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/ConstantFolding.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/MemoryBuiltins.h"
19 #include "llvm/Target/TargetData.h"
20 #include "llvm/Support/ConstantRange.h"
21 #include "llvm/Support/GetElementPtrTypeIterator.h"
22 #include "llvm/Support/PatternMatch.h"
24 using namespace PatternMatch;
26 static ConstantInt *getOne(Constant *C) {
27 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
30 /// AddOne - Add one to a ConstantInt
31 static Constant *AddOne(Constant *C) {
32 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
34 /// SubOne - Subtract one from a ConstantInt
35 static Constant *SubOne(Constant *C) {
36 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
39 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
40 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
43 static bool HasAddOverflow(ConstantInt *Result,
44 ConstantInt *In1, ConstantInt *In2,
47 return Result->getValue().ult(In1->getValue());
49 if (In2->isNegative())
50 return Result->getValue().sgt(In1->getValue());
51 return Result->getValue().slt(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 (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 return Result->getValue().ugt(In1->getValue());
83 if (In2->isNegative())
84 return Result->getValue().slt(In1->getValue());
86 return Result->getValue().sgt(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 (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->isMaxValue(true);
132 case ICmpInst::ICMP_UGE:
133 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
135 return RHS->getValue().isSignBit();
141 // isHighOnes - Return true if the constant is of the form 1+0+.
142 // This is the same as lowones(~X).
143 static bool isHighOnes(const ConstantInt *CI) {
144 return (~CI->getValue() + 1).isPowerOf2();
147 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
148 /// set of known zero and one bits, compute the maximum and minimum values that
149 /// could have the specified known zero and known one bits, returning them in
151 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
152 const APInt& KnownOne,
153 APInt& Min, APInt& Max) {
154 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
155 KnownZero.getBitWidth() == Min.getBitWidth() &&
156 KnownZero.getBitWidth() == Max.getBitWidth() &&
157 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
158 APInt UnknownBits = ~(KnownZero|KnownOne);
160 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
161 // bit if it is unknown.
163 Max = KnownOne|UnknownBits;
165 if (UnknownBits.isNegative()) { // Sign bit is unknown
166 Min.setBit(Min.getBitWidth()-1);
167 Max.clearBit(Max.getBitWidth()-1);
171 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
172 // a set of known zero and one bits, compute the maximum and minimum values that
173 // could have the specified known zero and known one bits, returning them in
175 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
176 const APInt &KnownOne,
177 APInt &Min, APInt &Max) {
178 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
179 KnownZero.getBitWidth() == Min.getBitWidth() &&
180 KnownZero.getBitWidth() == Max.getBitWidth() &&
181 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
182 APInt UnknownBits = ~(KnownZero|KnownOne);
184 // The minimum value is when the unknown bits are all zeros.
186 // The maximum value is when the unknown bits are all ones.
187 Max = KnownOne|UnknownBits;
192 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
193 /// cmp pred (load (gep GV, ...)), cmpcst
194 /// where GV is a global variable with a constant initializer. Try to simplify
195 /// this into some simple computation that does not need the load. For example
196 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
198 /// If AndCst is non-null, then the loaded value is masked with that constant
199 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
200 Instruction *InstCombiner::
201 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
202 CmpInst &ICI, ConstantInt *AndCst) {
203 // We need TD information to know the pointer size unless this is inbounds.
204 if (!GEP->isInBounds() && TD == 0) return 0;
206 Constant *Init = GV->getInitializer();
207 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
210 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
211 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays.
213 // There are many forms of this optimization we can handle, for now, just do
214 // the simple index into a single-dimensional array.
216 // Require: GEP GV, 0, i {{, constant indices}}
217 if (GEP->getNumOperands() < 3 ||
218 !isa<ConstantInt>(GEP->getOperand(1)) ||
219 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
220 isa<Constant>(GEP->getOperand(2)))
223 // Check that indices after the variable are constants and in-range for the
224 // type they index. Collect the indices. This is typically for arrays of
226 SmallVector<unsigned, 4> LaterIndices;
228 Type *EltTy = Init->getType()->getArrayElementType();
229 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
230 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
231 if (Idx == 0) return 0; // Variable index.
233 uint64_t IdxVal = Idx->getZExtValue();
234 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
236 if (StructType *STy = dyn_cast<StructType>(EltTy))
237 EltTy = STy->getElementType(IdxVal);
238 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
239 if (IdxVal >= ATy->getNumElements()) return 0;
240 EltTy = ATy->getElementType();
242 return 0; // Unknown type.
245 LaterIndices.push_back(IdxVal);
248 enum { Overdefined = -3, Undefined = -2 };
250 // Variables for our state machines.
252 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
253 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
254 // and 87 is the second (and last) index. FirstTrueElement is -2 when
255 // undefined, otherwise set to the first true element. SecondTrueElement is
256 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
257 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
259 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
260 // form "i != 47 & i != 87". Same state transitions as for true elements.
261 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
263 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
264 /// define a state machine that triggers for ranges of values that the index
265 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
266 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
267 /// index in the range (inclusive). We use -2 for undefined here because we
268 /// use relative comparisons and don't want 0-1 to match -1.
269 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
271 // MagicBitvector - This is a magic bitvector where we set a bit if the
272 // comparison is true for element 'i'. If there are 64 elements or less in
273 // the array, this will fully represent all the comparison results.
274 uint64_t MagicBitvector = 0;
277 // Scan the array and see if one of our patterns matches.
278 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
279 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
280 Constant *Elt = Init->getAggregateElement(i);
281 if (Elt == 0) return 0;
283 // If this is indexing an array of structures, get the structure element.
284 if (!LaterIndices.empty())
285 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
287 // If the element is masked, handle it.
288 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
290 // Find out if the comparison would be true or false for the i'th element.
291 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
292 CompareRHS, TD, TLI);
293 // If the result is undef for this element, ignore it.
294 if (isa<UndefValue>(C)) {
295 // Extend range state machines to cover this element in case there is an
296 // undef in the middle of the range.
297 if (TrueRangeEnd == (int)i-1)
299 if (FalseRangeEnd == (int)i-1)
304 // If we can't compute the result for any of the elements, we have to give
305 // up evaluating the entire conditional.
306 if (!isa<ConstantInt>(C)) return 0;
308 // Otherwise, we know if the comparison is true or false for this element,
309 // update our state machines.
310 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
312 // State machine for single/double/range index comparison.
314 // Update the TrueElement state machine.
315 if (FirstTrueElement == Undefined)
316 FirstTrueElement = TrueRangeEnd = i; // First true element.
318 // Update double-compare state machine.
319 if (SecondTrueElement == Undefined)
320 SecondTrueElement = i;
322 SecondTrueElement = Overdefined;
324 // Update range state machine.
325 if (TrueRangeEnd == (int)i-1)
328 TrueRangeEnd = Overdefined;
331 // Update the FalseElement state machine.
332 if (FirstFalseElement == Undefined)
333 FirstFalseElement = FalseRangeEnd = i; // First false element.
335 // Update double-compare state machine.
336 if (SecondFalseElement == Undefined)
337 SecondFalseElement = i;
339 SecondFalseElement = Overdefined;
341 // Update range state machine.
342 if (FalseRangeEnd == (int)i-1)
345 FalseRangeEnd = Overdefined;
350 // If this element is in range, update our magic bitvector.
351 if (i < 64 && IsTrueForElt)
352 MagicBitvector |= 1ULL << i;
354 // If all of our states become overdefined, bail out early. Since the
355 // predicate is expensive, only check it every 8 elements. This is only
356 // really useful for really huge arrays.
357 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
358 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
359 FalseRangeEnd == Overdefined)
363 // Now that we've scanned the entire array, emit our new comparison(s). We
364 // order the state machines in complexity of the generated code.
365 Value *Idx = GEP->getOperand(2);
367 // If the index is larger than the pointer size of the target, truncate the
368 // index down like the GEP would do implicitly. We don't have to do this for
369 // an inbounds GEP because the index can't be out of range.
370 if (!GEP->isInBounds() &&
371 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
372 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
374 // If the comparison is only true for one or two elements, emit direct
376 if (SecondTrueElement != Overdefined) {
377 // None true -> false.
378 if (FirstTrueElement == Undefined)
379 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
381 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
383 // True for one element -> 'i == 47'.
384 if (SecondTrueElement == Undefined)
385 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
387 // True for two elements -> 'i == 47 | i == 72'.
388 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
389 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
390 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
391 return BinaryOperator::CreateOr(C1, C2);
394 // If the comparison is only false for one or two elements, emit direct
396 if (SecondFalseElement != Overdefined) {
397 // None false -> true.
398 if (FirstFalseElement == Undefined)
399 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
401 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
403 // False for one element -> 'i != 47'.
404 if (SecondFalseElement == Undefined)
405 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
407 // False for two elements -> 'i != 47 & i != 72'.
