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 // If we're comparing GEPs with two base pointers that only differ in type
611 // and both GEPs have only constant indices or just one use, then fold
612 // the compare with the adjusted indices.
613 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
614 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
615 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
616 PtrBase->stripPointerCasts() ==
617 GEPRHS->getOperand(0)->stripPointerCasts()) {
618 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
619 EmitGEPOffset(GEPLHS),
620 EmitGEPOffset(GEPRHS));
621 return ReplaceInstUsesWith(I, Cmp);
624 // Otherwise, the base pointers are different and the indices are
625 // different, bail out.
629 // If one of the GEPs has all zero indices, recurse.
630 bool AllZeros = true;
631 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
632 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
633 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
638 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
639 ICmpInst::getSwappedPredicate(Cond), I);
641 // If the other GEP has all zero indices, recurse.
643 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
644 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
645 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
650 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
652 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
653 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
654 // If the GEPs only differ by one index, compare it.
655 unsigned NumDifferences = 0; // Keep track of # differences.
656 unsigned DiffOperand = 0; // The operand that differs.
657 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
658 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
659 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
660 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
661 // Irreconcilable differences.
665 if (NumDifferences++) break;
670 if (NumDifferences == 0) // SAME GEP?
671 return ReplaceInstUsesWith(I, // No comparison is needed here.
672 ConstantInt::get(Type::getInt1Ty(I.getContext()),
673 ICmpInst::isTrueWhenEqual(Cond)));
675 else if (NumDifferences == 1 && GEPsInBounds) {
676 Value *LHSV = GEPLHS->getOperand(DiffOperand);
677 Value *RHSV = GEPRHS->getOperand(DiffOperand);
678 // Make sure we do a signed comparison here.
679 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
683 // Only lower this if the icmp is the only user of the GEP or if we expect
684 // the result to fold to a constant!
687 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
688 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
689 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
690 Value *L = EmitGEPOffset(GEPLHS);
691 Value *R = EmitGEPOffset(GEPRHS);
692 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
698 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
699 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
700 Value *X, ConstantInt *CI,
701 ICmpInst::Predicate Pred,
703 // If we have X+0, exit early (simplifying logic below) and let it get folded
704 // elsewhere. icmp X+0, X -> icmp X, X
706 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
707 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
710 // (X+4) == X -> false.
711 if (Pred == ICmpInst::ICMP_EQ)
712 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
714 // (X+4) != X -> true.
715 if (Pred == ICmpInst::ICMP_NE)
716 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
718 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
719 // so the values can never be equal. Similarly for all other "or equals"
722 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
723 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
724 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
725 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
727 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
728 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
731 // (X+1) >u X --> X <u (0-1) --> X != 255
732 // (X+2) >u X --> X <u (0-2) --> X <u 254
733 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
734 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
735 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
737 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
738 ConstantInt *SMax = ConstantInt::get(X->getContext(),
739 APInt::getSignedMaxValue(BitWidth));
741 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
742 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
743 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
744 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
745 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
746 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
747 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
748 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
750 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
751 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
752 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
753 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
754 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
755 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
757 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
758 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
759 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
762 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
763 /// and CmpRHS are both known to be integer constants.
764 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
765 ConstantInt *DivRHS) {
766 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
767 const APInt &CmpRHSV = CmpRHS->getValue();
769 // FIXME: If the operand types don't match the type of the divide
770 // then don't attempt this transform. The code below doesn't have the
771 // logic to deal with a signed divide and an unsigned compare (and
772 // vice versa). This is because (x /s C1) <s C2 produces different
773 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
774 // (x /u C1) <u C2. Simply casting the operands and result won't
775 // work. :( The if statement below tests that condition and bails
777 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
778 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
780 if (DivRHS->isZero())
781 return 0; // The ProdOV computation fails on divide by zero.
782 if (DivIsSigned && DivRHS->isAllOnesValue())
783 return 0; // The overflow computation also screws up here
784 if (DivRHS->isOne()) {
785 // This eliminates some funny cases with INT_MIN.
786 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
790 // Compute Prod = CI * DivRHS. We are essentially solving an equation
791 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
792 // C2 (CI). By solving for X we can turn this into a range check
793 // instead of computing a divide.
794 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
796 // Determine if the product overflows by seeing if the product is
797 // not equal to the divide. Make sure we do the same kind of divide
798 // as in the LHS instruction that we're folding.
799 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
800 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
802 // Get the ICmp opcode
803 ICmpInst::Predicate Pred = ICI.getPredicate();
805 /// If the division is known to be exact, then there is no remainder from the
806 /// divide, so the covered range size is unit, otherwise it is the divisor.
807 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
809 // Figure out the interval that is being checked. For example, a comparison
810 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
811 // Compute this interval based on the constants involved and the signedness of
812 // the compare/divide. This computes a half-open interval, keeping track of
813 // whether either value in the interval overflows. After analysis each
814 // overflow variable is set to 0 if it's corresponding bound variable is valid
815 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
816 int LoOverflow = 0, HiOverflow = 0;
817 Constant *LoBound = 0, *HiBound = 0;
819 if (!DivIsSigned) { // udiv
820 // e.g. X/5 op 3 --> [15, 20)
822 HiOverflow = LoOverflow = ProdOV;
824 // If this is not an exact divide, then many values in the range collapse
825 // to the same result value.
826 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
829 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
830 if (CmpRHSV == 0) { // (X / pos) op 0
831 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
832 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
834 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
835 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
836 HiOverflow = LoOverflow = ProdOV;
838 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
839 } else { // (X / pos) op neg
840 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
841 HiBound = AddOne(Prod);
842 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
844 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
845 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
848 } else if (DivRHS->isNegative()) { // Divisor is < 0.
