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, InstCombiner &IC) {
473 TargetData &TD = *IC.getTargetData();
474 gep_type_iterator GTI = gep_type_begin(GEP);
476 // Check to see if this gep only has a single variable index. If so, and if
477 // any constant indices are a multiple of its scale, then we can compute this
478 // in terms of the scale of the variable index. For example, if the GEP
479 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
480 // because the expression will cross zero at the same point.
481 unsigned i, e = GEP->getNumOperands();
483 for (i = 1; i != e; ++i, ++GTI) {
484 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
485 // Compute the aggregate offset of constant indices.
486 if (CI->isZero()) continue;
488 // Handle a struct index, which adds its field offset to the pointer.
489 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
490 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
492 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
493 Offset += Size*CI->getSExtValue();
496 // Found our variable index.
501 // If there are no variable indices, we must have a constant offset, just
502 // evaluate it the general way.
503 if (i == e) return 0;
505 Value *VariableIdx = GEP->getOperand(i);
506 // Determine the scale factor of the variable element. For example, this is
507 // 4 if the variable index is into an array of i32.
508 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
510 // Verify that there are no other variable indices. If so, emit the hard way.
511 for (++i, ++GTI; i != e; ++i, ++GTI) {
512 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
515 // Compute the aggregate offset of constant indices.
516 if (CI->isZero()) continue;
518 // Handle a struct index, which adds its field offset to the pointer.
519 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
520 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
522 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
523 Offset += Size*CI->getSExtValue();
527 // Okay, we know we have a single variable index, which must be a
528 // pointer/array/vector index. If there is no offset, life is simple, return
530 unsigned IntPtrWidth = TD.getPointerSizeInBits();
532 // Cast to intptrty in case a truncation occurs. If an extension is needed,
533 // we don't need to bother extending: the extension won't affect where the
534 // computation crosses zero.
535 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
536 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
537 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
542 // Otherwise, there is an index. The computation we will do will be modulo
543 // the pointer size, so get it.
544 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
546 Offset &= PtrSizeMask;
547 VariableScale &= PtrSizeMask;
549 // To do this transformation, any constant index must be a multiple of the
550 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
551 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
552 // multiple of the variable scale.
553 int64_t NewOffs = Offset / (int64_t)VariableScale;
554 if (Offset != NewOffs*(int64_t)VariableScale)
557 // Okay, we can do this evaluation. Start by converting the index to intptr.
558 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
559 if (VariableIdx->getType() != IntPtrTy)
560 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
562 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
563 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
566 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
567 /// else. At this point we know that the GEP is on the LHS of the comparison.
568 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
569 ICmpInst::Predicate Cond,
571 // Look through bitcasts.
572 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
573 RHS = BCI->getOperand(0);
575 Value *PtrBase = GEPLHS->getOperand(0);
576 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
577 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
578 // This transformation (ignoring the base and scales) is valid because we
579 // know pointers can't overflow since the gep is inbounds. See if we can
580 // output an optimized form.
581 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
583 // If not, synthesize the offset the hard way.
585 Offset = EmitGEPOffset(GEPLHS);
586 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
587 Constant::getNullValue(Offset->getType()));
588 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
589 // If the base pointers are different, but the indices are the same, just
590 // compare the base pointer.
591 if (PtrBase != GEPRHS->getOperand(0)) {
592 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
593 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
594 GEPRHS->getOperand(0)->getType();
596 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
597 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
598 IndicesTheSame = false;
602 // If all indices are the same, just compare the base pointers.
604 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
605 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
607 // Otherwise, the base pointers are different and the indices are
608 // different, bail out.
612 // If one of the GEPs has all zero indices, recurse.
613 bool AllZeros = true;
614 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
615 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
616 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
621 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
622 ICmpInst::getSwappedPredicate(Cond), I);
624 // If the other GEP has all zero indices, recurse.
626 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
627 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
628 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
633 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
635 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
636 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
637 // If the GEPs only differ by one index, compare it.
638 unsigned NumDifferences = 0; // Keep track of # differences.
639 unsigned DiffOperand = 0; // The operand that differs.
640 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
641 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
642 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
643 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
644 // Irreconcilable differences.
648 if (NumDifferences++) break;
653 if (NumDifferences == 0) // SAME GEP?
654 return ReplaceInstUsesWith(I, // No comparison is needed here.
655 ConstantInt::get(Type::getInt1Ty(I.getContext()),
656 ICmpInst::isTrueWhenEqual(Cond)));
658 else if (NumDifferences == 1 && GEPsInBounds) {
659 Value *LHSV = GEPLHS->getOperand(DiffOperand);
660 Value *RHSV = GEPRHS->getOperand(DiffOperand);
661 // Make sure we do a signed comparison here.
662 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
666 // Only lower this if the icmp is the only user of the GEP or if we expect
667 // the result to fold to a constant!
670 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
671 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
672 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
673 Value *L = EmitGEPOffset(GEPLHS);
674 Value *R = EmitGEPOffset(GEPRHS);
675 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
681 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
682 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
683 Value *X, ConstantInt *CI,
684 ICmpInst::Predicate Pred,
686 // If we have X+0, exit early (simplifying logic below) and let it get folded
687 // elsewhere. icmp X+0, X -> icmp X, X
689 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
690 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
693 // (X+4) == X -> false.
694 if (Pred == ICmpInst::ICMP_EQ)
695 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
697 // (X+4) != X -> true.
698 if (Pred == ICmpInst::ICMP_NE)
699 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
701 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
702 // so the values can never be equal. Similarly for all other "or equals"
705 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
706 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
707 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
708 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
710 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
711 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
714 // (X+1) >u X --> X <u (0-1) --> X != 255
715 // (X+2) >u X --> X <u (0-2) --> X <u 254
716 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
717 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
718 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
720 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
721 ConstantInt *SMax = ConstantInt::get(X->getContext(),
722 APInt::getSignedMaxValue(BitWidth));
724 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
725 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
726 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
727 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
728 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
729 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
730 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
731 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
733 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
734 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
735 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
736 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
737 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
738 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
740 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
741 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
742 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
745 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
746 /// and CmpRHS are both known to be integer constants.
747 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
748 ConstantInt *DivRHS) {
749 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
750 const APInt &CmpRHSV = CmpRHS->getValue();
752 // FIXME: If the operand types don't match the type of the divide
753 // then don't attempt this transform. The code below doesn't have the
754 // logic to deal with a signed divide and an unsigned compare (and
755 // vice versa). This is because (x /s C1) <s C2 produces different
756 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
757 // (x /u C1) <u C2. Simply casting the operands and result won't
758 // work. :( The if statement below tests that condition and bails
760 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
761 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
763 if (DivRHS->isZero())
764 return 0; // The ProdOV computation fails on divide by zero.
765 if (DivIsSigned && DivRHS->isAllOnesValue())
766 return 0; // The overflow computation also screws up here
767 if (DivRHS->isOne()) {
768 // This eliminates some funny cases with INT_MIN.
769 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
773 // Compute Prod = CI * DivRHS. We are essentially solving an equation
774 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
775 // C2 (CI). By solving for X we can turn this into a range check
776 // instead of computing a divide.
777 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
779 // Determine if the product overflows by seeing if the product is
780 // not equal to the divide. Make sure we do the same kind of divide
781 // as in the LHS instruction that we're folding.
782 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
783 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
785 // Get the ICmp opcode
786 ICmpInst::Predicate Pred = ICI.getPredicate();
788 /// If the division is known to be exact, then there is no remainder from the
789 /// divide, so the covered range size is unit, otherwise it is the divisor.