408 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
409 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
410 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
411 return BinaryOperator::CreateAnd(C1, C2);
414 // If the comparison can be replaced with a range comparison for the elements
415 // where it is true, emit the range check.
416 if (TrueRangeEnd != Overdefined) {
417 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
419 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
420 if (FirstTrueElement) {
421 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
422 Idx = Builder->CreateAdd(Idx, Offs);
425 Value *End = ConstantInt::get(Idx->getType(),
426 TrueRangeEnd-FirstTrueElement+1);
427 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
430 // False range check.
431 if (FalseRangeEnd != Overdefined) {
432 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
433 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
434 if (FirstFalseElement) {
435 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
436 Idx = Builder->CreateAdd(Idx, Offs);
439 Value *End = ConstantInt::get(Idx->getType(),
440 FalseRangeEnd-FirstFalseElement);
441 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
445 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
446 // of this load, replace it with computation that does:
447 // ((magic_cst >> i) & 1) != 0
448 if (ArrayElementCount <= 32 ||
449 (TD && ArrayElementCount <= 64 && TD->isLegalInteger(64))) {
451 if (ArrayElementCount <= 32)
452 Ty = Type::getInt32Ty(Init->getContext());
454 Ty = Type::getInt64Ty(Init->getContext());
455 Value *V = Builder->CreateIntCast(Idx, Ty, false);
456 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
457 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
458 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
465 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
466 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
467 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
468 /// be complex, and scales are involved. The above expression would also be
469 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
470 /// This later form is less amenable to optimization though, and we are allowed
471 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
473 /// If we can't emit an optimized form for this expression, this returns null.
475 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
476 TargetData &TD = *IC.getTargetData();
477 gep_type_iterator GTI = gep_type_begin(GEP);
479 // Check to see if this gep only has a single variable index. If so, and if
480 // any constant indices are a multiple of its scale, then we can compute this
481 // in terms of the scale of the variable index. For example, if the GEP
482 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
483 // because the expression will cross zero at the same point.
484 unsigned i, e = GEP->getNumOperands();
486 for (i = 1; i != e; ++i, ++GTI) {
487 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
488 // Compute the aggregate offset of constant indices.
489 if (CI->isZero()) continue;
491 // Handle a struct index, which adds its field offset to the pointer.
492 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
493 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
495 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
496 Offset += Size*CI->getSExtValue();
499 // Found our variable index.
504 // If there are no variable indices, we must have a constant offset, just
505 // evaluate it the general way.
506 if (i == e) return 0;
508 Value *VariableIdx = GEP->getOperand(i);
509 // Determine the scale factor of the variable element. For example, this is
510 // 4 if the variable index is into an array of i32.
511 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
513 // Verify that there are no other variable indices. If so, emit the hard way.
514 for (++i, ++GTI; i != e; ++i, ++GTI) {
515 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
518 // Compute the aggregate offset of constant indices.
519 if (CI->isZero()) continue;
521 // Handle a struct index, which adds its field offset to the pointer.
522 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
523 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
525 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
526 Offset += Size*CI->getSExtValue();
530 // Okay, we know we have a single variable index, which must be a
531 // pointer/array/vector index. If there is no offset, life is simple, return
533 unsigned IntPtrWidth = TD.getPointerSizeInBits();
535 // Cast to intptrty in case a truncation occurs. If an extension is needed,
536 // we don't need to bother extending: the extension won't affect where the
537 // computation crosses zero.
538 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
539 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
540 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
545 // Otherwise, there is an index. The computation we will do will be modulo
546 // the pointer size, so get it.
547 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
549 Offset &= PtrSizeMask;
550 VariableScale &= PtrSizeMask;
552 // To do this transformation, any constant index must be a multiple of the
553 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
554 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
555 // multiple of the variable scale.
556 int64_t NewOffs = Offset / (int64_t)VariableScale;
557 if (Offset != NewOffs*(int64_t)VariableScale)
560 // Okay, we can do this evaluation. Start by converting the index to intptr.
561 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
562 if (VariableIdx->getType() != IntPtrTy)
563 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
565 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
566 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
569 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
570 /// else. At this point we know that the GEP is on the LHS of the comparison.
571 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
572 ICmpInst::Predicate Cond,
574 // Look through bitcasts.
575 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
576 RHS = BCI->getOperand(0);
578 Value *PtrBase = GEPLHS->getOperand(0);
579 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
580 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
581 // This transformation (ignoring the base and scales) is valid because we
582 // know pointers can't overflow since the gep is inbounds. See if we can
583 // output an optimized form.
584 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
586 // If not, synthesize the offset the hard way.
588 Offset = EmitGEPOffset(GEPLHS);
589 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
590 Constant::getNullValue(Offset->getType()));
591 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
592 // If the base pointers are different, but the indices are the same, just
593 // compare the base pointer.
594 if (PtrBase != GEPRHS->getOperand(0)) {
595 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
596 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
597 GEPRHS->getOperand(0)->getType();
599 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
600 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
601 IndicesTheSame = false;
605 // If all indices are the same, just compare the base pointers.
607 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
608 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
610 // Otherwise, the base pointers are different and the indices are
611 // different, bail out.
615 // If one of the GEPs has all zero indices, recurse.
616 bool AllZeros = true;
617 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
618 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
619 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
624 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
625 ICmpInst::getSwappedPredicate(Cond), I);
627 // If the other GEP has all zero indices, recurse.
629 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
630 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
631 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
636 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
638 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
639 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
640 // If the GEPs only differ by one index, compare it.
641 unsigned NumDifferences = 0; // Keep track of # differences.
642 unsigned DiffOperand = 0; // The operand that differs.
643 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
644 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
645 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
646 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
647 // Irreconcilable differences.
651 if (NumDifferences++) break;
656 if (NumDifferences == 0) // SAME GEP?
657 return ReplaceInstUsesWith(I, // No comparison is needed here.
658 ConstantInt::get(Type::getInt1Ty(I.getContext()),
659 ICmpInst::isTrueWhenEqual(Cond)));
661 else if (NumDifferences == 1 && GEPsInBounds) {
662 Value *LHSV = GEPLHS->getOperand(DiffOperand);
663 Value *RHSV = GEPRHS->getOperand(DiffOperand);
664 // Make sure we do a signed comparison here.
665 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
669 // Only lower this if the icmp is the only user of the GEP or if we expect
670 // the result to fold to a constant!
673 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
674 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
675 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
676 Value *L = EmitGEPOffset(GEPLHS);
677 Value *R = EmitGEPOffset(GEPRHS);
678 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
684 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
685 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
686 Value *X, ConstantInt *CI,
687 ICmpInst::Predicate Pred,
689 // If we have X+0, exit early (simplifying logic below) and let it get folded
690 // elsewhere. icmp X+0, X -> icmp X, X
692 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
693 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
696 // (X+4) == X -> false.
697 if (Pred == ICmpInst::ICMP_EQ)
698 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
700 // (X+4) != X -> true.
701 if (Pred == ICmpInst::ICMP_NE)
702 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
704 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
705 // so the values can never be equal. Similarly for all other "or equals"
708 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
709 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
710 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
711 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
713 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
714 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
717 // (X+1) >u X --> X <u (0-1) --> X != 255
718 // (X+2) >u X --> X <u (0-2) --> X <u 254
719 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
720 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
721 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
723 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
724 ConstantInt *SMax = ConstantInt::get(X->getContext(),
725 APInt::getSignedMaxValue(BitWidth));
727 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
728 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
729 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
730 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
731 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
732 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
733 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
734 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
736 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
737 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
738 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
739 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
740 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
741 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
743 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
744 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
745 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
748 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
749 /// and CmpRHS are both known to be integer constants.
750 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
751 ConstantInt *DivRHS) {
752 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
753 const APInt &CmpRHSV = CmpRHS->getValue();
755 // FIXME: If the operand types don't match the type of the divide
756 // then don't attempt this transform. The code below doesn't have the
757 // logic to deal with a signed divide and an unsigned compare (and
758 // vice versa). This is because (x /s C1) <s C2 produces different
759 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
760 // (x /u C1) <u C2. Simply casting the operands and result won't
761 // work. :( The if statement below tests that condition and bails
763 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
764 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
766 if (DivRHS->isZero())
767 return 0; // The ProdOV computation fails on divide by zero.
768 if (DivIsSigned && DivRHS->isAllOnesValue())
769 return 0; // The overflow computation also screws up here
770 if (DivRHS->isOne()) {
771 // This eliminates some funny cases with INT_MIN.
772 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
776 // Compute Prod = CI * DivRHS. We are essentially solving an equation
777 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
778 // C2 (CI). By solving for X we can turn this into a range check
779 // instead of computing a divide.
780 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
782 // Determine if the product overflows by seeing if the product is
783 // not equal to the divide. Make sure we do the same kind of divide
784 // as in the LHS instruction that we're folding.
785 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
786 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
788 // Get the ICmp opcode
789 ICmpInst::Predicate Pred = ICI.getPredicate();
791 /// If the division is known to be exact, then there is no remainder from the
792 /// divide, so the covered range size is unit, otherwise it is the divisor.