850 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
851 if (CmpRHSV == 0) { // (X / neg) op 0
852 // e.g. X/-5 op 0 --> [-4, 5)
853 LoBound = AddOne(RangeSize);
854 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
855 if (HiBound == DivRHS) { // -INTMIN = INTMIN
856 HiOverflow = 1; // [INTMIN+1, overflow)
857 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
859 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
860 // e.g. X/-5 op 3 --> [-19, -14)
861 HiBound = AddOne(Prod);
862 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
864 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
865 } else { // (X / neg) op neg
866 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
867 LoOverflow = HiOverflow = ProdOV;
869 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
872 // Dividing by a negative swaps the condition. LT <-> GT
873 Pred = ICmpInst::getSwappedPredicate(Pred);
876 Value *X = DivI->getOperand(0);
878 default: llvm_unreachable("Unhandled icmp opcode!");
879 case ICmpInst::ICMP_EQ:
880 if (LoOverflow && HiOverflow)
881 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
883 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
884 ICmpInst::ICMP_UGE, X, LoBound);
886 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
887 ICmpInst::ICMP_ULT, X, HiBound);
888 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
890 case ICmpInst::ICMP_NE:
891 if (LoOverflow && HiOverflow)
892 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
894 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
895 ICmpInst::ICMP_ULT, X, LoBound);
897 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
898 ICmpInst::ICMP_UGE, X, HiBound);
899 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
900 DivIsSigned, false));
901 case ICmpInst::ICMP_ULT:
902 case ICmpInst::ICMP_SLT:
903 if (LoOverflow == +1) // Low bound is greater than input range.
904 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
905 if (LoOverflow == -1) // Low bound is less than input range.
906 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
907 return new ICmpInst(Pred, X, LoBound);
908 case ICmpInst::ICMP_UGT:
909 case ICmpInst::ICMP_SGT:
910 if (HiOverflow == +1) // High bound greater than input range.
911 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
912 if (HiOverflow == -1) // High bound less than input range.
913 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
914 if (Pred == ICmpInst::ICMP_UGT)
915 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
916 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
920 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
921 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
922 ConstantInt *ShAmt) {
923 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
925 // Check that the shift amount is in range. If not, don't perform
926 // undefined shifts. When the shift is visited it will be
928 uint32_t TypeBits = CmpRHSV.getBitWidth();
929 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
930 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
933 if (!ICI.isEquality()) {
934 // If we have an unsigned comparison and an ashr, we can't simplify this.
935 // Similarly for signed comparisons with lshr.
936 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
939 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
940 // by a power of 2. Since we already have logic to simplify these,
941 // transform to div and then simplify the resultant comparison.
942 if (Shr->getOpcode() == Instruction::AShr &&
943 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
946 // Revisit the shift (to delete it).
950 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
953 Shr->getOpcode() == Instruction::AShr ?
954 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
955 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
957 ICI.setOperand(0, Tmp);
959 // If the builder folded the binop, just return it.
960 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
964 // Otherwise, fold this div/compare.
965 assert(TheDiv->getOpcode() == Instruction::SDiv ||
966 TheDiv->getOpcode() == Instruction::UDiv);
968 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
969 assert(Res && "This div/cst should have folded!");
974 // If we are comparing against bits always shifted out, the
975 // comparison cannot succeed.
976 APInt Comp = CmpRHSV << ShAmtVal;
977 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
978 if (Shr->getOpcode() == Instruction::LShr)
979 Comp = Comp.lshr(ShAmtVal);
981 Comp = Comp.ashr(ShAmtVal);
983 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
984 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
985 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
987 return ReplaceInstUsesWith(ICI, Cst);
990 // Otherwise, check to see if the bits shifted out are known to be zero.
991 // If so, we can compare against the unshifted value:
992 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
993 if (Shr->hasOneUse() && Shr->isExact())
994 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
996 if (Shr->hasOneUse()) {
997 // Otherwise strength reduce the shift into an and.
998 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
999 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
1001 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1002 Mask, Shr->getName()+".mask");
1003 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1009 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1011 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1014 const APInt &RHSV = RHS->getValue();
1016 switch (LHSI->getOpcode()) {
1017 case Instruction::Trunc:
1018 if (ICI.isEquality() && LHSI->hasOneUse()) {
1019 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1020 // of the high bits truncated out of x are known.
1021 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1022 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1023 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
1024 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1025 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
1027 // If all the high bits are known, we can do this xform.
1028 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1029 // Pull in the high bits from known-ones set.
1030 APInt NewRHS = RHS->getValue().zext(SrcBits);
1032 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1033 ConstantInt::get(ICI.getContext(), NewRHS));
1038 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1039 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1040 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1042 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1043 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1044 Value *CompareVal = LHSI->getOperand(0);
1046 // If the sign bit of the XorCST is not set, there is no change to
1047 // the operation, just stop using the Xor.
1048 if (!XorCST->isNegative()) {
1049 ICI.setOperand(0, CompareVal);
1054 // Was the old condition true if the operand is positive?
1055 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1057 // If so, the new one isn't.
1058 isTrueIfPositive ^= true;
1060 if (isTrueIfPositive)
1061 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1064 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1068 if (LHSI->hasOneUse()) {
1069 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1070 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1071 const APInt &SignBit = XorCST->getValue();
1072 ICmpInst::Predicate Pred = ICI.isSigned()
1073 ? ICI.getUnsignedPredicate()
1074 : ICI.getSignedPredicate();
1075 return new ICmpInst(Pred, LHSI->getOperand(0),
1076 ConstantInt::get(ICI.getContext(),
1080 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1081 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1082 const APInt &NotSignBit = XorCST->getValue();
1083 ICmpInst::Predicate Pred = ICI.isSigned()
1084 ? ICI.getUnsignedPredicate()
1085 : ICI.getSignedPredicate();
1086 Pred = ICI.getSwappedPredicate(Pred);
1087 return new ICmpInst(Pred, LHSI->getOperand(0),
1088 ConstantInt::get(ICI.getContext(),
1089 RHSV ^ NotSignBit));
1094 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1095 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1096 LHSI->getOperand(0)->hasOneUse()) {
1097 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1099 // If the LHS is an AND of a truncating cast, we can widen the
1100 // and/compare to be the input width without changing the value
1101 // produced, eliminating a cast.