790 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
792 // Figure out the interval that is being checked. For example, a comparison
793 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
794 // Compute this interval based on the constants involved and the signedness of
795 // the compare/divide. This computes a half-open interval, keeping track of
796 // whether either value in the interval overflows. After analysis each
797 // overflow variable is set to 0 if it's corresponding bound variable is valid
798 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
799 int LoOverflow = 0, HiOverflow = 0;
800 Constant *LoBound = 0, *HiBound = 0;
802 if (!DivIsSigned) { // udiv
803 // e.g. X/5 op 3 --> [15, 20)
805 HiOverflow = LoOverflow = ProdOV;
807 // If this is not an exact divide, then many values in the range collapse
808 // to the same result value.
809 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
812 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
813 if (CmpRHSV == 0) { // (X / pos) op 0
814 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
815 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
817 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
818 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
819 HiOverflow = LoOverflow = ProdOV;
821 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
822 } else { // (X / pos) op neg
823 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
824 HiBound = AddOne(Prod);
825 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
827 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
828 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
831 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
833 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
834 if (CmpRHSV == 0) { // (X / neg) op 0
835 // e.g. X/-5 op 0 --> [-4, 5)
836 LoBound = AddOne(RangeSize);
837 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
838 if (HiBound == DivRHS) { // -INTMIN = INTMIN
839 HiOverflow = 1; // [INTMIN+1, overflow)
840 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
842 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
843 // e.g. X/-5 op 3 --> [-19, -14)
844 HiBound = AddOne(Prod);
845 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
847 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
848 } else { // (X / neg) op neg
849 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
850 LoOverflow = HiOverflow = ProdOV;
852 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
855 // Dividing by a negative swaps the condition. LT <-> GT
856 Pred = ICmpInst::getSwappedPredicate(Pred);
859 Value *X = DivI->getOperand(0);
861 default: llvm_unreachable("Unhandled icmp opcode!");
862 case ICmpInst::ICMP_EQ:
863 if (LoOverflow && HiOverflow)
864 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
866 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
867 ICmpInst::ICMP_UGE, X, LoBound);
869 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
870 ICmpInst::ICMP_ULT, X, HiBound);
871 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
873 case ICmpInst::ICMP_NE:
874 if (LoOverflow && HiOverflow)
875 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
877 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
878 ICmpInst::ICMP_ULT, X, LoBound);
880 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
881 ICmpInst::ICMP_UGE, X, HiBound);
882 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
883 DivIsSigned, false));
884 case ICmpInst::ICMP_ULT:
885 case ICmpInst::ICMP_SLT:
886 if (LoOverflow == +1) // Low bound is greater than input range.
887 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
888 if (LoOverflow == -1) // Low bound is less than input range.
889 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
890 return new ICmpInst(Pred, X, LoBound);
891 case ICmpInst::ICMP_UGT:
892 case ICmpInst::ICMP_SGT:
893 if (HiOverflow == +1) // High bound greater than input range.
894 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
895 if (HiOverflow == -1) // High bound less than input range.
896 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
897 if (Pred == ICmpInst::ICMP_UGT)
898 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
899 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
903 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
904 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
905 ConstantInt *ShAmt) {
906 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
908 // Check that the shift amount is in range. If not, don't perform
909 // undefined shifts. When the shift is visited it will be
911 uint32_t TypeBits = CmpRHSV.getBitWidth();
912 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
913 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
916 if (!ICI.isEquality()) {
917 // If we have an unsigned comparison and an ashr, we can't simplify this.
918 // Similarly for signed comparisons with lshr.
919 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
922 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
923 // by a power of 2. Since we already have logic to simplify these,
924 // transform to div and then simplify the resultant comparison.
925 if (Shr->getOpcode() == Instruction::AShr &&
926 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
929 // Revisit the shift (to delete it).
933 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
936 Shr->getOpcode() == Instruction::AShr ?
937 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
938 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
940 ICI.setOperand(0, Tmp);
942 // If the builder folded the binop, just return it.
943 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
947 // Otherwise, fold this div/compare.
948 assert(TheDiv->getOpcode() == Instruction::SDiv ||
949 TheDiv->getOpcode() == Instruction::UDiv);
951 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
952 assert(Res && "This div/cst should have folded!");
957 // If we are comparing against bits always shifted out, the
958 // comparison cannot succeed.
959 APInt Comp = CmpRHSV << ShAmtVal;
960 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
961 if (Shr->getOpcode() == Instruction::LShr)
962 Comp = Comp.lshr(ShAmtVal);
964 Comp = Comp.ashr(ShAmtVal);
966 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
967 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
968 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
970 return ReplaceInstUsesWith(ICI, Cst);
973 // Otherwise, check to see if the bits shifted out are known to be zero.
974 // If so, we can compare against the unshifted value:
975 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
976 if (Shr->hasOneUse() && Shr->isExact())
977 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
979 if (Shr->hasOneUse()) {
980 // Otherwise strength reduce the shift into an and.
981 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
982 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
984 Value *And = Builder->CreateAnd(Shr->getOperand(0),
985 Mask, Shr->getName()+".mask");
986 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
992 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
994 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
997 const APInt &RHSV = RHS->getValue();
999 switch (LHSI->getOpcode()) {
1000 case Instruction::Trunc:
1001 if (ICI.isEquality() && LHSI->hasOneUse()) {
1002 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1003 // of the high bits truncated out of x are known.
1004 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1005 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1006 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
1007 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1008 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
1010 // If all the high bits are known, we can do this xform.
1011 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1012 // Pull in the high bits from known-ones set.
1013 APInt NewRHS = RHS->getValue().zext(SrcBits);
1015 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1016 ConstantInt::get(ICI.getContext(), NewRHS));
1021 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1022 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1023 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1025 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1026 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1027 Value *CompareVal = LHSI->getOperand(0);
1029 // If the sign bit of the XorCST is not set, there is no change to
1030 // the operation, just stop using the Xor.
1031 if (!XorCST->getValue().isNegative()) {
1032 ICI.setOperand(0, CompareVal);
1037 // Was the old condition true if the operand is positive?
1038 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1040 // If so, the new one isn't.
1041 isTrueIfPositive ^= true;
1043 if (isTrueIfPositive)
1044 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1047 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1051 if (LHSI->hasOneUse()) {
1052 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1053 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1054 const APInt &SignBit = XorCST->getValue();
1055 ICmpInst::Predicate Pred = ICI.isSigned()
1056 ? ICI.getUnsignedPredicate()
1057 : ICI.getSignedPredicate();
1058 return new ICmpInst(Pred, LHSI->getOperand(0),
1059 ConstantInt::get(ICI.getContext(),
1063 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1064 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
1065 const APInt &NotSignBit = XorCST->getValue();
1066 ICmpInst::Predicate Pred = ICI.isSigned()
1067 ? ICI.getUnsignedPredicate()
1068 : ICI.getSignedPredicate();
1069 Pred = ICI.getSwappedPredicate(Pred);
1070 return new ICmpInst(Pred, LHSI->getOperand(0),
1071 ConstantInt::get(ICI.getContext(),
1072 RHSV ^ NotSignBit));
1077 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1078 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1079 LHSI->getOperand(0)->hasOneUse()) {
1080 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1082 // If the LHS is an AND of a truncating cast, we can widen the
1083 // and/compare to be the input width without changing the value
1084 // produced, eliminating a cast.