793 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
795 // Figure out the interval that is being checked. For example, a comparison
796 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
797 // Compute this interval based on the constants involved and the signedness of
798 // the compare/divide. This computes a half-open interval, keeping track of
799 // whether either value in the interval overflows. After analysis each
800 // overflow variable is set to 0 if it's corresponding bound variable is valid
801 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
802 int LoOverflow = 0, HiOverflow = 0;
803 Constant *LoBound = 0, *HiBound = 0;
805 if (!DivIsSigned) { // udiv
806 // e.g. X/5 op 3 --> [15, 20)
808 HiOverflow = LoOverflow = ProdOV;
810 // If this is not an exact divide, then many values in the range collapse
811 // to the same result value.
812 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
815 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
816 if (CmpRHSV == 0) { // (X / pos) op 0
817 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
818 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
820 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
821 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
822 HiOverflow = LoOverflow = ProdOV;
824 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
825 } else { // (X / pos) op neg
826 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
827 HiBound = AddOne(Prod);
828 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
830 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
831 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
834 } else if (DivRHS->isNegative()) { // Divisor is < 0.
836 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
837 if (CmpRHSV == 0) { // (X / neg) op 0
838 // e.g. X/-5 op 0 --> [-4, 5)
839 LoBound = AddOne(RangeSize);
840 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
841 if (HiBound == DivRHS) { // -INTMIN = INTMIN
842 HiOverflow = 1; // [INTMIN+1, overflow)
843 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
845 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
846 // e.g. X/-5 op 3 --> [-19, -14)
847 HiBound = AddOne(Prod);
848 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
850 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
851 } else { // (X / neg) op neg
852 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
853 LoOverflow = HiOverflow = ProdOV;
855 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
858 // Dividing by a negative swaps the condition. LT <-> GT
859 Pred = ICmpInst::getSwappedPredicate(Pred);
862 Value *X = DivI->getOperand(0);
864 default: llvm_unreachable("Unhandled icmp opcode!");
865 case ICmpInst::ICMP_EQ:
866 if (LoOverflow && HiOverflow)
867 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
869 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
870 ICmpInst::ICMP_UGE, X, LoBound);
872 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
873 ICmpInst::ICMP_ULT, X, HiBound);
874 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
876 case ICmpInst::ICMP_NE:
877 if (LoOverflow && HiOverflow)
878 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
880 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
881 ICmpInst::ICMP_ULT, X, LoBound);
883 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
884 ICmpInst::ICMP_UGE, X, HiBound);
885 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
886 DivIsSigned, false));
887 case ICmpInst::ICMP_ULT:
888 case ICmpInst::ICMP_SLT:
889 if (LoOverflow == +1) // Low bound is greater than input range.
890 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
891 if (LoOverflow == -1) // Low bound is less than input range.
892 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
893 return new ICmpInst(Pred, X, LoBound);
894 case ICmpInst::ICMP_UGT:
895 case ICmpInst::ICMP_SGT:
896 if (HiOverflow == +1) // High bound greater than input range.
897 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
898 if (HiOverflow == -1) // High bound less than input range.
899 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
900 if (Pred == ICmpInst::ICMP_UGT)
901 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
902 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
906 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
907 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
908 ConstantInt *ShAmt) {
909 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
911 // Check that the shift amount is in range. If not, don't perform
912 // undefined shifts. When the shift is visited it will be
914 uint32_t TypeBits = CmpRHSV.getBitWidth();
915 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
916 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
919 if (!ICI.isEquality()) {
920 // If we have an unsigned comparison and an ashr, we can't simplify this.
921 // Similarly for signed comparisons with lshr.
922 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
925 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
926 // by a power of 2. Since we already have logic to simplify these,
927 // transform to div and then simplify the resultant comparison.
928 if (Shr->getOpcode() == Instruction::AShr &&
929 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
932 // Revisit the shift (to delete it).
936 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
939 Shr->getOpcode() == Instruction::AShr ?
940 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
941 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
943 ICI.setOperand(0, Tmp);
945 // If the builder folded the binop, just return it.
946 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
950 // Otherwise, fold this div/compare.
951 assert(TheDiv->getOpcode() == Instruction::SDiv ||
952 TheDiv->getOpcode() == Instruction::UDiv);
954 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
955 assert(Res && "This div/cst should have folded!");
960 // If we are comparing against bits always shifted out, the
961 // comparison cannot succeed.
962 APInt Comp = CmpRHSV << ShAmtVal;
963 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
964 if (Shr->getOpcode() == Instruction::LShr)
965 Comp = Comp.lshr(ShAmtVal);
967 Comp = Comp.ashr(ShAmtVal);
969 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
970 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
971 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
973 return ReplaceInstUsesWith(ICI, Cst);
976 // Otherwise, check to see if the bits shifted out are known to be zero.
977 // If so, we can compare against the unshifted value:
978 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
979 if (Shr->hasOneUse() && Shr->isExact())
980 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
982 if (Shr->hasOneUse()) {
983 // Otherwise strength reduce the shift into an and.
984 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
985 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
987 Value *And = Builder->CreateAnd(Shr->getOperand(0),
988 Mask, Shr->getName()+".mask");
989 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
995 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
997 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1000 const APInt &RHSV = RHS->getValue();
1002 switch (LHSI->getOpcode()) {
1003 case Instruction::Trunc:
1004 if (ICI.isEquality() && LHSI->hasOneUse()) {
1005 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1006 // of the high bits truncated out of x are known.
1007 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1008 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1009 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
1010 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1011 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
1013 // If all the high bits are known, we can do this xform.
1014 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1015 // Pull in the high bits from known-ones set.
1016 APInt NewRHS = RHS->getValue().zext(SrcBits);
1018 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1019 ConstantInt::get(ICI.getContext(), NewRHS));
1024 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1025 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1026 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1028 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1029 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1030 Value *CompareVal = LHSI->getOperand(0);
1032 // If the sign bit of the XorCST is not set, there is no change to
1033 // the operation, just stop using the Xor.
1034 if (!XorCST->isNegative()) {
1035 ICI.setOperand(0, CompareVal);
1040 // Was the old condition true if the operand is positive?
1041 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1043 // If so, the new one isn't.
1044 isTrueIfPositive ^= true;
1046 if (isTrueIfPositive)
1047 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1050 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1054 if (LHSI->hasOneUse()) {
1055 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1056 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1057 const APInt &SignBit = XorCST->getValue();
1058 ICmpInst::Predicate Pred = ICI.isSigned()
1059 ? ICI.getUnsignedPredicate()
1060 : ICI.getSignedPredicate();
1061 return new ICmpInst(Pred, LHSI->getOperand(0),
1062 ConstantInt::get(ICI.getContext(),
1066 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1067 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1068 const APInt &NotSignBit = XorCST->getValue();
1069 ICmpInst::Predicate Pred = ICI.isSigned()
1070 ? ICI.getUnsignedPredicate()
1071 : ICI.getSignedPredicate();
1072 Pred = ICI.getSwappedPredicate(Pred);
1073 return new ICmpInst(Pred, LHSI->getOperand(0),
1074 ConstantInt::get(ICI.getContext(),
1075 RHSV ^ NotSignBit));
1080 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1081 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1082 LHSI->getOperand(0)->hasOneUse()) {
1083 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1085 // If the LHS is an AND of a truncating cast, we can widen the
1086 // and/compare to be the input width without changing the value
1087 // produced, eliminating a cast.
1088 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1089 // We can do this transformation if either the AND constant does not
1090 // have its sign bit set or if it is an equality comparison.
1091 // Extending a relational comparison when we're checking the sign
1092 // bit would not work.
1093 if (ICI.isEquality() ||
1094 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1096 Builder->CreateAnd(Cast->getOperand(0),
1097 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1098 NewAnd->takeName(LHSI);
1099 return new ICmpInst(ICI.getPredicate(), NewAnd,
1100 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1104 // If the LHS is an AND of a zext, and we have an equality compare, we can
1105 // shrink the and/compare to the smaller type, eliminating the cast.
1106 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1107 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1108 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1109 // should fold the icmp to true/false in that case.
1110 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1112 Builder->CreateAnd(Cast->getOperand(0),
1113 ConstantExpr::getTrunc(AndCST, Ty));
1114 NewAnd->takeName(LHSI);
1115 return new ICmpInst(ICI.getPredicate(), NewAnd,
1116 ConstantExpr::getTrunc(RHS, Ty));
1120 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1121 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1122 // happens a LOT in code produced by the C front-end, for bitfield
1124 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1125 if (Shift && !Shift->isShift())
1129 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1130 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1131 Type *AndTy = AndCST->getType(); // Type of the and.
1133 // We can fold this as long as we can't shift unknown bits
1134 // into the mask. This can only happen with signed shift
1135 // rights, as they sign-extend.