1102 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1103 // We can do this transformation if either the AND constant does not
1104 // have its sign bit set or if it is an equality comparison.
1105 // Extending a relational comparison when we're checking the sign
1106 // bit would not work.
1107 if (ICI.isEquality() ||
1108 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1110 Builder->CreateAnd(Cast->getOperand(0),
1111 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1112 NewAnd->takeName(LHSI);
1113 return new ICmpInst(ICI.getPredicate(), NewAnd,
1114 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1118 // If the LHS is an AND of a zext, and we have an equality compare, we can
1119 // shrink the and/compare to the smaller type, eliminating the cast.
1120 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1121 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1122 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1123 // should fold the icmp to true/false in that case.
1124 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1126 Builder->CreateAnd(Cast->getOperand(0),
1127 ConstantExpr::getTrunc(AndCST, Ty));
1128 NewAnd->takeName(LHSI);
1129 return new ICmpInst(ICI.getPredicate(), NewAnd,
1130 ConstantExpr::getTrunc(RHS, Ty));
1134 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1135 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1136 // happens a LOT in code produced by the C front-end, for bitfield
1138 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1139 if (Shift && !Shift->isShift())
1143 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1144 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1145 Type *AndTy = AndCST->getType(); // Type of the and.
1147 // We can fold this as long as we can't shift unknown bits
1148 // into the mask. This can only happen with signed shift
1149 // rights, as they sign-extend.
1151 bool CanFold = Shift->isLogicalShift();
1153 // To test for the bad case of the signed shr, see if any
1154 // of the bits shifted in could be tested after the mask.
1155 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1156 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1158 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1159 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1160 AndCST->getValue()) == 0)
1166 if (Shift->getOpcode() == Instruction::Shl)
1167 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1169 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1171 // Check to see if we are shifting out any of the bits being
1173 if (ConstantExpr::get(Shift->getOpcode(),
1174 NewCst, ShAmt) != RHS) {
1175 // If we shifted bits out, the fold is not going to work out.
1176 // As a special case, check to see if this means that the
1177 // result is always true or false now.
1178 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1179 return ReplaceInstUsesWith(ICI,
1180 ConstantInt::getFalse(ICI.getContext()));
1181 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1182 return ReplaceInstUsesWith(ICI,
1183 ConstantInt::getTrue(ICI.getContext()));
1185 ICI.setOperand(1, NewCst);
1186 Constant *NewAndCST;
1187 if (Shift->getOpcode() == Instruction::Shl)
1188 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1190 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1191 LHSI->setOperand(1, NewAndCST);
1192 LHSI->setOperand(0, Shift->getOperand(0));
1193 Worklist.Add(Shift); // Shift is dead.
1199 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1200 // preferable because it allows the C<<Y expression to be hoisted out
1201 // of a loop if Y is invariant and X is not.
1202 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1203 ICI.isEquality() && !Shift->isArithmeticShift() &&
1204 !isa<Constant>(Shift->getOperand(0))) {
1207 if (Shift->getOpcode() == Instruction::LShr) {
1208 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1210 // Insert a logical shift.
1211 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1214 // Compute X & (C << Y).
1216 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1218 ICI.setOperand(0, NewAnd);
1223 // Try to optimize things like "A[i]&42 == 0" to index computations.
1224 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1225 if (GetElementPtrInst *GEP =
1226 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1227 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1228 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1229 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1230 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1231 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1237 case Instruction::Or: {
1238 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1241 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1242 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1243 // -> and (icmp eq P, null), (icmp eq Q, null).
1244 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1245 Constant::getNullValue(P->getType()));
1246 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1247 Constant::getNullValue(Q->getType()));
1249 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1250 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1252 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1258 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1259 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1262 uint32_t TypeBits = RHSV.getBitWidth();
1264 // Check that the shift amount is in range. If not, don't perform
1265 // undefined shifts. When the shift is visited it will be
1267 if (ShAmt->uge(TypeBits))
1270 if (ICI.isEquality()) {
1271 // If we are comparing against bits always shifted out, the
1272 // comparison cannot succeed.
1274 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1276 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1277 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1279 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1280 return ReplaceInstUsesWith(ICI, Cst);
1283 // If the shift is NUW, then it is just shifting out zeros, no need for an
1285 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1286 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1287 ConstantExpr::getLShr(RHS, ShAmt));
1289 if (LHSI->hasOneUse()) {
1290 // Otherwise strength reduce the shift into an and.
1291 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1293 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1294 TypeBits-ShAmtVal));
1297 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1298 return new ICmpInst(ICI.getPredicate(), And,
1299 ConstantExpr::getLShr(RHS, ShAmt));
1303 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1304 bool TrueIfSigned = false;
1305 if (LHSI->hasOneUse() &&
1306 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1307 // (X << 31) <s 0 --> (X&1) != 0
1308 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1309 APInt::getOneBitSet(TypeBits,
1310 TypeBits-ShAmt->getZExtValue()-1));
1312 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1313 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1314 And, Constant::getNullValue(And->getType()));
1319 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1320 case Instruction::AShr: {
1321 // Handle equality comparisons of shift-by-constant.
1322 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1323 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1324 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1328 // Handle exact shr's.
1329 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1330 if (RHSV.isMinValue())
1331 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1336 case Instruction::SDiv:
1337 case Instruction::UDiv:
1338 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1339 // Fold this div into the comparison, producing a range check.
1340 // Determine, based on the divide type, what the range is being
1341 // checked. If there is an overflow on the low or high side, remember
1342 // it, otherwise compute the range [low, hi) bounding the new value.
1343 // See: InsertRangeTest above for the kinds of replacements possible.