1085 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1086 // We can do this transformation if either the AND constant does not
1087 // have its sign bit set or if it is an equality comparison.
1088 // Extending a relational comparison when we're checking the sign
1089 // bit would not work.
1090 if (ICI.isEquality() ||
1091 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative())) {
1093 Builder->CreateAnd(Cast->getOperand(0),
1094 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1095 NewAnd->takeName(LHSI);
1096 return new ICmpInst(ICI.getPredicate(), NewAnd,
1097 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1101 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1102 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1103 // happens a LOT in code produced by the C front-end, for bitfield
1105 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1106 if (Shift && !Shift->isShift())
1110 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1111 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1112 const Type *AndTy = AndCST->getType(); // Type of the and.
1114 // We can fold this as long as we can't shift unknown bits
1115 // into the mask. This can only happen with signed shift
1116 // rights, as they sign-extend.
1118 bool CanFold = Shift->isLogicalShift();
1120 // To test for the bad case of the signed shr, see if any
1121 // of the bits shifted in could be tested after the mask.
1122 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1123 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1125 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1126 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1127 AndCST->getValue()) == 0)
1133 if (Shift->getOpcode() == Instruction::Shl)
1134 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1136 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1138 // Check to see if we are shifting out any of the bits being
1140 if (ConstantExpr::get(Shift->getOpcode(),
1141 NewCst, ShAmt) != RHS) {
1142 // If we shifted bits out, the fold is not going to work out.
1143 // As a special case, check to see if this means that the
1144 // result is always true or false now.
1145 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1146 return ReplaceInstUsesWith(ICI,
1147 ConstantInt::getFalse(ICI.getContext()));
1148 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1149 return ReplaceInstUsesWith(ICI,
1150 ConstantInt::getTrue(ICI.getContext()));
1152 ICI.setOperand(1, NewCst);
1153 Constant *NewAndCST;
1154 if (Shift->getOpcode() == Instruction::Shl)
1155 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1157 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1158 LHSI->setOperand(1, NewAndCST);
1159 LHSI->setOperand(0, Shift->getOperand(0));
1160 Worklist.Add(Shift); // Shift is dead.
1166 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1167 // preferable because it allows the C<<Y expression to be hoisted out
1168 // of a loop if Y is invariant and X is not.
1169 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1170 ICI.isEquality() && !Shift->isArithmeticShift() &&
1171 !isa<Constant>(Shift->getOperand(0))) {
1174 if (Shift->getOpcode() == Instruction::LShr) {
1175 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
1177 // Insert a logical shift.
1178 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
1181 // Compute X & (C << Y).
1183 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1185 ICI.setOperand(0, NewAnd);
1190 // Try to optimize things like "A[i]&42 == 0" to index computations.
1191 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1192 if (GetElementPtrInst *GEP =
1193 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1194 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1195 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1196 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1197 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1198 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1204 case Instruction::Or: {
1205 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1208 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1209 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1210 // -> and (icmp eq P, null), (icmp eq Q, null).
1211 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1212 Constant::getNullValue(P->getType()));
1213 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1214 Constant::getNullValue(Q->getType()));
1216 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1217 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1219 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1225 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1226 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1229 uint32_t TypeBits = RHSV.getBitWidth();
1231 // Check that the shift amount is in range. If not, don't perform
1232 // undefined shifts. When the shift is visited it will be
1234 if (ShAmt->uge(TypeBits))
1237 if (ICI.isEquality()) {
1238 // If we are comparing against bits always shifted out, the
1239 // comparison cannot succeed.
1241 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1243 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1244 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1246 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1247 return ReplaceInstUsesWith(ICI, Cst);
1250 // If the shift is NUW, then it is just shifting out zeros, no need for an
1252 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1253 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1254 ConstantExpr::getLShr(RHS, ShAmt));
1256 if (LHSI->hasOneUse()) {
1257 // Otherwise strength reduce the shift into an and.
1258 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1260 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1261 TypeBits-ShAmtVal));
1264 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1265 return new ICmpInst(ICI.getPredicate(), And,
1266 ConstantExpr::getLShr(RHS, ShAmt));
1270 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1271 bool TrueIfSigned = false;
1272 if (LHSI->hasOneUse() &&
1273 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1274 // (X << 31) <s 0 --> (X&1) != 0
1275 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1276 APInt::getOneBitSet(TypeBits,
1277 TypeBits-ShAmt->getZExtValue()-1));
1279 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1280 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1281 And, Constant::getNullValue(And->getType()));
1286 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1287 case Instruction::AShr: {
1288 // Handle equality comparisons of shift-by-constant.
1289 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1290 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1291 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1295 // Handle exact shr's.
1296 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1297 if (RHSV.isMinValue())
1298 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1303 case Instruction::SDiv:
1304 case Instruction::UDiv:
1305 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1306 // Fold this div into the comparison, producing a range check.
1307 // Determine, based on the divide type, what the range is being
1308 // checked. If there is an overflow on the low or high side, remember
1309 // it, otherwise compute the range [low, hi) bounding the new value.
1310 // See: InsertRangeTest above for the kinds of replacements possible.
1311 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1312 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1317 case Instruction::Add:
1318 // Fold: icmp pred (add X, C1), C2
1319 if (!ICI.isEquality()) {
1320 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1322 const APInt &LHSV = LHSC->getValue();
1324 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1327 if (ICI.isSigned()) {
1328 if (CR.getLower().isSignBit()) {
1329 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1330 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1331 } else if (CR.getUpper().isSignBit()) {
1332 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1333 ConstantInt::get(ICI.getContext(),CR.getLower()));
1336 if (CR.getLower().isMinValue()) {
1337 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1338 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1339 } else if (CR.getUpper().isMinValue()) {
1340 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1341 ConstantInt::get(ICI.getContext(),CR.getLower()));
1348 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1349 if (ICI.isEquality()) {
1350 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1352 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1353 // the second operand is a constant, simplify a bit.
1354 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1355 switch (BO->getOpcode()) {
1356 case Instruction::SRem:
1357 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1358 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1359 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1360 if (V.sgt(1) && V.isPowerOf2()) {
1362 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1364 return new ICmpInst(ICI.getPredicate(), NewRem,
1365 Constant::getNullValue(BO->getType()));
1369 case Instruction::Add:
1370 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1371 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1372 if (BO->hasOneUse())
1373 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1374 ConstantExpr::getSub(RHS, BOp1C));
1375 } else if (RHSV == 0) {
1376 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1377 // efficiently invertible, or if the add has just this one use.
1378 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1380 if (Value *NegVal = dyn_castNegVal(BOp1))
1381 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1382 if (Value *NegVal = dyn_castNegVal(BOp0))
1383 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1384 if (BO->hasOneUse()) {
1385 Value *Neg = Builder->CreateNeg(BOp1);
1387 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1391 case Instruction::Xor:
1392 // For the xor case, we can xor two constants together, eliminating
1393 // the explicit xor.
1394 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
1395 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1396 ConstantExpr::getXor(RHS, BOC));
1399 case Instruction::Sub:
1400 // Replace (([sub|xor] A, B) != 0) with (A != B)
1402 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1406 case Instruction::Or:
1407 // If bits are being or'd in that are not present in the constant we
1408 // are comparing against, then the comparison could never succeed!