1137 bool CanFold = Shift->isLogicalShift();
1139 // To test for the bad case of the signed shr, see if any
1140 // of the bits shifted in could be tested after the mask.
1141 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1142 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1144 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1145 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1146 AndCST->getValue()) == 0)
1152 if (Shift->getOpcode() == Instruction::Shl)
1153 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1155 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1157 // Check to see if we are shifting out any of the bits being
1159 if (ConstantExpr::get(Shift->getOpcode(),
1160 NewCst, ShAmt) != RHS) {
1161 // If we shifted bits out, the fold is not going to work out.
1162 // As a special case, check to see if this means that the
1163 // result is always true or false now.
1164 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1165 return ReplaceInstUsesWith(ICI,
1166 ConstantInt::getFalse(ICI.getContext()));
1167 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1168 return ReplaceInstUsesWith(ICI,
1169 ConstantInt::getTrue(ICI.getContext()));
1171 ICI.setOperand(1, NewCst);
1172 Constant *NewAndCST;
1173 if (Shift->getOpcode() == Instruction::Shl)
1174 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1176 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1177 LHSI->setOperand(1, NewAndCST);
1178 LHSI->setOperand(0, Shift->getOperand(0));
1179 Worklist.Add(Shift); // Shift is dead.
1185 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1186 // preferable because it allows the C<<Y expression to be hoisted out
1187 // of a loop if Y is invariant and X is not.
1188 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1189 ICI.isEquality() && !Shift->isArithmeticShift() &&
1190 !isa<Constant>(Shift->getOperand(0))) {
1193 if (Shift->getOpcode() == Instruction::LShr) {
1194 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1196 // Insert a logical shift.
1197 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1200 // Compute X & (C << Y).
1202 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1204 ICI.setOperand(0, NewAnd);
1209 // Try to optimize things like "A[i]&42 == 0" to index computations.
1210 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1211 if (GetElementPtrInst *GEP =
1212 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1213 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1214 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1215 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1216 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1217 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1223 case Instruction::Or: {
1224 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1227 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1228 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1229 // -> and (icmp eq P, null), (icmp eq Q, null).
1230 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1231 Constant::getNullValue(P->getType()));
1232 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1233 Constant::getNullValue(Q->getType()));
1235 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1236 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1238 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1244 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1245 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1248 uint32_t TypeBits = RHSV.getBitWidth();
1250 // Check that the shift amount is in range. If not, don't perform
1251 // undefined shifts. When the shift is visited it will be
1253 if (ShAmt->uge(TypeBits))
1256 if (ICI.isEquality()) {
1257 // If we are comparing against bits always shifted out, the
1258 // comparison cannot succeed.
1260 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1262 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1263 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1265 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1266 return ReplaceInstUsesWith(ICI, Cst);
1269 // If the shift is NUW, then it is just shifting out zeros, no need for an
1271 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1272 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1273 ConstantExpr::getLShr(RHS, ShAmt));
1275 if (LHSI->hasOneUse()) {
1276 // Otherwise strength reduce the shift into an and.
1277 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1279 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1280 TypeBits-ShAmtVal));
1283 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1284 return new ICmpInst(ICI.getPredicate(), And,
1285 ConstantExpr::getLShr(RHS, ShAmt));
1289 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1290 bool TrueIfSigned = false;
1291 if (LHSI->hasOneUse() &&
1292 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1293 // (X << 31) <s 0 --> (X&1) != 0
1294 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1295 APInt::getOneBitSet(TypeBits,
1296 TypeBits-ShAmt->getZExtValue()-1));
1298 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1299 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1300 And, Constant::getNullValue(And->getType()));
1305 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1306 case Instruction::AShr: {
1307 // Handle equality comparisons of shift-by-constant.
1308 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1309 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1310 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1314 // Handle exact shr's.
1315 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1316 if (RHSV.isMinValue())
1317 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1322 case Instruction::SDiv:
1323 case Instruction::UDiv:
1324 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1325 // Fold this div into the comparison, producing a range check.
1326 // Determine, based on the divide type, what the range is being
1327 // checked. If there is an overflow on the low or high side, remember
1328 // it, otherwise compute the range [low, hi) bounding the new value.
1329 // See: InsertRangeTest above for the kinds of replacements possible.
1330 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1331 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1336 case Instruction::Add:
1337 // Fold: icmp pred (add X, C1), C2
1338 if (!ICI.isEquality()) {
1339 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1341 const APInt &LHSV = LHSC->getValue();
1343 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1346 if (ICI.isSigned()) {
1347 if (CR.getLower().isSignBit()) {
1348 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1349 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1350 } else if (CR.getUpper().isSignBit()) {
1351 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1352 ConstantInt::get(ICI.getContext(),CR.getLower()));
1355 if (CR.getLower().isMinValue()) {
1356 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1357 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1358 } else if (CR.getUpper().isMinValue()) {
1359 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1360 ConstantInt::get(ICI.getContext(),CR.getLower()));
1367 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1368 if (ICI.isEquality()) {
1369 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1371 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1372 // the second operand is a constant, simplify a bit.
1373 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1374 switch (BO->getOpcode()) {
1375 case Instruction::SRem:
1376 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1377 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1378 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1379 if (V.sgt(1) && V.isPowerOf2()) {
1381 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1383 return new ICmpInst(ICI.getPredicate(), NewRem,
1384 Constant::getNullValue(BO->getType()));
1388 case Instruction::Add:
1389 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1390 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1391 if (BO->hasOneUse())
1392 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1393 ConstantExpr::getSub(RHS, BOp1C));
1394 } else if (RHSV == 0) {
1395 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1396 // efficiently invertible, or if the add has just this one use.
1397 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1399 if (Value *NegVal = dyn_castNegVal(BOp1))
1400 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1401 if (Value *NegVal = dyn_castNegVal(BOp0))
1402 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1403 if (BO->hasOneUse()) {
1404 Value *Neg = Builder->CreateNeg(BOp1);
1406 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1410 case Instruction::Xor:
1411 // For the xor case, we can xor two constants together, eliminating
1412 // the explicit xor.
1413 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1414 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1415 ConstantExpr::getXor(RHS, BOC));
1416 } else if (RHSV == 0) {
1417 // Replace ((xor A, B) != 0) with (A != B)
1418 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1422 case Instruction::Sub:
1423 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1424 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1425 if (BO->hasOneUse())
1426 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1427 ConstantExpr::getSub(BOp0C, RHS));
1428 } else if (RHSV == 0) {
1429 // Replace ((sub A, B) != 0) with (A != B)
1430 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1434 case Instruction::Or:
1435 // If bits are being or'd in that are not present in the constant we
1436 // are comparing against, then the comparison could never succeed!
1437 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1438 Constant *NotCI = ConstantExpr::getNot(RHS);
1439 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1440 return ReplaceInstUsesWith(ICI,
1441 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1446 case Instruction::And:
1447 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1448 // If bits are being compared against that are and'd out, then the
1449 // comparison can never succeed!
1450 if ((RHSV & ~BOC->getValue()) != 0)
1451 return ReplaceInstUsesWith(ICI,
1452 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1455 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1456 if (RHS == BOC && RHSV.isPowerOf2())
1457 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1458 ICmpInst::ICMP_NE, LHSI,
1459 Constant::getNullValue(RHS->getType()));
1461 // Don't perform the following transforms if the AND has multiple uses
1462 if (!BO->hasOneUse())
1465 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1466 if (BOC->getValue().isSignBit()) {
1467 Value *X = BO->getOperand(0);
1468 Constant *Zero = Constant::getNullValue(X->getType());
1469 ICmpInst::Predicate pred = isICMP_NE ?
1470 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1471 return new ICmpInst(pred, X, Zero);
1474 // ((X & ~7) == 0) --> X < 8
1475 if (RHSV == 0 && isHighOnes(BOC)) {
1476 Value *X = BO->getOperand(0);
1477 Constant *NegX = ConstantExpr::getNeg(BOC);
1478 ICmpInst::Predicate pred = isICMP_NE ?
1479 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1480 return new ICmpInst(pred, X, NegX);
1485 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1486 // Handle icmp {eq|ne} <intrinsic>, intcst.
1487 switch (II->getIntrinsicID()) {
1488 case Intrinsic::bswap:
1490 ICI.setOperand(0, II->getArgOperand(0));
1491 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1493 case Intrinsic::ctlz:
1494 case Intrinsic::cttz:
1495 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1496 if (RHSV == RHS->getType()->getBitWidth()) {
1498 ICI.setOperand(0, II->getArgOperand(0));
1499 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1503 case Intrinsic::ctpop:
1504 // popcount(A) == 0 -> A == 0 and likewise for !=
1505 if (RHS->isZero()) {
1507 ICI.setOperand(0, II->getArgOperand(0));
1508 ICI.setOperand(1, RHS);
1520 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1521 /// We only handle extending casts so far.