1344 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1345 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1350 case Instruction::Add:
1351 // Fold: icmp pred (add X, C1), C2
1352 if (!ICI.isEquality()) {
1353 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1355 const APInt &LHSV = LHSC->getValue();
1357 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1360 if (ICI.isSigned()) {
1361 if (CR.getLower().isSignBit()) {
1362 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1363 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1364 } else if (CR.getUpper().isSignBit()) {
1365 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1366 ConstantInt::get(ICI.getContext(),CR.getLower()));
1369 if (CR.getLower().isMinValue()) {
1370 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1371 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1372 } else if (CR.getUpper().isMinValue()) {
1373 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1374 ConstantInt::get(ICI.getContext(),CR.getLower()));
1381 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1382 if (ICI.isEquality()) {
1383 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1385 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1386 // the second operand is a constant, simplify a bit.
1387 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1388 switch (BO->getOpcode()) {
1389 case Instruction::SRem:
1390 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1391 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1392 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1393 if (V.sgt(1) && V.isPowerOf2()) {
1395 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1397 return new ICmpInst(ICI.getPredicate(), NewRem,
1398 Constant::getNullValue(BO->getType()));
1402 case Instruction::Add:
1403 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1404 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1405 if (BO->hasOneUse())
1406 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1407 ConstantExpr::getSub(RHS, BOp1C));
1408 } else if (RHSV == 0) {
1409 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1410 // efficiently invertible, or if the add has just this one use.
1411 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1413 if (Value *NegVal = dyn_castNegVal(BOp1))
1414 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1415 if (Value *NegVal = dyn_castNegVal(BOp0))
1416 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1417 if (BO->hasOneUse()) {
1418 Value *Neg = Builder->CreateNeg(BOp1);
1420 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1424 case Instruction::Xor:
1425 // For the xor case, we can xor two constants together, eliminating
1426 // the explicit xor.
1427 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1428 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1429 ConstantExpr::getXor(RHS, BOC));
1430 } else if (RHSV == 0) {
1431 // Replace ((xor A, B) != 0) with (A != B)
1432 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1436 case Instruction::Sub:
1437 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1438 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1439 if (BO->hasOneUse())
1440 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1441 ConstantExpr::getSub(BOp0C, RHS));
1442 } else if (RHSV == 0) {
1443 // Replace ((sub A, B) != 0) with (A != B)
1444 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1448 case Instruction::Or:
1449 // If bits are being or'd in that are not present in the constant we
1450 // are comparing against, then the comparison could never succeed!
1451 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1452 Constant *NotCI = ConstantExpr::getNot(RHS);
1453 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1454 return ReplaceInstUsesWith(ICI,
1455 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1460 case Instruction::And:
1461 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1462 // If bits are being compared against that are and'd out, then the
1463 // comparison can never succeed!
1464 if ((RHSV & ~BOC->getValue()) != 0)
1465 return ReplaceInstUsesWith(ICI,
1466 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1469 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1470 if (RHS == BOC && RHSV.isPowerOf2())
1471 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1472 ICmpInst::ICMP_NE, LHSI,
1473 Constant::getNullValue(RHS->getType()));
1475 // Don't perform the following transforms if the AND has multiple uses
1476 if (!BO->hasOneUse())
1479 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1480 if (BOC->getValue().isSignBit()) {
1481 Value *X = BO->getOperand(0);
1482 Constant *Zero = Constant::getNullValue(X->getType());
1483 ICmpInst::Predicate pred = isICMP_NE ?
1484 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1485 return new ICmpInst(pred, X, Zero);
1488 // ((X & ~7) == 0) --> X < 8
1489 if (RHSV == 0 && isHighOnes(BOC)) {
1490 Value *X = BO->getOperand(0);
1491 Constant *NegX = ConstantExpr::getNeg(BOC);
1492 ICmpInst::Predicate pred = isICMP_NE ?
1493 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1494 return new ICmpInst(pred, X, NegX);
1499 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1500 // Handle icmp {eq|ne} <intrinsic>, intcst.
1501 switch (II->getIntrinsicID()) {
1502 case Intrinsic::bswap:
1504 ICI.setOperand(0, II->getArgOperand(0));
1505 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1507 case Intrinsic::ctlz:
1508 case Intrinsic::cttz:
1509 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1510 if (RHSV == RHS->getType()->getBitWidth()) {
1512 ICI.setOperand(0, II->getArgOperand(0));
1513 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1517 case Intrinsic::ctpop:
1518 // popcount(A) == 0 -> A == 0 and likewise for !=
1519 if (RHS->isZero()) {
1521 ICI.setOperand(0, II->getArgOperand(0));
1522 ICI.setOperand(1, RHS);
1534 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1535 /// We only handle extending casts so far.
1537 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1538 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1539 Value *LHSCIOp = LHSCI->getOperand(0);
1540 Type *SrcTy = LHSCIOp->getType();
1541 Type *DestTy = LHSCI->getType();
1544 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1545 // integer type is the same size as the pointer type.
1546 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1547 TD->getPointerSizeInBits() ==
1548 cast<IntegerType>(DestTy)->getBitWidth()) {
1550 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1551 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1552 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1553 RHSOp = RHSC->getOperand(0);
1554 // If the pointer types don't match, insert a bitcast.
1555 if (LHSCIOp->getType() != RHSOp->getType())
1556 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1560 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1563 // The code below only handles extension cast instructions, so far.
1565 if (LHSCI->getOpcode() != Instruction::ZExt &&
1566 LHSCI->getOpcode() != Instruction::SExt)
1569 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1570 bool isSignedCmp = ICI.isSigned();
1572 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1573 // Not an extension from the same type?
1574 RHSCIOp = CI->getOperand(0);
1575 if (RHSCIOp->getType() != LHSCIOp->getType())
1578 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1579 // and the other is a zext), then we can't handle this.