1409 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1410 Constant *NotCI = ConstantExpr::getNot(RHS);
1411 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1412 return ReplaceInstUsesWith(ICI,
1413 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1418 case Instruction::And:
1419 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1420 // If bits are being compared against that are and'd out, then the
1421 // comparison can never succeed!
1422 if ((RHSV & ~BOC->getValue()) != 0)
1423 return ReplaceInstUsesWith(ICI,
1424 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1427 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1428 if (RHS == BOC && RHSV.isPowerOf2())
1429 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1430 ICmpInst::ICMP_NE, LHSI,
1431 Constant::getNullValue(RHS->getType()));
1433 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1434 if (BOC->getValue().isSignBit()) {
1435 Value *X = BO->getOperand(0);
1436 Constant *Zero = Constant::getNullValue(X->getType());
1437 ICmpInst::Predicate pred = isICMP_NE ?
1438 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1439 return new ICmpInst(pred, X, Zero);
1442 // ((X & ~7) == 0) --> X < 8
1443 if (RHSV == 0 && isHighOnes(BOC)) {
1444 Value *X = BO->getOperand(0);
1445 Constant *NegX = ConstantExpr::getNeg(BOC);
1446 ICmpInst::Predicate pred = isICMP_NE ?
1447 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1448 return new ICmpInst(pred, X, NegX);
1453 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1454 // Handle icmp {eq|ne} <intrinsic>, intcst.
1455 switch (II->getIntrinsicID()) {
1456 case Intrinsic::bswap:
1458 ICI.setOperand(0, II->getArgOperand(0));
1459 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1461 case Intrinsic::ctlz:
1462 case Intrinsic::cttz:
1463 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1464 if (RHSV == RHS->getType()->getBitWidth()) {
1466 ICI.setOperand(0, II->getArgOperand(0));
1467 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1471 case Intrinsic::ctpop:
1472 // popcount(A) == 0 -> A == 0 and likewise for !=
1473 if (RHS->isZero()) {
1475 ICI.setOperand(0, II->getArgOperand(0));
1476 ICI.setOperand(1, RHS);
1488 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1489 /// We only handle extending casts so far.
1491 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1492 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1493 Value *LHSCIOp = LHSCI->getOperand(0);
1494 const Type *SrcTy = LHSCIOp->getType();
1495 const Type *DestTy = LHSCI->getType();
1498 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1499 // integer type is the same size as the pointer type.
1500 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1501 TD->getPointerSizeInBits() ==
1502 cast<IntegerType>(DestTy)->getBitWidth()) {
1504 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1505 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1506 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1507 RHSOp = RHSC->getOperand(0);
1508 // If the pointer types don't match, insert a bitcast.
1509 if (LHSCIOp->getType() != RHSOp->getType())
1510 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1514 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1517 // The code below only handles extension cast instructions, so far.
1519 if (LHSCI->getOpcode() != Instruction::ZExt &&
1520 LHSCI->getOpcode() != Instruction::SExt)
1523 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1524 bool isSignedCmp = ICI.isSigned();
1526 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1527 // Not an extension from the same type?
1528 RHSCIOp = CI->getOperand(0);
1529 if (RHSCIOp->getType() != LHSCIOp->getType())
1532 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1533 // and the other is a zext), then we can't handle this.
1534 if (CI->getOpcode() != LHSCI->getOpcode())
1537 // Deal with equality cases early.
1538 if (ICI.isEquality())
1539 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1541 // A signed comparison of sign extended values simplifies into a
1542 // signed comparison.
1543 if (isSignedCmp && isSignedExt)
1544 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1546 // The other three cases all fold into an unsigned comparison.
1547 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1550 // If we aren't dealing with a constant on the RHS, exit early
1551 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1555 // Compute the constant that would happen if we truncated to SrcTy then
1556 // reextended to DestTy.
1557 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1558 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1561 // If the re-extended constant didn't change...
1563 // Deal with equality cases early.
1564 if (ICI.isEquality())
1565 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1567 // A signed comparison of sign extended values simplifies into a
1568 // signed comparison.
1569 if (isSignedExt && isSignedCmp)
1570 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1572 // The other three cases all fold into an unsigned comparison.
1573 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1576 // The re-extended constant changed so the constant cannot be represented
1577 // in the shorter type. Consequently, we cannot emit a simple comparison.
1578 // All the cases that fold to true or false will have already been handled
1579 // by SimplifyICmpInst, so only deal with the tricky case.
1581 if (isSignedCmp || !isSignedExt)
1584 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1585 // should have been folded away previously and not enter in here.
1587 // We're performing an unsigned comp with a sign extended value.
1588 // This is true if the input is >= 0. [aka >s -1]
1589 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1590 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1592 // Finally, return the value computed.
1593 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1594 return ReplaceInstUsesWith(ICI, Result);
1596 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1597 return BinaryOperator::CreateNot(Result);
1600 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1601 /// I = icmp ugt (add (add A, B), CI2), CI1
1602 /// If this is of the form:
1604 /// if (sum+128 >u 255)
1605 /// Then replace it with llvm.sadd.with.overflow.i8.
1607 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1608 ConstantInt *CI2, ConstantInt *CI1,
1610 // The transformation we're trying to do here is to transform this into an
1611 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1612 // with a narrower add, and discard the add-with-constant that is part of the
1613 // range check (if we can't eliminate it, this isn't profitable).
1615 // In order to eliminate the add-with-constant, the compare can be its only
1617 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1618 if (!AddWithCst->hasOneUse()) return 0;
1620 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1621 if (!CI2->getValue().isPowerOf2()) return 0;
1622 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1623 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1625 // The width of the new add formed is 1 more than the bias.
1628 // Check to see that CI1 is an all-ones value with NewWidth bits.
1629 if (CI1->getBitWidth() == NewWidth ||
1630 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1633 // In order to replace the original add with a narrower
1634 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1635 // and truncates that discard the high bits of the add. Verify that this is
1637 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1638 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1640 if (*UI == AddWithCst) continue;
1642 // Only accept truncates for now. We would really like a nice recursive
1643 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1644 // chain to see which bits of a value are actually demanded. If the
1645 // original add had another add which was then immediately truncated, we
1646 // could still do the transformation.
1647 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1649 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1652 // If the pattern matches, truncate the inputs to the narrower type and
1653 // use the sadd_with_overflow intrinsic to efficiently compute both the
1654 // result and the overflow bit.
1655 Module *M = I.getParent()->getParent()->getParent();
1657 const Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1658 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1661 InstCombiner::BuilderTy *Builder = IC.Builder;
1663 // Put the new code above the original add, in case there are any uses of the
1664 // add between the add and the compare.
1665 Builder->SetInsertPoint(OrigAdd);
1667 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1668 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1669 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1670 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1671 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1673 // The inner add was the result of the narrow add, zero extended to the
1674 // wider type. Replace it with the result computed by the intrinsic.
1675 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1677 // The original icmp gets replaced with the overflow value.
1678 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1681 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1683 // Don't bother doing this transformation for pointers, don't do it for
1685 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1687 // If the add is a constant expr, then we don't bother transforming it.
1688 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1689 if (OrigAdd == 0) return 0;
1691 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1693 // Put the new code above the original add, in case there are any uses of the
1694 // add between the add and the compare.
1695 InstCombiner::BuilderTy *Builder = IC.Builder;
1696 Builder->SetInsertPoint(OrigAdd);
1698 Module *M = I.getParent()->getParent()->getParent();
1699 const Type *Ty = LHS->getType();
1700 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, &Ty,1);
1701 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1702 Value *Add = Builder->CreateExtractValue(Call, 0);
1704 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1706 // The original icmp gets replaced with the overflow value.
