llvm_unreachable("Unexpected predicate!");
2591 case FCmpInst::FCMP_UEQ:
2592 case FCmpInst::FCMP_OEQ:
2593 Pred = ICmpInst::ICMP_EQ;
2595 case FCmpInst::FCMP_UGT:
2596 case FCmpInst::FCMP_OGT:
2597 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2599 case FCmpInst::FCMP_UGE:
2600 case FCmpInst::FCMP_OGE:
2601 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2603 case FCmpInst::FCMP_ULT:
2604 case FCmpInst::FCMP_OLT:
2605 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2607 case FCmpInst::FCMP_ULE:
2608 case FCmpInst::FCMP_OLE:
2609 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2611 case FCmpInst::FCMP_UNE:
2612 case FCmpInst::FCMP_ONE:
2613 Pred = ICmpInst::ICMP_NE;
2615 case FCmpInst::FCMP_ORD:
2616 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2617 case FCmpInst::FCMP_UNO:
2618 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2621 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2623 // Now we know that the APFloat is a normal number, zero or inf.
2625 // See if the FP constant is too large for the integer. For example,
2626 // comparing an i8 to 300.0.
2627 unsigned IntWidth = IntTy->getScalarSizeInBits();
2630 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2631 // and large values.
2632 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2633 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2634 APFloat::rmNearestTiesToEven);
2635 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2636 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2637 Pred == ICmpInst::ICMP_SLE)
2638 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2639 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2642 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2643 // +INF and large values.
2644 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2645 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2646 APFloat::rmNearestTiesToEven);
2647 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2648 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2649 Pred == ICmpInst::ICMP_ULE)
2650 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2651 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2656 // See if the RHS value is < SignedMin.
2657 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2658 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2659 APFloat::rmNearestTiesToEven);
2660 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2661 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2662 Pred == ICmpInst::ICMP_SGE)
2663 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2664 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2668 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2669 // [0, UMAX], but it may still be fractional. See if it is fractional by
2670 // casting the FP value to the integer value and back, checking for equality.
2671 // Don't do this for zero, because -0.0 is not fractional.
2672 Constant *RHSInt = LHSUnsigned
2673 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2674 : ConstantExpr::getFPToSI(RHSC, IntTy);
2675 if (!RHS.isZero()) {
2676 bool Equal = LHSUnsigned
2677 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2678 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2680 // If we had a comparison against a fractional value, we have to adjust
2681 // the compare predicate and sometimes the value. RHSC is rounded towards
2682 // zero at this point.
2684 default: llvm_unreachable("Unexpected integer comparison!");
2685 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2686 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2687 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2688 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2689 case ICmpInst::ICMP_ULE:
2690 // (float)int <= 4.4 --> int <= 4
2691 // (float)int <= -4.4 --> false
2692 if (RHS.isNegative())
2693 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2695 case ICmpInst::ICMP_SLE:
2696 // (float)int <= 4.4 --> int <= 4
2697 // (float)int <= -4.4 --> int < -4
2698 if (RHS.isNegative())
2699 Pred = ICmpInst::ICMP_SLT;
2701 case ICmpInst::ICMP_ULT:
2702 // (float)int < -4.4 --> false
2703 // (float)int < 4.4 --> int <= 4
2704 if (RHS.isNegative())
2705 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2706 Pred = ICmpInst::ICMP_ULE;
2708 case ICmpInst::ICMP_SLT:
2709 // (float)int < -4.4 --> int < -4
2710 // (float)int < 4.4 --> int <= 4
2711 if (!RHS.isNegative())
2712 Pred = ICmpInst::ICMP_SLE;
2714 case ICmpInst::ICMP_UGT:
2715 // (float)int > 4.4 --> int > 4
2716 // (float)int > -4.4 --> true
2717 if (RHS.isNegative())
2718 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2720 case ICmpInst::ICMP_SGT:
2721 // (float)int > 4.4 --> int > 4
2722 // (float)int > -4.4 --> int >= -4
2723 if (RHS.isNegative())
2724 Pred = ICmpInst::ICMP_SGE;
2726 case ICmpInst::ICMP_UGE:
2727 // (float)int >= -4.4 --> true
2728 // (float)int >= 4.4 --> int > 4
2729 if (!RHS.isNegative())
2730 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2731 Pred = ICmpInst::ICMP_UGT;
2733 case ICmpInst::ICMP_SGE:
2734 // (float)int >= -4.4 --> int >= -4
2735 // (float)int >= 4.4 --> int > 4
2736 if (!RHS.isNegative())
2737 Pred = ICmpInst::ICMP_SGT;
2743 // Lower this FP comparison into an appropriate integer version of the
2745 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2748 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2749 bool Changed = false;
2751 /// Orders the operands of the compare so that they are listed from most
2752 /// complex to least complex. This puts constants before unary operators,
2753 /// before binary operators.