1580 if (CI->getOpcode() != LHSCI->getOpcode())
1583 // Deal with equality cases early.
1584 if (ICI.isEquality())
1585 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1587 // A signed comparison of sign extended values simplifies into a
1588 // signed comparison.
1589 if (isSignedCmp && isSignedExt)
1590 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1592 // The other three cases all fold into an unsigned comparison.
1593 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1596 // If we aren't dealing with a constant on the RHS, exit early
1597 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1601 // Compute the constant that would happen if we truncated to SrcTy then
1602 // reextended to DestTy.
1603 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1604 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1607 // If the re-extended constant didn't change...
1609 // Deal with equality cases early.
1610 if (ICI.isEquality())
1611 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1613 // A signed comparison of sign extended values simplifies into a
1614 // signed comparison.
1615 if (isSignedExt && isSignedCmp)
1616 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1618 // The other three cases all fold into an unsigned comparison.
1619 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1622 // The re-extended constant changed so the constant cannot be represented
1623 // in the shorter type. Consequently, we cannot emit a simple comparison.
1624 // All the cases that fold to true or false will have already been handled
1625 // by SimplifyICmpInst, so only deal with the tricky case.
1627 if (isSignedCmp || !isSignedExt)
1630 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1631 // should have been folded away previously and not enter in here.
1633 // We're performing an unsigned comp with a sign extended value.
1634 // This is true if the input is >= 0. [aka >s -1]
1635 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1636 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1638 // Finally, return the value computed.
1639 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1640 return ReplaceInstUsesWith(ICI, Result);
1642 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1643 return BinaryOperator::CreateNot(Result);
1646 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1647 /// I = icmp ugt (add (add A, B), CI2), CI1
1648 /// If this is of the form:
1650 /// if (sum+128 >u 255)
1651 /// Then replace it with llvm.sadd.with.overflow.i8.
1653 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1654 ConstantInt *CI2, ConstantInt *CI1,
1656 // The transformation we're trying to do here is to transform this into an
1657 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1658 // with a narrower add, and discard the add-with-constant that is part of the
1659 // range check (if we can't eliminate it, this isn't profitable).
1661 // In order to eliminate the add-with-constant, the compare can be its only
1663 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1664 if (!AddWithCst->hasOneUse()) return 0;
1666 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1667 if (!CI2->getValue().isPowerOf2()) return 0;
1668 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1669 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1671 // The width of the new add formed is 1 more than the bias.
1674 // Check to see that CI1 is an all-ones value with NewWidth bits.
1675 if (CI1->getBitWidth() == NewWidth ||
1676 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1679 // This is only really a signed overflow check if the inputs have been
1680 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1681 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1682 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1683 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1684 IC.ComputeNumSignBits(B) < NeededSignBits)
1687 // In order to replace the original add with a narrower
1688 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1689 // and truncates that discard the high bits of the add. Verify that this is
1691 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1692 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1694 if (*UI == AddWithCst) continue;
1696 // Only accept truncates for now. We would really like a nice recursive
1697 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1698 // chain to see which bits of a value are actually demanded. If the
1699 // original add had another add which was then immediately truncated, we
1700 // could still do the transformation.
1701 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1703 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1706 // If the pattern matches, truncate the inputs to the narrower type and
1707 // use the sadd_with_overflow intrinsic to efficiently compute both the
1708 // result and the overflow bit.
1709 Module *M = I.getParent()->getParent()->getParent();
1711 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1712 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1715 InstCombiner::BuilderTy *Builder = IC.Builder;
1717 // Put the new code above the original add, in case there are any uses of the
1718 // add between the add and the compare.
1719 Builder->SetInsertPoint(OrigAdd);
1721 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1722 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1723 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1724 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1725 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1727 // The inner add was the result of the narrow add, zero extended to the
1728 // wider type. Replace it with the result computed by the intrinsic.
1729 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1731 // The original icmp gets replaced with the overflow value.
1732 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1735 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1737 // Don't bother doing this transformation for pointers, don't do it for
1739 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1741 // If the add is a constant expr, then we don't bother transforming it.
1742 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1743 if (OrigAdd == 0) return 0;
1745 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1747 // Put the new code above the original add, in case there are any uses of the
1748 // add between the add and the compare.
1749 InstCombiner::BuilderTy *Builder = IC.Builder;
1750 Builder->SetInsertPoint(OrigAdd);
1752 Module *M = I.getParent()->getParent()->getParent();
1753 Type *Ty = LHS->getType();
1754 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1755 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1756 Value *Add = Builder->CreateExtractValue(Call, 0);
1758 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1760 // The original icmp gets replaced with the overflow value.
1761 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1764 // DemandedBitsLHSMask - When performing a comparison against a constant,
1765 // it is possible that not all the bits in the LHS are demanded. This helper
1766 // method computes the mask that IS demanded.
1767 static APInt DemandedBitsLHSMask(ICmpInst &I,
1768 unsigned BitWidth, bool isSignCheck) {
1770 return APInt::getSignBit(BitWidth);
1772 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1773 if (!CI) return APInt::getAllOnesValue(BitWidth);
1774 const APInt &RHS = CI->getValue();
1776 switch (I.getPredicate()) {
1777 // For a UGT comparison, we don't care about any bits that
1778 // correspond to the trailing ones of the comparand. The value of these
1779 // bits doesn't impact the outcome of the comparison, because any value
1780 // greater than the RHS must differ in a bit higher than these due to carry.
1781 case ICmpInst::ICMP_UGT: {
1782 unsigned trailingOnes = RHS.countTrailingOnes();
1783 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1787 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1788 // Any value less than the RHS must differ in a higher bit because of carries.
1789 case ICmpInst::ICMP_ULT: {
1790 unsigned trailingZeros = RHS.countTrailingZeros();
1791 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1796 return APInt::getAllOnesValue(BitWidth);
1801 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1802 bool Changed = false;
1803 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1805 /// Orders the operands of the compare so that they are listed from most
1806 /// complex to least complex. This puts constants before unary operators,
1807 /// before binary operators.