1707 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1710 // DemandedBitsLHSMask - When performing a comparison against a constant,
1711 // it is possible that not all the bits in the LHS are demanded. This helper
1712 // method computes the mask that IS demanded.
1713 static APInt DemandedBitsLHSMask(ICmpInst &I,
1714 unsigned BitWidth, bool isSignCheck) {
1716 return APInt::getSignBit(BitWidth);
1718 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1719 if (!CI) return APInt::getAllOnesValue(BitWidth);
1720 const APInt &RHS = CI->getValue();
1722 switch (I.getPredicate()) {
1723 // For a UGT comparison, we don't care about any bits that
1724 // correspond to the trailing ones of the comparand. The value of these
1725 // bits doesn't impact the outcome of the comparison, because any value
1726 // greater than the RHS must differ in a bit higher than these due to carry.
1727 case ICmpInst::ICMP_UGT: {
1728 unsigned trailingOnes = RHS.countTrailingOnes();
1729 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1733 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1734 // Any value less than the RHS must differ in a higher bit because of carries.
1735 case ICmpInst::ICMP_ULT: {
1736 unsigned trailingZeros = RHS.countTrailingZeros();
1737 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1742 return APInt::getAllOnesValue(BitWidth);
1747 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1748 bool Changed = false;
1749 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1751 /// Orders the operands of the compare so that they are listed from most
1752 /// complex to least complex. This puts constants before unary operators,
1753 /// before binary operators.
1754 if (getComplexity(Op0) < getComplexity(Op1)) {
1756 std::swap(Op0, Op1);
1760 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1761 return ReplaceInstUsesWith(I, V);
1763 const Type *Ty = Op0->getType();
1765 // icmp's with boolean values can always be turned into bitwise operations
1766 if (Ty->isIntegerTy(1)) {
1767 switch (I.getPredicate()) {
1768 default: llvm_unreachable("Invalid icmp instruction!");
1769 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1770 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1771 return BinaryOperator::CreateNot(Xor);
1773 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1774 return BinaryOperator::CreateXor(Op0, Op1);
1776 case ICmpInst::ICMP_UGT:
1777 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1779 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1780 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1781 return BinaryOperator::CreateAnd(Not, Op1);
1783 case ICmpInst::ICMP_SGT:
1784 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1786 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1787 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1788 return BinaryOperator::CreateAnd(Not, Op0);
1790 case ICmpInst::ICMP_UGE:
1791 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1793 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1794 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1795 return BinaryOperator::CreateOr(Not, Op1);
1797 case ICmpInst::ICMP_SGE:
1798 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1800 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1801 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1802 return BinaryOperator::CreateOr(Not, Op0);
1807 unsigned BitWidth = 0;
1808 if (Ty->isIntOrIntVectorTy())
1809 BitWidth = Ty->getScalarSizeInBits();
1810 else if (TD) // Pointers require TD info to get their size.
1811 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1813 bool isSignBit = false;
1815 // See if we are doing a comparison with a constant.
1816 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1817 Value *A = 0, *B = 0;
1819 // Match the following pattern, which is a common idiom when writing
1820 // overflow-safe integer arithmetic function. The source performs an
1821 // addition in wider type, and explicitly checks for overflow using
1822 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1823 // sadd_with_overflow intrinsic.
1825 // TODO: This could probably be generalized to handle other overflow-safe
1826 // operations if we worked out the formulas to compute the appropriate
1830 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1832 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1833 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1834 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1835 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1839 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1840 if (I.isEquality() && CI->isZero() &&
1841 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1842 // (icmp cond A B) if cond is equality
1843 return new ICmpInst(I.getPredicate(), A, B);
1846 // If we have an icmp le or icmp ge instruction, turn it into the
1847 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1848 // them being folded in the code below. The SimplifyICmpInst code has
1849 // already handled the edge cases for us, so we just assert on them.
1850 switch (I.getPredicate()) {
1852 case ICmpInst::ICMP_ULE:
1853 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1854 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1855 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1856 case ICmpInst::ICMP_SLE:
1857 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1858 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1859 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1860 case ICmpInst::ICMP_UGE:
1861 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1862 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1863 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1864 case ICmpInst::ICMP_SGE:
1865 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1866 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1867 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1870 // If this comparison is a normal comparison, it demands all
1871 // bits, if it is a sign bit comparison, it only demands the sign bit.
1873 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1876 // See if we can fold the comparison based on range information we can get
1877 // by checking whether bits are known to be zero or one in the input.
1878 if (BitWidth != 0) {
1879 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1880 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1882 if (SimplifyDemandedBits(I.getOperandUse(0),
1883 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1884 Op0KnownZero, Op0KnownOne, 0))
1886 if (SimplifyDemandedBits(I.getOperandUse(1),
1887 APInt::getAllOnesValue(BitWidth),
1888 Op1KnownZero, Op1KnownOne, 0))
1891 // Given the known and unknown bits, compute a range that the LHS could be
1892 // in. Compute the Min, Max and RHS values based on the known bits. For the
1893 // EQ and NE we use unsigned values.
1894 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1895 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1897 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1899 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1902 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1904 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1908 // If Min and Max are known to be the same, then SimplifyDemandedBits
1909 // figured out that the LHS is a constant. Just constant fold this now so
1910 // that code below can assume that Min != Max.
1911 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1912 return new ICmpInst(I.getPredicate(),
1913 ConstantInt::get(Op0->getType(), Op0Min), Op1);
1914 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1915 return new ICmpInst(I.getPredicate(), Op0,
1916 ConstantInt::get(Op1->getType(), Op1Min));
1918 // Based on the range information we know about the LHS, see if we can
1919 // simplify this comparison. For example, (x&4) < 8 is always true.
1920 switch (I.getPredicate()) {
1921 default: llvm_unreachable("Unknown icmp opcode!");
1922 case ICmpInst::ICMP_EQ: {
1923 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1924 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
1926 // If all bits are known zero except for one, then we know at most one
1927 // bit is set. If the comparison is against zero, then this is a check
1928 // to see if *that* bit is set.
1929 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1930 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1931 // If the LHS is an AND with the same constant, look through it.
1933 ConstantInt *LHSC = 0;
1934 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1935 LHSC->getValue() != Op0KnownZeroInverted)
1938 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1939 // then turn "((1 << x)&8) == 0" into "x != 3".
1941 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1942 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1943 return new ICmpInst(ICmpInst::ICMP_NE, X,
1944 ConstantInt::get(X->getType(), CmpVal));
1947 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1948 // then turn "((8 >>u x)&1) == 0" into "x != 3".
1950 if (Op0KnownZeroInverted == 1 &&
1951 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1952 return new ICmpInst(ICmpInst::ICMP_NE, X,
1953 ConstantInt::get(X->getType(),
1954 CI->countTrailingZeros()));
1959 case ICmpInst::ICMP_NE: {
1960 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1961 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
1963 // If all bits are known zero except for one, then we know at most one
1964 // bit is set. If the comparison is against zero, then this is a check
1965 // to see if *that* bit is set.
1966 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1967 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1968 // If the LHS is an AND with the same constant, look through it.
1970 ConstantInt *LHSC = 0;
1971 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1972 LHSC->getValue() != Op0KnownZeroInverted)
1975 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1976 // then turn "((1 << x)&8) != 0" into "x == 3".