2754 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2759 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2761 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2762 return ReplaceInstUsesWith(I, V);
2764 // Simplify 'fcmp pred X, X'
2766 switch (I.getPredicate()) {
2767 default: llvm_unreachable("Unknown predicate!");
2768 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2769 case FCmpInst::FCMP_ULT: // True if unordered or less than
2770 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2771 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2772 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2773 I.setPredicate(FCmpInst::FCMP_UNO);
2774 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2777 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2778 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2779 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2780 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2781 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2782 I.setPredicate(FCmpInst::FCMP_ORD);
2783 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2788 // Handle fcmp with constant RHS
2789 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2790 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2791 switch (LHSI->getOpcode()) {
2792 case Instruction::FPExt: {
2793 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2794 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2795 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2799 // We can't convert a PPC double double.
2800 if (RHSF->getType()->isPPC_FP128Ty())
2803 const fltSemantics *Sem;
2804 // FIXME: This shouldn't be here.
2805 if (LHSExt->getSrcTy()->isFloatTy())
2806 Sem = &APFloat::IEEEsingle;
2807 else if (LHSExt->getSrcTy()->isDoubleTy())
2808 Sem = &APFloat::IEEEdouble;
2809 else if (LHSExt->getSrcTy()->isFP128Ty())
2810 Sem = &APFloat::IEEEquad;
2811 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2812 Sem = &APFloat::x87DoubleExtended;
2817 APFloat F = RHSF->getValueAPF();
2818 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2820 // Avoid lossy conversions and denormals.
2822 F.compare(APFloat::getSmallestNormalized(*Sem)) !=
2823 APFloat::cmpLessThan)
2824 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2825 ConstantFP::get(RHSC->getContext(), F));
2828 case Instruction::PHI:
2829 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2830 // block. If in the same block, we're encouraging jump threading. If
2831 // not, we are just pessimizing the code by making an i1 phi.
2832 if (LHSI->getParent() == I.getParent())
2833 if (Instruction *NV = FoldOpIntoPhi(I))
2836 case Instruction::SIToFP:
2837 case Instruction::UIToFP:
2838 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2841 case Instruction::Select: {
2842 // If either operand of the select is a constant, we can fold the
2843 // comparison into the select arms, which will cause one to be
2844 // constant folded and the select turned into a bitwise or.
2845 Value *Op1 = 0, *Op2 = 0;
2846 if (LHSI->hasOneUse()) {
2847 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2848 // Fold the known value into the constant operand.
2849 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2850 // Insert a new FCmp of the other select operand.
2851 Op2 = Builder->CreateFCmp(I.getPredicate(),
2852 LHSI->getOperand(2), RHSC, I.getName());
2853 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2854 // Fold the known value into the constant operand.
2855 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2856 // Insert a new FCmp of the other select operand.
2857 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2863 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2866 case Instruction::FSub: {
2867 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2869 if (match(LHSI, m_FNeg(m_Value(Op))))
2870 return new FCmpInst(I.getSwappedPredicate(), Op,
2871 ConstantExpr::getFNeg(RHSC));
2874 case Instruction::Load:
2875 if (GetElementPtrInst *GEP =
2876 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2877 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2878 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2879 !cast<LoadInst>(LHSI)->isVolatile())
2880 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2887 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
2889 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
2890 return new FCmpInst(I.getSwappedPredicate(), X, Y);
2892 // fcmp (fpext x), (fpext y) -> fcmp x, y
2893 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
2894 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
2895 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
2896 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2897 RHSExt->getOperand(0));
2899 return Changed ? &I : 0;