1808 if (getComplexity(Op0) < getComplexity(Op1)) {
1810 std::swap(Op0, Op1);
1814 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1815 return ReplaceInstUsesWith(I, V);
1817 // comparing -val or val with non-zero is the same as just comparing val
1818 // ie, abs(val) != 0 -> val != 0
1819 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
1821 Value *Cond, *SelectTrue, *SelectFalse;
1822 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
1823 m_Value(SelectFalse)))) {
1824 if (Value *V = dyn_castNegVal(SelectTrue)) {
1825 if (V == SelectFalse)
1826 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1828 else if (Value *V = dyn_castNegVal(SelectFalse)) {
1829 if (V == SelectTrue)
1830 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1835 Type *Ty = Op0->getType();
1837 // icmp's with boolean values can always be turned into bitwise operations
1838 if (Ty->isIntegerTy(1)) {
1839 switch (I.getPredicate()) {
1840 default: llvm_unreachable("Invalid icmp instruction!");
1841 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1842 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1843 return BinaryOperator::CreateNot(Xor);
1845 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1846 return BinaryOperator::CreateXor(Op0, Op1);
1848 case ICmpInst::ICMP_UGT:
1849 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1851 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1852 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1853 return BinaryOperator::CreateAnd(Not, Op1);
1855 case ICmpInst::ICMP_SGT:
1856 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1858 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1859 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1860 return BinaryOperator::CreateAnd(Not, Op0);
1862 case ICmpInst::ICMP_UGE:
1863 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1865 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1866 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1867 return BinaryOperator::CreateOr(Not, Op1);
1869 case ICmpInst::ICMP_SGE:
1870 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1872 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1873 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1874 return BinaryOperator::CreateOr(Not, Op0);
1879 unsigned BitWidth = 0;
1880 if (Ty->isIntOrIntVectorTy())
1881 BitWidth = Ty->getScalarSizeInBits();
1882 else if (TD) // Pointers require TD info to get their size.
1883 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1885 bool isSignBit = false;
1887 // See if we are doing a comparison with a constant.
1888 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1889 Value *A = 0, *B = 0;
1891 // Match the following pattern, which is a common idiom when writing
1892 // overflow-safe integer arithmetic function. The source performs an
1893 // addition in wider type, and explicitly checks for overflow using
1894 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1895 // sadd_with_overflow intrinsic.
1897 // TODO: This could probably be generalized to handle other overflow-safe
1898 // operations if we worked out the formulas to compute the appropriate
1902 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1904 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1905 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1906 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1907 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1911 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1912 if (I.isEquality() && CI->isZero() &&
1913 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1914 // (icmp cond A B) if cond is equality
1915 return new ICmpInst(I.getPredicate(), A, B);
1918 // If we have an icmp le or icmp ge instruction, turn it into the
1919 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1920 // them being folded in the code below. The SimplifyICmpInst code has
1921 // already handled the edge cases for us, so we just assert on them.
1922 switch (I.getPredicate()) {
1924 case ICmpInst::ICMP_ULE:
1925 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1926 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1927 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1928 case ICmpInst::ICMP_SLE:
1929 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1930 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1931 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1932 case ICmpInst::ICMP_UGE:
1933 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1934 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1935 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1936 case ICmpInst::ICMP_SGE:
1937 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1938 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1939 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1942 // If this comparison is a normal comparison, it demands all
1943 // bits, if it is a sign bit comparison, it only demands the sign bit.
1945 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1948 // See if we can fold the comparison based on range information we can get
1949 // by checking whether bits are known to be zero or one in the input.
1950 if (BitWidth != 0) {
1951 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1952 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1954 if (SimplifyDemandedBits(I.getOperandUse(0),
1955 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1956 Op0KnownZero, Op0KnownOne, 0))
1958 if (SimplifyDemandedBits(I.getOperandUse(1),
1959 APInt::getAllOnesValue(BitWidth),
1960 Op1KnownZero, Op1KnownOne, 0))
1963 // Given the known and unknown bits, compute a range that the LHS could be
1964 // in. Compute the Min, Max and RHS values based on the known bits. For the
1965 // EQ and NE we use unsigned values.
1966 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1967 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1969 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1971 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1974 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1976 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1980 // If Min and Max are known to be the same, then SimplifyDemandedBits
1981 // figured out that the LHS is a constant. Just constant fold this now so
1982 // that code below can assume that Min != Max.
1983 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1984 return new ICmpInst(I.getPredicate(),
1985 ConstantInt::get(Op0->getType(), Op0Min), Op1);
1986 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1987 return new ICmpInst(I.getPredicate(), Op0,
1988 ConstantInt::get(Op1->getType(), Op1Min));
1990 // Based on the range information we know about the LHS, see if we can
1991 // simplify this comparison. For example, (x&4) < 8 is always true.
1992 switch (I.getPredicate()) {
1993 default: llvm_unreachable("Unknown icmp opcode!");
1994 case ICmpInst::ICMP_EQ: {
1995 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1996 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
1998 // If all bits are known zero except for one, then we know at most one
1999 // bit is set. If the comparison is against zero, then this is a check
2000 // to see if *that* bit is set.
2001 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2002 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2003 // If the LHS is an AND with the same constant, look through it.
2005 ConstantInt *LHSC = 0;
2006 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2007 LHSC->getValue() != Op0KnownZeroInverted)
2010 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2011 // then turn "((1 << x)&8) == 0" into "x != 3".