1978 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1979 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1980 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1981 ConstantInt::get(X->getType(), CmpVal));
1984 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1985 // then turn "((8 >>u x)&1) != 0" into "x == 3".
1987 if (Op0KnownZeroInverted == 1 &&
1988 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1989 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1990 ConstantInt::get(X->getType(),
1991 CI->countTrailingZeros()));
1996 case ICmpInst::ICMP_ULT:
1997 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
1998 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
1999 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2000 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2001 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2002 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2003 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2004 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2005 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2006 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2008 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2009 if (CI->isMinValue(true))
2010 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2011 Constant::getAllOnesValue(Op0->getType()));
2014 case ICmpInst::ICMP_UGT:
2015 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2016 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2017 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2018 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2020 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2021 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2022 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2023 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2024 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2025 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2027 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2028 if (CI->isMaxValue(true))
2029 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2030 Constant::getNullValue(Op0->getType()));
2033 case ICmpInst::ICMP_SLT:
2034 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2035 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2036 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2037 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2038 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2039 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2040 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2041 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2042 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2043 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2046 case ICmpInst::ICMP_SGT:
2047 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2048 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2049 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2050 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2052 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2053 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2054 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2055 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2056 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2057 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2060 case ICmpInst::ICMP_SGE:
2061 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2062 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2063 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2064 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2065 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2067 case ICmpInst::ICMP_SLE:
2068 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2069 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2070 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2071 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2072 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2074 case ICmpInst::ICMP_UGE:
2075 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2076 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2077 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2078 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2079 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2081 case ICmpInst::ICMP_ULE:
2082 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2083 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2084 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2085 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2086 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2090 // Turn a signed comparison into an unsigned one if both operands
2091 // are known to have the same sign.
2093 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2094 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2095 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2098 // Test if the ICmpInst instruction is used exclusively by a select as
2099 // part of a minimum or maximum operation. If so, refrain from doing
2100 // any other folding. This helps out other analyses which understand
2101 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2102 // and CodeGen. And in this case, at least one of the comparison
2103 // operands has at least one user besides the compare (the select),
2104 // which would often largely negate the benefit of folding anyway.
2106 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2107 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2108 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2111 // See if we are doing a comparison between a constant and an instruction that
2112 // can be folded into the comparison.
2113 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2114 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2115 // instruction, see if that instruction also has constants so that the
2116 // instruction can be folded into the icmp
2117 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2118 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2122 // Handle icmp with constant (but not simple integer constant) RHS
2123 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2124 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2125 switch (LHSI->getOpcode()) {
2126 case Instruction::GetElementPtr:
2127 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2128 if (RHSC->isNullValue() &&
2129 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2130 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2131 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2133 case Instruction::PHI:
2134 // Only fold icmp into the PHI if the phi and icmp are in the same
2135 // block. If in the same block, we're encouraging jump threading. If
2136 // not, we are just pessimizing the code by making an i1 phi.
2137 if (LHSI->getParent() == I.getParent())
2138 if (Instruction *NV = FoldOpIntoPhi(I))
2141 case Instruction::Select: {
2142 // If either operand of the select is a constant, we can fold the
2143 // comparison into the select arms, which will cause one to be
2144 // constant folded and the select turned into a bitwise or.
2145 Value *Op1 = 0, *Op2 = 0;
2146 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2147 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2148 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2149 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2151 // We only want to perform this transformation if it will not lead to
2152 // additional code. This is true if either both sides of the select
2153 // fold to a constant (in which case the icmp is replaced with a select
2154 // which will usually simplify) or this is the only user of the
2155 // select (in which case we are trading a select+icmp for a simpler
2157 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2159 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2162 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2164 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2168 case Instruction::IntToPtr:
2169 // icmp pred inttoptr(X), null -> icmp pred X, 0
2170 if (RHSC->isNullValue() && TD &&
2171 TD->getIntPtrType(RHSC->getContext()) ==
2172 LHSI->getOperand(0)->getType())
2173 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2174 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2177 case Instruction::Load:
2178 // Try to optimize things like "A[i] > 4" to index computations.
2179 if (GetElementPtrInst *GEP =
2180 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2181 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2182 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2183 !cast<LoadInst>(LHSI)->isVolatile())
2184 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2191 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2192 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2193 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2195 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2196 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2197 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2200 // Test to see if the operands of the icmp are casted versions of other
2201 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2203 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2204 if (Op0->getType()->isPointerTy() &&
2205 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2206 // We keep moving the cast from the left operand over to the right
2207 // operand, where it can often be eliminated completely.
2208 Op0 = CI->getOperand(0);
2210 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2211 // so eliminate it as well.
2212 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2213 Op1 = CI2->getOperand(0);
2215 // If Op1 is a constant, we can fold the cast into the constant.
2216 if (Op0->getType() != Op1->getType()) {
2217 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2218 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2220 // Otherwise, cast the RHS right before the icmp
2221 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2224 return new ICmpInst(I.getPredicate(), Op0, Op1);
2228 if (isa<CastInst>(Op0)) {
2229 // Handle the special case of: icmp (cast bool to X), <cst>
2230 // This comes up when you have code like
2233 // For generality, we handle any zero-extension of any operand comparison
2234 // with a constant or another cast from the same type.
2235 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2236 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2240 // Special logic for binary operators.
2241 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2242 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2244 CmpInst::Predicate Pred = I.getPredicate();
2245 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2246 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2247 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2248 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2249 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2250 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2251 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2252 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2253 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2255 // Analyze the case when either Op0 or Op1 is an add instruction.
2256 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2257 Value *A = 0, *B = 0, *C = 0, *D = 0;
2258 if (BO0 && BO0->getOpcode() == Instruction::Add)
2259 A = BO0->getOperand(0), B = BO0->getOperand(1);
2260 if (BO1 && BO1->getOpcode() == Instruction::Add)
2261 C = BO1->getOperand(0), D = BO1->getOperand(1);
2263 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2264 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2265 return new ICmpInst(Pred, A == Op1 ? B : A,
2266 Constant::getNullValue(Op1->getType()));
2268 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2269 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2270 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2273 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2274 if (A && C && (A == C || A == D || B == C || B == D) &&
2275 NoOp0WrapProblem && NoOp1WrapProblem &&
2276 // Try not to increase register pressure.
2277 BO0->hasOneUse() && BO1->hasOneUse()) {
2278 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2279 Value *Y = (A == C || A == D) ? B : A;
2280 Value *Z = (C == A || C == B) ? D : C;
2281 return new ICmpInst(Pred, Y, Z);
2284 // Analyze the case when either Op0 or Op1 is a sub instruction.
2285 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2286 A = 0; B = 0; C = 0; D = 0;
2287 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2288 A = BO0->getOperand(0), B = BO0->getOperand(1);
2289 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2290 C = BO1->getOperand(0), D = BO1->getOperand(1);
2292 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2293 if (A == Op1 && NoOp0WrapProblem)
2294 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2296 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2297 if (C == Op0 && NoOp1WrapProblem)
2298 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2300 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2301 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2302 // Try not to increase register pressure.
2303 BO0->hasOneUse() && BO1->hasOneUse())
2304 return new ICmpInst(Pred, A, C);
2306 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2307 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2308 // Try not to increase register pressure.
2309 BO0->hasOneUse() && BO1->hasOneUse())
2310 return new ICmpInst(Pred, D, B);
2312 BinaryOperator *SRem = NULL;
2313 // icmp (srem X, Y), Y
2314 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2315 Op1 == BO0->getOperand(1))
2317 // icmp Y, (srem X, Y)
2318 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2319 Op0 == BO1->getOperand(1))
2322 // We don't check hasOneUse to avoid increasing register pressure because
2323 // the value we use is the same value this instruction was already using.