1523 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1524 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1525 Value *LHSCIOp = LHSCI->getOperand(0);
1526 Type *SrcTy = LHSCIOp->getType();
1527 Type *DestTy = LHSCI->getType();
1530 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1531 // integer type is the same size as the pointer type.
1532 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1533 TD->getPointerSizeInBits() ==
1534 cast<IntegerType>(DestTy)->getBitWidth()) {
1536 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1537 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1538 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1539 RHSOp = RHSC->getOperand(0);
1540 // If the pointer types don't match, insert a bitcast.
1541 if (LHSCIOp->getType() != RHSOp->getType())
1542 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1546 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1549 // The code below only handles extension cast instructions, so far.
1551 if (LHSCI->getOpcode() != Instruction::ZExt &&
1552 LHSCI->getOpcode() != Instruction::SExt)
1555 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1556 bool isSignedCmp = ICI.isSigned();
1558 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1559 // Not an extension from the same type?
1560 RHSCIOp = CI->getOperand(0);
1561 if (RHSCIOp->getType() != LHSCIOp->getType())
1564 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1565 // and the other is a zext), then we can't handle this.
1566 if (CI->getOpcode() != LHSCI->getOpcode())
1569 // Deal with equality cases early.
1570 if (ICI.isEquality())
1571 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1573 // A signed comparison of sign extended values simplifies into a
1574 // signed comparison.
1575 if (isSignedCmp && isSignedExt)
1576 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1578 // The other three cases all fold into an unsigned comparison.
1579 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1582 // If we aren't dealing with a constant on the RHS, exit early
1583 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1587 // Compute the constant that would happen if we truncated to SrcTy then
1588 // reextended to DestTy.
1589 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1590 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1593 // If the re-extended constant didn't change...
1595 // Deal with equality cases early.
1596 if (ICI.isEquality())
1597 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1599 // A signed comparison of sign extended values simplifies into a
1600 // signed comparison.
1601 if (isSignedExt && isSignedCmp)
1602 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1604 // The other three cases all fold into an unsigned comparison.
1605 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1608 // The re-extended constant changed so the constant cannot be represented
1609 // in the shorter type. Consequently, we cannot emit a simple comparison.
1610 // All the cases that fold to true or false will have already been handled
1611 // by SimplifyICmpInst, so only deal with the tricky case.
1613 if (isSignedCmp || !isSignedExt)
1616 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1617 // should have been folded away previously and not enter in here.
1619 // We're performing an unsigned comp with a sign extended value.
1620 // This is true if the input is >= 0. [aka >s -1]
1621 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1622 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1624 // Finally, return the value computed.
1625 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1626 return ReplaceInstUsesWith(ICI, Result);
1628 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1629 return BinaryOperator::CreateNot(Result);
1632 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1633 /// I = icmp ugt (add (add A, B), CI2), CI1
1634 /// If this is of the form:
1636 /// if (sum+128 >u 255)
1637 /// Then replace it with llvm.sadd.with.overflow.i8.
1639 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1640 ConstantInt *CI2, ConstantInt *CI1,
1642 // The transformation we're trying to do here is to transform this into an
1643 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1644 // with a narrower add, and discard the add-with-constant that is part of the
1645 // range check (if we can't eliminate it, this isn't profitable).
1647 // In order to eliminate the add-with-constant, the compare can be its only
1649 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1650 if (!AddWithCst->hasOneUse()) return 0;
1652 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1653 if (!CI2->getValue().isPowerOf2()) return 0;
1654 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1655 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1657 // The width of the new add formed is 1 more than the bias.
1660 // Check to see that CI1 is an all-ones value with NewWidth bits.
1661 if (CI1->getBitWidth() == NewWidth ||
1662 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1665 // This is only really a signed overflow check if the inputs have been
1666 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1667 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1668 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1669 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1670 IC.ComputeNumSignBits(B) < NeededSignBits)
1673 // In order to replace the original add with a narrower
1674 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1675 // and truncates that discard the high bits of the add. Verify that this is
1677 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1678 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1680 if (*UI == AddWithCst) continue;
1682 // Only accept truncates for now. We would really like a nice recursive
1683 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1684 // chain to see which bits of a value are actually demanded. If the
1685 // original add had another add which was then immediately truncated, we
1686 // could still do the transformation.
1687 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1689 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1692 // If the pattern matches, truncate the inputs to the narrower type and
1693 // use the sadd_with_overflow intrinsic to efficiently compute both the
1694 // result and the overflow bit.
1695 Module *M = I.getParent()->getParent()->getParent();
1697 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1698 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1701 InstCombiner::BuilderTy *Builder = IC.Builder;
1703 // Put the new code above the original add, in case there are any uses of the
1704 // add between the add and the compare.
1705 Builder->SetInsertPoint(OrigAdd);
1707 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1708 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1709 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1710 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1711 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1713 // The inner add was the result of the narrow add, zero extended to the
1714 // wider type. Replace it with the result computed by the intrinsic.
1715 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1717 // The original icmp gets replaced with the overflow value.
1718 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1721 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1723 // Don't bother doing this transformation for pointers, don't do it for
1725 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1727 // If the add is a constant expr, then we don't bother transforming it.
1728 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1729 if (OrigAdd == 0) return 0;
1731 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1733 // Put the new code above the original add, in case there are any uses of the
1734 // add between the add and the compare.
1735 InstCombiner::BuilderTy *Builder = IC.Builder;
1736 Builder->SetInsertPoint(OrigAdd);
1738 Module *M = I.getParent()->getParent()->getParent();
1739 Type *Ty = LHS->getType();
1740 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1741 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1742 Value *Add = Builder->CreateExtractValue(Call, 0);
1744 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1746 // The original icmp gets replaced with the overflow value.
1747 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1750 // DemandedBitsLHSMask - When performing a comparison against a constant,
1751 // it is possible that not all the bits in the LHS are demanded. This helper
1752 // method computes the mask that IS demanded.
1753 static APInt DemandedBitsLHSMask(ICmpInst &I,
1754 unsigned BitWidth, bool isSignCheck) {
1756 return APInt::getSignBit(BitWidth);
1758 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1759 if (!CI) return APInt::getAllOnesValue(BitWidth);
1760 const APInt &RHS = CI->getValue();
1762 switch (I.getPredicate()) {
1763 // For a UGT comparison, we don't care about any bits that
1764 // correspond to the trailing ones of the comparand. The value of these
1765 // bits doesn't impact the outcome of the comparison, because any value
1766 // greater than the RHS must differ in a bit higher than these due to carry.
1767 case ICmpInst::ICMP_UGT: {
1768 unsigned trailingOnes = RHS.countTrailingOnes();
1769 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1773 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1774 // Any value less than the RHS must differ in a higher bit because of carries.
1775 case ICmpInst::ICMP_ULT: {
1776 unsigned trailingZeros = RHS.countTrailingZeros();
1777 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1782 return APInt::getAllOnesValue(BitWidth);
1787 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1788 bool Changed = false;
1789 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1791 /// Orders the operands of the compare so that they are listed from most
1792 /// complex to least complex. This puts constants before unary operators,
1793 /// before binary operators.
1794 if (getComplexity(Op0) < getComplexity(Op1)) {
1796 std::swap(Op0, Op1);
1800 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1801 return ReplaceInstUsesWith(I, V);
1803 // comparing -val or val with non-zero is the same as just comparing val
1804 // ie, abs(val) != 0 -> val != 0
1805 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
1807 Value *Cond, *SelectTrue, *SelectFalse;
1808 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
1809 m_Value(SelectFalse)))) {
1810 if (Value *V = dyn_castNegVal(SelectTrue)) {
1811 if (V == SelectFalse)
1812 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1814 else if (Value *V = dyn_castNegVal(SelectFalse)) {
1815 if (V == SelectTrue)
1816 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1821 Type *Ty = Op0->getType();
1823 // icmp's with boolean values can always be turned into bitwise operations
1824 if (Ty->isIntegerTy(1)) {
1825 switch (I.getPredicate()) {
1826 default: llvm_unreachable("Invalid icmp instruction!");
1827 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1828 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1829 return BinaryOperator::CreateNot(Xor);
1831 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1832 return BinaryOperator::CreateXor(Op0, Op1);
1834 case ICmpInst::ICMP_UGT:
1835 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1837 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1838 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1839 return BinaryOperator::CreateAnd(Not, Op1);
1841 case ICmpInst::ICMP_SGT:
1842 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1844 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1845 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1846 return BinaryOperator::CreateAnd(Not, Op0);
1848 case ICmpInst::ICMP_UGE:
1849 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1851 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1852 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1853 return BinaryOperator::CreateOr(Not, Op1);
1855 case ICmpInst::ICMP_SGE:
1856 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1858 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1859 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1860 return BinaryOperator::CreateOr(Not, Op0);
1865 unsigned BitWidth = 0;
1866 if (Ty->isIntOrIntVectorTy())
1867 BitWidth = Ty->getScalarSizeInBits();
1868 else if (TD) // Pointers require TD info to get their size.