2013 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2014 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2015 return new ICmpInst(ICmpInst::ICMP_NE, X,
2016 ConstantInt::get(X->getType(), CmpVal));
2019 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2020 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2022 if (Op0KnownZeroInverted == 1 &&
2023 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2024 return new ICmpInst(ICmpInst::ICMP_NE, X,
2025 ConstantInt::get(X->getType(),
2026 CI->countTrailingZeros()));
2031 case ICmpInst::ICMP_NE: {
2032 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2033 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2035 // If all bits are known zero except for one, then we know at most one
2036 // bit is set. If the comparison is against zero, then this is a check
2037 // to see if *that* bit is set.
2038 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2039 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2040 // If the LHS is an AND with the same constant, look through it.
2042 ConstantInt *LHSC = 0;
2043 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2044 LHSC->getValue() != Op0KnownZeroInverted)
2047 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2048 // then turn "((1 << x)&8) != 0" into "x == 3".
2050 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2051 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2052 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2053 ConstantInt::get(X->getType(), CmpVal));
2056 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2057 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2059 if (Op0KnownZeroInverted == 1 &&
2060 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2061 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2062 ConstantInt::get(X->getType(),
2063 CI->countTrailingZeros()));
2068 case ICmpInst::ICMP_ULT:
2069 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2070 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2071 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2072 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2073 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2074 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2075 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2076 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2077 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2078 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2080 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2081 if (CI->isMinValue(true))
2082 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2083 Constant::getAllOnesValue(Op0->getType()));
2086 case ICmpInst::ICMP_UGT:
2087 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2088 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2089 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2090 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2092 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2093 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2094 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2095 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2096 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2097 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2099 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2100 if (CI->isMaxValue(true))
2101 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2102 Constant::getNullValue(Op0->getType()));
2105 case ICmpInst::ICMP_SLT:
2106 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2107 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2108 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2109 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2110 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2111 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2112 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2113 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2114 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2115 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2118 case ICmpInst::ICMP_SGT:
2119 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2120 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2121 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2122 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2124 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2125 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2126 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2127 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2128 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2129 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2132 case ICmpInst::ICMP_SGE:
2133 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2134 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2135 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2136 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2137 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2139 case ICmpInst::ICMP_SLE:
2140 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2141 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2142 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2143 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2144 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2146 case ICmpInst::ICMP_UGE:
2147 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2148 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2149 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2150 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2151 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2153 case ICmpInst::ICMP_ULE:
2154 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2155 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2156 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2157 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2158 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2162 // Turn a signed comparison into an unsigned one if both operands
2163 // are known to have the same sign.
2165 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2166 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2167 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2170 // Test if the ICmpInst instruction is used exclusively by a select as
2171 // part of a minimum or maximum operation. If so, refrain from doing
2172 // any other folding. This helps out other analyses which understand
2173 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2174 // and CodeGen. And in this case, at least one of the comparison
2175 // operands has at least one user besides the compare (the select),
2176 // which would often largely negate the benefit of folding anyway.
2178 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2179 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2180 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2183 // See if we are doing a comparison between a constant and an instruction that
2184 // can be folded into the comparison.
2185 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2186 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2187 // instruction, see if that instruction also has constants so that the
2188 // instruction can be folded into the icmp
2189 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2190 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2194 // Handle icmp with constant (but not simple integer constant) RHS
2195 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2196 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2197 switch (LHSI->getOpcode()) {
2198 case Instruction::GetElementPtr:
2199 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2200 if (RHSC->isNullValue() &&
2201 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2202 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2203 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2205 case Instruction::PHI:
2206 // Only fold icmp into the PHI if the phi and icmp are in the same
2207 // block. If in the same block, we're encouraging jump threading. If
2208 // not, we are just pessimizing the code by making an i1 phi.
2209 if (LHSI->getParent() == I.getParent())
2210 if (Instruction *NV = FoldOpIntoPhi(I))
2213 case Instruction::Select: {
2214 // If either operand of the select is a constant, we can fold the
2215 // comparison into the select arms, which will cause one to be
2216 // constant folded and the select turned into a bitwise or.
2217 Value *Op1 = 0, *Op2 = 0;
2218 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2219 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2220 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2221 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2223 // We only want to perform this transformation if it will not lead to
2224 // additional code. This is true if either both sides of the select
2225 // fold to a constant (in which case the icmp is replaced with a select
2226 // which will usually simplify) or this is the only user of the
2227 // select (in which case we are trading a select+icmp for a simpler
2229 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2231 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2234 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2236 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2240 case Instruction::IntToPtr:
2241 // icmp pred inttoptr(X), null -> icmp pred X, 0
2242 if (RHSC->isNullValue() && TD &&
2243 TD->getIntPtrType(RHSC->getContext()) ==
2244 LHSI->getOperand(0)->getType())
2245 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2246 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2249 case Instruction::Load:
2250 // Try to optimize things like "A[i] > 4" to index computations.
2251 if (GetElementPtrInst *GEP =
2252 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2253 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2254 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2255 !cast<LoadInst>(LHSI)->isVolatile())
2256 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2263 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2264 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2265 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2267 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2268 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2269 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2272 // Test to see if the operands of the icmp are casted versions of other
2273 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2275 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2276 if (Op0->getType()->isPointerTy() &&
2277 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2278 // We keep moving the cast from the left operand over to the right
2279 // operand, where it can often be eliminated completely.
2280 Op0 = CI->getOperand(0);
2282 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2283 // so eliminate it as well.
2284 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2285 Op1 = CI2->getOperand(0);
2287 // If Op1 is a constant, we can fold the cast into the constant.
2288 if (Op0->getType() != Op1->getType()) {
2289 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2290 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2292 // Otherwise, cast the RHS right before the icmp
2293 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2296 return new ICmpInst(I.getPredicate(), Op0, Op1);
2300 if (isa<CastInst>(Op0)) {
2301 // Handle the special case of: icmp (cast bool to X), <cst>
2302 // This comes up when you have code like
2305 // For generality, we handle any zero-extension of any operand comparison
2306 // with a constant or another cast from the same type.
2307 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2308 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2312 // Special logic for binary operators.