2324 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2326 case ICmpInst::ICMP_EQ:
2327 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2328 case ICmpInst::ICMP_NE:
2329 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2330 case ICmpInst::ICMP_SGT:
2331 case ICmpInst::ICMP_SGE:
2332 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2333 Constant::getAllOnesValue(SRem->getType()));
2334 case ICmpInst::ICMP_SLT:
2335 case ICmpInst::ICMP_SLE:
2336 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2337 Constant::getNullValue(SRem->getType()));
2341 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2342 BO0->hasOneUse() && BO1->hasOneUse() &&
2343 BO0->getOperand(1) == BO1->getOperand(1)) {
2344 switch (BO0->getOpcode()) {
2346 case Instruction::Add:
2347 case Instruction::Sub:
2348 case Instruction::Xor:
2349 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2350 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2351 BO1->getOperand(0));
2352 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2353 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2354 if (CI->getValue().isSignBit()) {
2355 ICmpInst::Predicate Pred = I.isSigned()
2356 ? I.getUnsignedPredicate()
2357 : I.getSignedPredicate();
2358 return new ICmpInst(Pred, BO0->getOperand(0),
2359 BO1->getOperand(0));
2362 if (CI->getValue().isMaxSignedValue()) {
2363 ICmpInst::Predicate Pred = I.isSigned()
2364 ? I.getUnsignedPredicate()
2365 : I.getSignedPredicate();
2366 Pred = I.getSwappedPredicate(Pred);
2367 return new ICmpInst(Pred, BO0->getOperand(0),
2368 BO1->getOperand(0));
2372 case Instruction::Mul:
2373 if (!I.isEquality())
2376 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2377 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2378 // Mask = -1 >> count-trailing-zeros(Cst).
2379 if (!CI->isZero() && !CI->isOne()) {
2380 const APInt &AP = CI->getValue();
2381 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2382 APInt::getLowBitsSet(AP.getBitWidth(),
2384 AP.countTrailingZeros()));
2385 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2386 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2387 return new ICmpInst(I.getPredicate(), And1, And2);
2391 case Instruction::UDiv:
2392 case Instruction::LShr:
2396 case Instruction::SDiv:
2397 case Instruction::AShr:
2398 if (!BO0->isExact() || !BO1->isExact())
2400 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2401 BO1->getOperand(0));
2402 case Instruction::Shl: {
2403 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2404 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2407 if (!NSW && I.isSigned())
2409 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2410 BO1->getOperand(0));
2417 // ~x < ~y --> y < x
2418 // ~x < cst --> ~cst < x
2419 if (match(Op0, m_Not(m_Value(A)))) {
2420 if (match(Op1, m_Not(m_Value(B))))
2421 return new ICmpInst(I.getPredicate(), B, A);
2422 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2423 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2426 // (a+b) <u a --> llvm.uadd.with.overflow.
2427 // (a+b) <u b --> llvm.uadd.with.overflow.
2428 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2429 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2430 (Op1 == A || Op1 == B))
2431 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2434 // a >u (a+b) --> llvm.uadd.with.overflow.
2435 // b >u (a+b) --> llvm.uadd.with.overflow.
2436 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2437 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2438 (Op0 == A || Op0 == B))
2439 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2443 if (I.isEquality()) {
2444 Value *A, *B, *C, *D;
2446 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2447 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2448 Value *OtherVal = A == Op1 ? B : A;
2449 return new ICmpInst(I.getPredicate(), OtherVal,
2450 Constant::getNullValue(A->getType()));
2453 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2454 // A^c1 == C^c2 --> A == C^(c1^c2)
2455 ConstantInt *C1, *C2;
2456 if (match(B, m_ConstantInt(C1)) &&
2457 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2458 Constant *NC = ConstantInt::get(I.getContext(),
2459 C1->getValue() ^ C2->getValue());
2460 Value *Xor = Builder->CreateXor(C, NC, "tmp");
2461 return new ICmpInst(I.getPredicate(), A, Xor);
2464 // A^B == A^D -> B == D
2465 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2466 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2467 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2468 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2472 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2473 (A == Op0 || B == Op0)) {
2474 // A == (A^B) -> B == 0
2475 Value *OtherVal = A == Op0 ? B : A;
2476 return new ICmpInst(I.getPredicate(), OtherVal,
2477 Constant::getNullValue(A->getType()));
2480 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2481 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2482 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2483 Value *X = 0, *Y = 0, *Z = 0;
2486 X = B; Y = D; Z = A;
2487 } else if (A == D) {
2488 X = B; Y = C; Z = A;
2489 } else if (B == C) {
2490 X = A; Y = D; Z = B;
2491 } else if (B == D) {
2492 X = A; Y = C; Z = B;
2495 if (X) { // Build (X^Y) & Z
2496 Op1 = Builder->CreateXor(X, Y, "tmp");
2497 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
2498 I.setOperand(0, Op1);
2499 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2504 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2505 // "icmp (and X, mask), cst"
2508 if (Op0->hasOneUse() &&
2509 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2510 m_ConstantInt(ShAmt))))) &&
2511 match(Op1, m_ConstantInt(Cst1)) &&
2512 // Only do this when A has multiple uses. This is most important to do
2513 // when it exposes other optimizations.
2515 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2517 if (ShAmt < ASize) {
2519 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2522 APInt CmpV = Cst1->getValue().zext(ASize);
2525 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2526 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2532 Value *X; ConstantInt *Cst;
2534 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2535 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2538 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2539 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2541 return Changed ? &I : 0;
2549 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2551 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2554 if (!isa<ConstantFP>(RHSC)) return 0;
2555 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2557 // Get the width of the mantissa. We don't want to hack on conversions that
2558 // might lose information from the integer, e.g. "i64 -> float"
2559 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2560 if (MantissaWidth == -1) return 0; // Unknown.
2562 // Check to see that the input is converted from an integer type that is small
2563 // enough that preserves all bits. TODO: check here for "known" sign bits.
2564 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2565 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2567 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2568 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2572 // If the conversion would lose info, don't hack on this.
2573 if ((int)InputSize > MantissaWidth)
2576 // Otherwise, we can potentially simplify the comparison. We know that it
2577 // will always come through as an integer value and we know the constant is
2578 // not a NAN (it would have been previously simplified).
2579 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2581 ICmpInst::Predicate Pred;
2582 switch (I.getPredicate()) {
2583 default: llvm_unreachable("Unexpected predicate!");
2584 case FCmpInst::FCMP_UEQ:
2585 case FCmpInst::FCMP_OEQ:
2586 Pred = ICmpInst::ICMP_EQ;
2588 case FCmpInst::FCMP_UGT:
2589 case FCmpInst::FCMP_OGT:
2590 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2592 case FCmpInst::FCMP_UGE:
2593 case FCmpInst::FCMP_OGE:
2594 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2596 case FCmpInst::FCMP_ULT:
2597 case FCmpInst::FCMP_OLT:
2598 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2600 case FCmpInst::FCMP_ULE:
2601 case FCmpInst::FCMP_OLE:
2602 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2604 case FCmpInst::FCMP_UNE:
2605 case FCmpInst::FCMP_ONE:
2606 Pred = ICmpInst::ICMP_NE;
2608 case FCmpInst::FCMP_ORD:
2609 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2610 case FCmpInst::FCMP_UNO:
2611 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2614 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2616 // Now we know that the APFloat is a normal number, zero or inf.