1869 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1871 bool isSignBit = false;
1873 // See if we are doing a comparison with a constant.
1874 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1875 Value *A = 0, *B = 0;
1877 // Match the following pattern, which is a common idiom when writing
1878 // overflow-safe integer arithmetic function. The source performs an
1879 // addition in wider type, and explicitly checks for overflow using
1880 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1881 // sadd_with_overflow intrinsic.
1883 // TODO: This could probably be generalized to handle other overflow-safe
1884 // operations if we worked out the formulas to compute the appropriate
1888 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1890 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1891 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1892 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1893 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1897 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1898 if (I.isEquality() && CI->isZero() &&
1899 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1900 // (icmp cond A B) if cond is equality
1901 return new ICmpInst(I.getPredicate(), A, B);
1904 // If we have an icmp le or icmp ge instruction, turn it into the
1905 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1906 // them being folded in the code below. The SimplifyICmpInst code has
1907 // already handled the edge cases for us, so we just assert on them.
1908 switch (I.getPredicate()) {
1910 case ICmpInst::ICMP_ULE:
1911 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1912 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1913 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1914 case ICmpInst::ICMP_SLE:
1915 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1916 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1917 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1918 case ICmpInst::ICMP_UGE:
1919 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1920 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1921 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1922 case ICmpInst::ICMP_SGE:
1923 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1924 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1925 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1928 // If this comparison is a normal comparison, it demands all
1929 // bits, if it is a sign bit comparison, it only demands the sign bit.
1931 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1934 // See if we can fold the comparison based on range information we can get
1935 // by checking whether bits are known to be zero or one in the input.
1936 if (BitWidth != 0) {
1937 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1938 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1940 if (SimplifyDemandedBits(I.getOperandUse(0),
1941 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1942 Op0KnownZero, Op0KnownOne, 0))
1944 if (SimplifyDemandedBits(I.getOperandUse(1),
1945 APInt::getAllOnesValue(BitWidth),
1946 Op1KnownZero, Op1KnownOne, 0))
1949 // Given the known and unknown bits, compute a range that the LHS could be
1950 // in. Compute the Min, Max and RHS values based on the known bits. For the
1951 // EQ and NE we use unsigned values.
1952 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1953 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1955 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1957 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1960 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1962 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1966 // If Min and Max are known to be the same, then SimplifyDemandedBits
1967 // figured out that the LHS is a constant. Just constant fold this now so
1968 // that code below can assume that Min != Max.
1969 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1970 return new ICmpInst(I.getPredicate(),
1971 ConstantInt::get(Op0->getType(), Op0Min), Op1);
1972 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1973 return new ICmpInst(I.getPredicate(), Op0,
1974 ConstantInt::get(Op1->getType(), Op1Min));
1976 // Based on the range information we know about the LHS, see if we can
1977 // simplify this comparison. For example, (x&4) < 8 is always true.
1978 switch (I.getPredicate()) {
1979 default: llvm_unreachable("Unknown icmp opcode!");
1980 case ICmpInst::ICMP_EQ: {
1981 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1982 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
1984 // If all bits are known zero except for one, then we know at most one
1985 // bit is set. If the comparison is against zero, then this is a check
1986 // to see if *that* bit is set.
1987 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1988 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1989 // If the LHS is an AND with the same constant, look through it.
1991 ConstantInt *LHSC = 0;
1992 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1993 LHSC->getValue() != Op0KnownZeroInverted)
1996 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1997 // then turn "((1 << x)&8) == 0" into "x != 3".
1999 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2000 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2001 return new ICmpInst(ICmpInst::ICMP_NE, X,
2002 ConstantInt::get(X->getType(), CmpVal));
2005 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2006 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2008 if (Op0KnownZeroInverted == 1 &&
2009 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2010 return new ICmpInst(ICmpInst::ICMP_NE, X,
2011 ConstantInt::get(X->getType(),
2012 CI->countTrailingZeros()));
2017 case ICmpInst::ICMP_NE: {
2018 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2019 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2021 // If all bits are known zero except for one, then we know at most one
2022 // bit is set. If the comparison is against zero, then this is a check
2023 // to see if *that* bit is set.
2024 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2025 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2026 // If the LHS is an AND with the same constant, look through it.
2028 ConstantInt *LHSC = 0;
2029 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2030 LHSC->getValue() != Op0KnownZeroInverted)
2033 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2034 // then turn "((1 << x)&8) != 0" into "x == 3".
2036 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2037 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2038 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2039 ConstantInt::get(X->getType(), CmpVal));
2042 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2043 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2045 if (Op0KnownZeroInverted == 1 &&
2046 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2047 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2048 ConstantInt::get(X->getType(),
2049 CI->countTrailingZeros()));
2054 case ICmpInst::ICMP_ULT:
2055 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2056 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2057 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2058 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2059 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2060 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2061 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2062 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2063 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2064 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2066 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2067 if (CI->isMinValue(true))
2068 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2069 Constant::getAllOnesValue(Op0->getType()));
2072 case ICmpInst::ICMP_UGT:
2073 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2074 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2075 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2076 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2078 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2079 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2080 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2081 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2082 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2083 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2085 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2086 if (CI->isMaxValue(true))
2087 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2088 Constant::getNullValue(Op0->getType()));
2091 case ICmpInst::ICMP_SLT:
2092 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2093 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2094 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2095 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2096 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2097 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2098 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2099 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2100 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2101 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2104 case ICmpInst::ICMP_SGT:
2105 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2106 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2107 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2108 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2110 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2111 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2112 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2113 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2114 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2115 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2118 case ICmpInst::ICMP_SGE:
2119 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2120 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2121 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2122 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2123 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2125 case ICmpInst::ICMP_SLE:
2126 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2127 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2128 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2129 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2130 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2132 case ICmpInst::ICMP_UGE:
2133 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2134 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2135 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2136 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2137 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2139 case ICmpInst::ICMP_ULE:
2140 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2141 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2142 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2143 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2144 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2148 // Turn a signed comparison into an unsigned one if both operands
2149 // are known to have the same sign.
2151 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2152 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2153 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2156 // Test if the ICmpInst instruction is used exclusively by a select as
2157 // part of a minimum or maximum operation. If so, refrain from doing
2158 // any other folding. This helps out other analyses which understand
2159 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2160 // and CodeGen. And in this case, at least one of the comparison
2161 // operands has at least one user besides the compare (the select),
2162 // which would often largely negate the benefit of folding anyway.
2164 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2165 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2166 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2169 // See if we are doing a comparison between a constant and an instruction that
2170 // can be folded into the comparison.
2171 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2172 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2173 // instruction, see if that instruction also has constants so that the
2174 // instruction can be folded into the icmp
2175 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2176 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2180 // Handle icmp with constant (but not simple integer constant) RHS
2181 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2182 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2183 switch (LHSI->getOpcode()) {
2184 case Instruction::GetElementPtr:
2185 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2186 if (RHSC->isNullValue() &&
2187 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2188 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2189 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2191 case Instruction::PHI:
2192 // Only fold icmp into the PHI if the phi and icmp are in the same
2193 // block. If in the same block, we're encouraging jump threading. If
2194 // not, we are just pessimizing the code by making an i1 phi.
2195 if (LHSI->getParent() == I.getParent())
2196 if (Instruction *NV = FoldOpIntoPhi(I))
2199 case Instruction::Select: {
2200 // If either operand of the select is a constant, we can fold the
2201 // comparison into the select arms, which will cause one to be
2202 // constant folded and the select turned into a bitwise or.
2203 Value *Op1 = 0, *Op2 = 0;
2204 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2205 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2206 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2207 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2209 // We only want to perform this transformation if it will not lead to
2210 // additional code. This is true if either both sides of the select
2211 // fold to a constant (in which case the icmp is replaced with a select
2212 // which will usually simplify) or this is the only user of the
2213 // select (in which case we are trading a select+icmp for a simpler
2215 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2217 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2220 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2222 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2226 case Instruction::IntToPtr:
2227 // icmp pred inttoptr(X), null -> icmp pred X, 0
2228 if (RHSC->isNullValue() && TD &&
2229 TD->getIntPtrType(RHSC->getContext()) ==
2230 LHSI->getOperand(0)->getType())
2231 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2232 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2235 case Instruction::Load:
2236 // Try to optimize things like "A[i] > 4" to index computations.
2237 if (GetElementPtrInst *GEP =
2238 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2239 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2240 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2241 !cast<LoadInst>(LHSI)->isVolatile())
2242 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2249 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2250 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2251 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2253 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2254 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2255 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2258 // Test to see if the operands of the icmp are casted versions of other
2259 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2261 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2262 if (Op0->getType()->isPointerTy() &&
2263 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2264 // We keep moving the cast from the left operand over to the right
2265 // operand, where it can often be eliminated completely.
2266 Op0 = CI->getOperand(0);
2268 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2269 // so eliminate it as well.