2618 // See if the FP constant is too large for the integer. For example,
2619 // comparing an i8 to 300.0.
2620 unsigned IntWidth = IntTy->getScalarSizeInBits();
2623 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2624 // and large values.
2625 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2626 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2627 APFloat::rmNearestTiesToEven);
2628 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2629 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2630 Pred == ICmpInst::ICMP_SLE)
2631 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2632 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2635 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2636 // +INF and large values.
2637 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2638 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2639 APFloat::rmNearestTiesToEven);
2640 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2641 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2642 Pred == ICmpInst::ICMP_ULE)
2643 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2644 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2649 // See if the RHS value is < SignedMin.
2650 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2651 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2652 APFloat::rmNearestTiesToEven);
2653 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2654 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2655 Pred == ICmpInst::ICMP_SGE)
2656 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2657 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2661 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2662 // [0, UMAX], but it may still be fractional. See if it is fractional by
2663 // casting the FP value to the integer value and back, checking for equality.
2664 // Don't do this for zero, because -0.0 is not fractional.
2665 Constant *RHSInt = LHSUnsigned
2666 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2667 : ConstantExpr::getFPToSI(RHSC, IntTy);
2668 if (!RHS.isZero()) {
2669 bool Equal = LHSUnsigned
2670 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2671 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2673 // If we had a comparison against a fractional value, we have to adjust
2674 // the compare predicate and sometimes the value. RHSC is rounded towards
2675 // zero at this point.
2677 default: llvm_unreachable("Unexpected integer comparison!");
2678 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2679 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2680 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2681 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2682 case ICmpInst::ICMP_ULE:
2683 // (float)int <= 4.4 --> int <= 4
2684 // (float)int <= -4.4 --> false
2685 if (RHS.isNegative())
2686 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2688 case ICmpInst::ICMP_SLE:
2689 // (float)int <= 4.4 --> int <= 4
2690 // (float)int <= -4.4 --> int < -4
2691 if (RHS.isNegative())
2692 Pred = ICmpInst::ICMP_SLT;
2694 case ICmpInst::ICMP_ULT:
2695 // (float)int < -4.4 --> false
2696 // (float)int < 4.4 --> int <= 4
2697 if (RHS.isNegative())
2698 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2699 Pred = ICmpInst::ICMP_ULE;
2701 case ICmpInst::ICMP_SLT:
2702 // (float)int < -4.4 --> int < -4
2703 // (float)int < 4.4 --> int <= 4
2704 if (!RHS.isNegative())
2705 Pred = ICmpInst::ICMP_SLE;
2707 case ICmpInst::ICMP_UGT:
2708 // (float)int > 4.4 --> int > 4
2709 // (float)int > -4.4 --> true
2710 if (RHS.isNegative())
2711 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2713 case ICmpInst::ICMP_SGT:
2714 // (float)int > 4.4 --> int > 4
2715 // (float)int > -4.4 --> int >= -4
2716 if (RHS.isNegative())
2717 Pred = ICmpInst::ICMP_SGE;
2719 case ICmpInst::ICMP_UGE:
2720 // (float)int >= -4.4 --> true
2721 // (float)int >= 4.4 --> int > 4
2722 if (!RHS.isNegative())
2723 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2724 Pred = ICmpInst::ICMP_UGT;
2726 case ICmpInst::ICMP_SGE:
2727 // (float)int >= -4.4 --> int >= -4
2728 // (float)int >= 4.4 --> int > 4
2729 if (!RHS.isNegative())
2730 Pred = ICmpInst::ICMP_SGT;
2736 // Lower this FP comparison into an appropriate integer version of the
2738 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2741 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2742 bool Changed = false;
2744 /// Orders the operands of the compare so that they are listed from most
2745 /// complex to least complex. This puts constants before unary operators,
2746 /// before binary operators.
2747 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2752 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2754 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2755 return ReplaceInstUsesWith(I, V);
2757 // Simplify 'fcmp pred X, X'
2759 switch (I.getPredicate()) {
2760 default: llvm_unreachable("Unknown predicate!");
2761 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2762 case FCmpInst::FCMP_ULT: // True if unordered or less than
2763 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2764 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2765 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2766 I.setPredicate(FCmpInst::FCMP_UNO);
2767 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2770 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2771 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2772 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2773 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2774 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2775 I.setPredicate(FCmpInst::FCMP_ORD);
2776 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2781 // Handle fcmp with constant RHS
2782 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2783 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2784 switch (LHSI->getOpcode()) {
2785 case Instruction::FPExt: {
2786 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2787 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2788 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2792 // We can't convert a PPC double double.
2793 if (RHSF->getType()->isPPC_FP128Ty())
2796 const fltSemantics *Sem;
2797 // FIXME: This shouldn't be here.
2798 if (LHSExt->getSrcTy()->isFloatTy())
2799 Sem = &APFloat::IEEEsingle;
2800 else if (LHSExt->getSrcTy()->isDoubleTy())
2801 Sem = &APFloat::IEEEdouble;
2802 else if (LHSExt->getSrcTy()->isFP128Ty())
2803 Sem = &APFloat::IEEEquad;
2804 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2805 Sem = &APFloat::x87DoubleExtended;
2810 APFloat F = RHSF->getValueAPF();
2811 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2813 // Avoid lossy conversions and denormals.
2815 F.compare(APFloat::getSmallestNormalized(*Sem)) !=
2816 APFloat::cmpLessThan)
2817 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2818 ConstantFP::get(RHSC->getContext(), F));
2821 case Instruction::PHI:
2822 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2823 // block. If in the same block, we're encouraging jump threading. If
2824 // not, we are just pessimizing the code by making an i1 phi.
2825 if (LHSI->getParent() == I.getParent())
2826 if (Instruction *NV = FoldOpIntoPhi(I))
2829 case Instruction::SIToFP:
2830 case Instruction::UIToFP:
2831 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2834 case Instruction::Select: {
2835 // If either operand of the select is a constant, we can fold the
2836 // comparison into the select arms, which will cause one to be
2837 // constant folded and the select turned into a bitwise or.
2838 Value *Op1 = 0, *Op2 = 0;
2839 if (LHSI->hasOneUse()) {
2840 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2841 // Fold the known value into the constant operand.
2842 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2843 // Insert a new FCmp of the other select operand.
2844 Op2 = Builder->CreateFCmp(I.getPredicate(),
2845 LHSI->getOperand(2), RHSC, I.getName());
2846 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2847 // Fold the known value into the constant operand.
2848 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2849 // Insert a new FCmp of the other select operand.
2850 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2856 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2859 case Instruction::FSub: {
2860 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2862 if (match(LHSI, m_FNeg(m_Value(Op))))
2863 return new FCmpInst(I.getSwappedPredicate(), Op,
2864 ConstantExpr::getFNeg(RHSC));
2867 case Instruction::Load:
2868 if (GetElementPtrInst *GEP =
2869 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2870 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2871 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2872 !cast<LoadInst>(LHSI)->isVolatile())
2873 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2880 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
2882 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
2883 return new FCmpInst(I.getSwappedPredicate(), X, Y);
2885 // fcmp (fpext x), (fpext y) -> fcmp x, y
2886 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
2887 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
2888 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
2889 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2890 RHSExt->getOperand(0));
2892 return Changed ? &I : 0;