2270 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2271 Op1 = CI2->getOperand(0);
2273 // If Op1 is a constant, we can fold the cast into the constant.
2274 if (Op0->getType() != Op1->getType()) {
2275 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2276 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2278 // Otherwise, cast the RHS right before the icmp
2279 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2282 return new ICmpInst(I.getPredicate(), Op0, Op1);
2286 if (isa<CastInst>(Op0)) {
2287 // Handle the special case of: icmp (cast bool to X), <cst>
2288 // This comes up when you have code like
2291 // For generality, we handle any zero-extension of any operand comparison
2292 // with a constant or another cast from the same type.
2293 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2294 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2298 // Special logic for binary operators.
2299 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2300 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2302 CmpInst::Predicate Pred = I.getPredicate();
2303 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2304 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2305 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2306 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2307 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2308 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2309 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2310 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2311 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2313 // Analyze the case when either Op0 or Op1 is an add instruction.
2314 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2315 Value *A = 0, *B = 0, *C = 0, *D = 0;
2316 if (BO0 && BO0->getOpcode() == Instruction::Add)
2317 A = BO0->getOperand(0), B = BO0->getOperand(1);
2318 if (BO1 && BO1->getOpcode() == Instruction::Add)
2319 C = BO1->getOperand(0), D = BO1->getOperand(1);
2321 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2322 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2323 return new ICmpInst(Pred, A == Op1 ? B : A,
2324 Constant::getNullValue(Op1->getType()));
2326 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2327 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2328 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2331 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2332 if (A && C && (A == C || A == D || B == C || B == D) &&
2333 NoOp0WrapProblem && NoOp1WrapProblem &&
2334 // Try not to increase register pressure.
2335 BO0->hasOneUse() && BO1->hasOneUse()) {
2336 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2337 Value *Y = (A == C || A == D) ? B : A;
2338 Value *Z = (C == A || C == B) ? D : C;
2339 return new ICmpInst(Pred, Y, Z);
2342 // Analyze the case when either Op0 or Op1 is a sub instruction.
2343 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2344 A = 0; B = 0; C = 0; D = 0;
2345 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2346 A = BO0->getOperand(0), B = BO0->getOperand(1);
2347 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2348 C = BO1->getOperand(0), D = BO1->getOperand(1);
2350 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2351 if (A == Op1 && NoOp0WrapProblem)
2352 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2354 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2355 if (C == Op0 && NoOp1WrapProblem)
2356 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2358 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2359 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2360 // Try not to increase register pressure.
2361 BO0->hasOneUse() && BO1->hasOneUse())
2362 return new ICmpInst(Pred, A, C);
2364 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2365 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2366 // Try not to increase register pressure.
2367 BO0->hasOneUse() && BO1->hasOneUse())
2368 return new ICmpInst(Pred, D, B);
2370 BinaryOperator *SRem = NULL;
2371 // icmp (srem X, Y), Y
2372 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2373 Op1 == BO0->getOperand(1))
2375 // icmp Y, (srem X, Y)
2376 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2377 Op0 == BO1->getOperand(1))
2380 // We don't check hasOneUse to avoid increasing register pressure because
2381 // the value we use is the same value this instruction was already using.
2382 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2384 case ICmpInst::ICMP_EQ:
2385 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2386 case ICmpInst::ICMP_NE:
2387 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2388 case ICmpInst::ICMP_SGT:
2389 case ICmpInst::ICMP_SGE:
2390 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2391 Constant::getAllOnesValue(SRem->getType()));
2392 case ICmpInst::ICMP_SLT:
2393 case ICmpInst::ICMP_SLE:
2394 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2395 Constant::getNullValue(SRem->getType()));
2399 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2400 BO0->hasOneUse() && BO1->hasOneUse() &&
2401 BO0->getOperand(1) == BO1->getOperand(1)) {
2402 switch (BO0->getOpcode()) {
2404 case Instruction::Add:
2405 case Instruction::Sub:
2406 case Instruction::Xor:
2407 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2408 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2409 BO1->getOperand(0));
2410 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2411 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2412 if (CI->getValue().isSignBit()) {
2413 ICmpInst::Predicate Pred = I.isSigned()
2414 ? I.getUnsignedPredicate()
2415 : I.getSignedPredicate();
2416 return new ICmpInst(Pred, BO0->getOperand(0),
2417 BO1->getOperand(0));
2420 if (CI->isMaxValue(true)) {
2421 ICmpInst::Predicate Pred = I.isSigned()
2422 ? I.getUnsignedPredicate()
2423 : I.getSignedPredicate();
2424 Pred = I.getSwappedPredicate(Pred);
2425 return new ICmpInst(Pred, BO0->getOperand(0),
2426 BO1->getOperand(0));
2430 case Instruction::Mul:
2431 if (!I.isEquality())
2434 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2435 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2436 // Mask = -1 >> count-trailing-zeros(Cst).
2437 if (!CI->isZero() && !CI->isOne()) {
2438 const APInt &AP = CI->getValue();
2439 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2440 APInt::getLowBitsSet(AP.getBitWidth(),
2442 AP.countTrailingZeros()));
2443 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2444 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2445 return new ICmpInst(I.getPredicate(), And1, And2);
2449 case Instruction::UDiv:
2450 case Instruction::LShr:
2454 case Instruction::SDiv:
2455 case Instruction::AShr:
2456 if (!BO0->isExact() || !BO1->isExact())
2458 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2459 BO1->getOperand(0));
2460 case Instruction::Shl: {
2461 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2462 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2465 if (!NSW && I.isSigned())
2467 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2468 BO1->getOperand(0));
2475 // ~x < ~y --> y < x
2476 // ~x < cst --> ~cst < x
2477 if (match(Op0, m_Not(m_Value(A)))) {
2478 if (match(Op1, m_Not(m_Value(B))))
2479 return new ICmpInst(I.getPredicate(), B, A);
2480 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2481 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2484 // (a+b) <u a --> llvm.uadd.with.overflow.
2485 // (a+b) <u b --> llvm.uadd.with.overflow.
2486 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2487 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2488 (Op1 == A || Op1 == B))
2489 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2492 // a >u (a+b) --> llvm.uadd.with.overflow.
2493 // b >u (a+b) --> llvm.uadd.with.overflow.
2494 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2495 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2496 (Op0 == A || Op0 == B))
2497 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2501 if (I.isEquality()) {
2502 Value *A, *B, *C, *D;
2504 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2505 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2506 Value *OtherVal = A == Op1 ? B : A;
2507 return new ICmpInst(I.getPredicate(), OtherVal,
2508 Constant::getNullValue(A->getType()));
2511 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2512 // A^c1 == C^c2 --> A == C^(c1^c2)
2513 ConstantInt *C1, *C2;
2514 if (match(B, m_ConstantInt(C1)) &&
2515 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2516 Constant *NC = ConstantInt::get(I.getContext(),
2517 C1->getValue() ^ C2->getValue());
2518 Value *Xor = Builder->CreateXor(C, NC);
2519 return new ICmpInst(I.getPredicate(), A, Xor);
2522 // A^B == A^D -> B == D
2523 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2524 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2525 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2526 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2530 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2531 (A == Op0 || B == Op0)) {
2532 // A == (A^B) -> B == 0
2533 Value *OtherVal = A == Op0 ? B : A;
2534 return new ICmpInst(I.getPredicate(), OtherVal,
2535 Constant::getNullValue(A->getType()));
2538 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2539 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2540 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2541 Value *X = 0, *Y = 0, *Z = 0;
2544 X = B; Y = D; Z = A;
2545 } else if (A == D) {
2546 X = B; Y = C; Z = A;
2547 } else if (B == C) {
2548 X = A; Y = D; Z = B;
2549 } else if (B == D) {
2550 X = A; Y = C; Z = B;
2553 if (X) { // Build (X^Y) & Z
2554 Op1 = Builder->CreateXor(X, Y);
2555 Op1 = Builder->CreateAnd(Op1, Z);
2556 I.setOperand(0, Op1);
2557 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2562 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2563 // "icmp (and X, mask), cst"
2566 if (Op0->hasOneUse() &&
2567 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2568 m_ConstantInt(ShAmt))))) &&
2569 match(Op1, m_ConstantInt(Cst1)) &&
2570 // Only do this when A has multiple uses. This is most important to do
2571 // when it exposes other optimizations.
2573 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2575 if (ShAmt < ASize) {
2577 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2580 APInt CmpV = Cst1->getValue().zext(ASize);
2583 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2584 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2590 Value *X; ConstantInt *Cst;
2592 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2593 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2596 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2597 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2599 return Changed ? &I : 0;
2607 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2609 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2612 if (!isa<ConstantFP>(RHSC)) return 0;
2613 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2615 // Get the width of the mantissa. We don't want to hack on conversions that
2616 // might lose information from the integer, e.g. "i64 -> float"
2617 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2618 if (MantissaWidth == -1) return 0; // Unknown.
2620 // Check to see that the input is converted from an integer type that is small
2621 // enough that preserves all bits. TODO: check here for "known" sign bits.
2622 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2623 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2625 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2626 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2630 // If the conversion would lose info, don't hack on this.
2631 if ((int)InputSize > MantissaWidth)
2634 // Otherwise, we can potentially simplify the comparison. We know that it
2635 // will always come through as an integer value and we know the constant is
2636 // not a NAN (it would have been previously simplified).
2637 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2639 ICmpInst::Predicate Pred;
2640 switch (I.getPredicate()) {
2641 default: llvm_unreachable("Unexpected predicate!");
2642 case FCmpInst::FCMP_UEQ:
2643 case FCmpInst::FCMP_OEQ:
2644 Pred = ICmpInst::ICMP_EQ;
2646 case FCmpInst::FCMP_UGT:
2647 case FCmpInst::FCMP_OGT:
2648 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2650 case FCmpInst::FCMP_UGE:
2651 case FCmpInst::FCMP_OGE:
2652 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2654 case FCmpInst::FCMP_ULT:
2655 case FCmpInst::FCMP_OLT:
2656 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2658 case FCmpInst::FCMP_ULE:
2659 case FCmpInst::FCMP_OLE:
2660 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2662 case FCmpInst::FCMP_UNE:
2663 case FCmpInst::FCMP_ONE:
2664 Pred = ICmpInst::ICMP_NE;
2666 case FCmpInst::FCMP_ORD:
2667 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2668 case FCmpInst::FCMP_UNO:
2669 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2672 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2674 // Now we know that the APFloat is a normal number, zero or inf.
2676 // See if the FP constant is too large for the integer. For example,
2677 // comparing an i8 to 300.0.
2678 unsigned IntWidth = IntTy->getScalarSizeInBits();
2681 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2682 // and large values.
2683 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2684 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2685 APFloat::rmNearestTiesToEven);
2686 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2687 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2688 Pred == ICmpInst::ICMP_SLE)
2689 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2690 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2693 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2694 // +INF and large values.
2695 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2696 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2697 APFloat::rmNearestTiesToEven);
2698 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2699 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2700 Pred == ICmpInst::ICMP_ULE)
2701 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2702 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2707 // See if the RHS value is < SignedMin.
2708 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2709 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2710 APFloat::rmNearestTiesToEven);
2711 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2712 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2713 Pred == ICmpInst::ICMP_SGE)
2714 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2715 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2719 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2720 // [0, UMAX], but it may still be fractional. See if it is fractional by
2721 // casting the FP value to the integer value and back, checking for equality.
2722 // Don't do this for zero, because -0.0 is not fractional.
2723 Constant *RHSInt = LHSUnsigned
2724 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2725 : ConstantExpr::getFPToSI(RHSC, IntTy);
2726 if (!RHS.isZero()) {
2727 bool Equal = LHSUnsigned
2728 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2729 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2731 // If we had a comparison against a fractional value, we have to adjust
2732 // the compare predicate and sometimes the value. RHSC is rounded towards
2733 // zero at this point.
2735 default: llvm_unreachable("Unexpected integer comparison!");
2736 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2737 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2738 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2739 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2740 case ICmpInst::ICMP_ULE:
2741 // (float)int <= 4.4 --> int <= 4
2742 // (float)int <= -4.4 --> false
2743 if (RHS.isNegative())
2744 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2746 case ICmpInst::ICMP_SLE:
2747 // (float)int <= 4.4 --> int <= 4
2748 // (float)int <= -4.4 --> int < -4
2749 if (RHS.isNegative())
2750 Pred = ICmpInst::ICMP_SLT;
2752 case ICmpInst::ICMP_ULT:
2753 // (float)int < -4.4 --> false
2754 // (float)int < 4.4 --> int <= 4
2755 if (RHS.isNegative())
2756 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2757 Pred = ICmpInst::ICMP_ULE;
2759 case ICmpInst::ICMP_SLT:
2760 // (float)int < -4.4 --> int < -4
2761 // (float)int < 4.4 --> int <= 4
2762 if (!RHS.isNegative())
2763 Pred = ICmpInst::ICMP_SLE;
2765 case ICmpInst::ICMP_UGT:
2766 // (float)int > 4.4 --> int > 4
2767 // (float)int > -4.4 --> true
2768 if (RHS.isNegative())
2769 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2771 case ICmpInst::ICMP_SGT:
2772 // (float)int > 4.4 --> int > 4
2773 // (float)int > -4.4 --> int >= -4
2774 if (RHS.isNegative())
2775 Pred = ICmpInst::ICMP_SGE;
2777 case ICmpInst::ICMP_UGE:
2778 // (float)int >= -4.4 --> true
2779 // (float)int >= 4.4 --> int > 4
2780 if (!RHS.isNegative())
2781 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2782 Pred = ICmpInst::ICMP_UGT;
2784 case ICmpInst::ICMP_SGE:
2785 // (float)int >= -4.4 --> int >= -4
2786 // (float)int >= 4.4 --> int > 4
2787 if (!RHS.isNegative())
2788 Pred = ICmpInst::ICMP_SGT;
2794 // Lower this FP comparison into an appropriate integer version of the
2796 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2799 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2800 bool Changed = false;
2802 /// Orders the operands of the compare so that they are listed from most
2803 /// complex to least complex. This puts constants before unary operators,
2804 /// before binary operators.
2805 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2810 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2812 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2813 return ReplaceInstUsesWith(I, V);
2815 // Simplify 'fcmp pred X, X'
2817 switch (I.getPredicate()) {
2818 default: llvm_unreachable("Unknown predicate!");
2819 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2820 case FCmpInst::FCMP_ULT: // True if unordered or less than
2821 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2822 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2823 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2824 I.setPredicate(FCmpInst::FCMP_UNO);
2825 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2828 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2829 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2830 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2831 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2832 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2833 I.setPredicate(FCmpInst::FCMP_ORD);
2834 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2839 // Handle fcmp with constant RHS
2840 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2841 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2842 switch (LHSI->getOpcode()) {
2843 case Instruction::FPExt: {
2844 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2845 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2846 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2850 // We can't convert a PPC double double.
2851 if (RHSF->getType()->isPPC_FP128Ty())
2854 const fltSemantics *Sem;
2855 // FIXME: This shouldn't be here.
2856 if (LHSExt->getSrcTy()->isHalfTy())
2857 Sem = &APFloat::IEEEhalf;
2858 else if (LHSExt->getSrcTy()->isFloatTy())
2859 Sem = &APFloat::IEEEsingle;
2860 else if (LHSExt->getSrcTy()->isDoubleTy())
2861 Sem = &APFloat::IEEEdouble;
2862 else if (LHSExt->getSrcTy()->isFP128Ty())
2863 Sem = &APFloat::IEEEquad;
2864 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2865 Sem = &APFloat::x87DoubleExtended;
2870 APFloat F = RHSF->getValueAPF();
2871 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2873 // Avoid lossy conversions and denormals. Zero is a special case
2874 // that's OK to convert.
2878 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
2879 APFloat::cmpLessThan) || Fabs.isZero()))
2881 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2882 ConstantFP::get(RHSC->getContext(), F));
2885 case Instruction::PHI:
2886 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2887 // block. If in the same block, we're encouraging jump threading. If
2888 // not, we are just pessimizing the code by making an i1 phi.
2889 if (LHSI->getParent() == I.getParent())
2890 if (Instruction *NV = FoldOpIntoPhi(I))
2893 case Instruction::SIToFP:
2894 case Instruction::UIToFP:
2895 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2898 case Instruction::Select: {
2899 // If either operand of the select is a constant, we can fold the
2900 // comparison into the select arms, which will cause one to be
2901 // constant folded and the select turned into a bitwise or.
2902 Value *Op1 = 0, *Op2 = 0;
2903 if (LHSI->hasOneUse()) {
2904 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2905 // Fold the known value into the constant operand.
2906 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2907 // Insert a new FCmp of the other select operand.
2908 Op2 = Builder->CreateFCmp(I.getPredicate(),
2909 LHSI->getOperand(2), RHSC, I.getName());
2910 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2911 // Fold the known value into the constant operand.
2912 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2913 // Insert a new FCmp of the other select operand.
2914 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2920 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2923 case Instruction::FSub: {
2924 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2926 if (match(LHSI, m_FNeg(m_Value(Op))))
2927 return new FCmpInst(I.getSwappedPredicate(), Op,
2928 ConstantExpr::getFNeg(RHSC));
2931 case Instruction::Load:
2932 if (GetElementPtrInst *GEP =
2933 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2934 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2935 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2936 !cast<LoadInst>(LHSI)->isVolatile())
2937 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2944 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
2946 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
2947 return new FCmpInst(I.getSwappedPredicate(), X, Y);
2949 // fcmp (fpext x), (fpext y) -> fcmp x, y
2950 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
2951 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
2952 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
2953 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2954 RHSExt->getOperand(0));
2956 return Changed ? &I : 0;