2313 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2314 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2316 CmpInst::Predicate Pred = I.getPredicate();
2317 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2318 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2319 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2320 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2321 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2322 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2323 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2324 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2325 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2327 // Analyze the case when either Op0 or Op1 is an add instruction.
2328 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2329 Value *A = 0, *B = 0, *C = 0, *D = 0;
2330 if (BO0 && BO0->getOpcode() == Instruction::Add)
2331 A = BO0->getOperand(0), B = BO0->getOperand(1);
2332 if (BO1 && BO1->getOpcode() == Instruction::Add)
2333 C = BO1->getOperand(0), D = BO1->getOperand(1);
2335 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2336 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2337 return new ICmpInst(Pred, A == Op1 ? B : A,
2338 Constant::getNullValue(Op1->getType()));
2340 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2341 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2342 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2345 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2346 if (A && C && (A == C || A == D || B == C || B == D) &&
2347 NoOp0WrapProblem && NoOp1WrapProblem &&
2348 // Try not to increase register pressure.
2349 BO0->hasOneUse() && BO1->hasOneUse()) {
2350 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2351 Value *Y = (A == C || A == D) ? B : A;
2352 Value *Z = (C == A || C == B) ? D : C;
2353 return new ICmpInst(Pred, Y, Z);
2356 // Analyze the case when either Op0 or Op1 is a sub instruction.
2357 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2358 A = 0; B = 0; C = 0; D = 0;
2359 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2360 A = BO0->getOperand(0), B = BO0->getOperand(1);
2361 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2362 C = BO1->getOperand(0), D = BO1->getOperand(1);
2364 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2365 if (A == Op1 && NoOp0WrapProblem)
2366 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2368 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2369 if (C == Op0 && NoOp1WrapProblem)
2370 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2372 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2373 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2374 // Try not to increase register pressure.
2375 BO0->hasOneUse() && BO1->hasOneUse())
2376 return new ICmpInst(Pred, A, C);
2378 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2379 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2380 // Try not to increase register pressure.
2381 BO0->hasOneUse() && BO1->hasOneUse())
2382 return new ICmpInst(Pred, D, B);
2384 BinaryOperator *SRem = NULL;
2385 // icmp (srem X, Y), Y
2386 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2387 Op1 == BO0->getOperand(1))
2389 // icmp Y, (srem X, Y)
2390 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2391 Op0 == BO1->getOperand(1))
2394 // We don't check hasOneUse to avoid increasing register pressure because
2395 // the value we use is the same value this instruction was already using.
2396 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2398 case ICmpInst::ICMP_EQ:
2399 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2400 case ICmpInst::ICMP_NE:
2401 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2402 case ICmpInst::ICMP_SGT:
2403 case ICmpInst::ICMP_SGE:
2404 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2405 Constant::getAllOnesValue(SRem->getType()));
2406 case ICmpInst::ICMP_SLT:
2407 case ICmpInst::ICMP_SLE:
2408 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2409 Constant::getNullValue(SRem->getType()));
2413 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2414 BO0->hasOneUse() && BO1->hasOneUse() &&
2415 BO0->getOperand(1) == BO1->getOperand(1)) {
2416 switch (BO0->getOpcode()) {
2418 case Instruction::Add:
2419 case Instruction::Sub:
2420 case Instruction::Xor:
2421 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2422 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2423 BO1->getOperand(0));
2424 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2425 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2426 if (CI->getValue().isSignBit()) {
2427 ICmpInst::Predicate Pred = I.isSigned()
2428 ? I.getUnsignedPredicate()
2429 : I.getSignedPredicate();
2430 return new ICmpInst(Pred, BO0->getOperand(0),
2431 BO1->getOperand(0));
2434 if (CI->isMaxValue(true)) {
2435 ICmpInst::Predicate Pred = I.isSigned()
2436 ? I.getUnsignedPredicate()
2437 : I.getSignedPredicate();
2438 Pred = I.getSwappedPredicate(Pred);
2439 return new ICmpInst(Pred, BO0->getOperand(0),
2440 BO1->getOperand(0));
2444 case Instruction::Mul:
2445 if (!I.isEquality())
2448 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2449 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2450 // Mask = -1 >> count-trailing-zeros(Cst).
2451 if (!CI->isZero() && !CI->isOne()) {
2452 const APInt &AP = CI->getValue();
2453 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2454 APInt::getLowBitsSet(AP.getBitWidth(),
2456 AP.countTrailingZeros()));
2457 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2458 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2459 return new ICmpInst(I.getPredicate(), And1, And2);
2463 case Instruction::UDiv:
2464 case Instruction::LShr:
2468 case Instruction::SDiv:
2469 case Instruction::AShr:
2470 if (!BO0->isExact() || !BO1->isExact())
2472 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2473 BO1->getOperand(0));
2474 case Instruction::Shl: {
2475 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2476 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2479 if (!NSW && I.isSigned())
2481 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2482 BO1->getOperand(0));
2489 // ~x < ~y --> y < x
2490 // ~x < cst --> ~cst < x
2491 if (match(Op0, m_Not(m_Value(A)))) {
2492 if (match(Op1, m_Not(m_Value(B))))
2493 return new ICmpInst(I.getPredicate(), B, A);
2494 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2495 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2498 // (a+b) <u a --> llvm.uadd.with.overflow.
2499 // (a+b) <u b --> llvm.uadd.with.overflow.
2500 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2501 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2502 (Op1 == A || Op1 == B))
2503 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2506 // a >u (a+b) --> llvm.uadd.with.overflow.
2507 // b >u (a+b) --> llvm.uadd.with.overflow.
2508 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2509 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2510 (Op0 == A || Op0 == B))
2511 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2515 if (I.isEquality()) {
2516 Value *A, *B, *C, *D;
2518 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2519 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2520 Value *OtherVal = A == Op1 ? B : A;
2521 return new ICmpInst(I.getPredicate(), OtherVal,
2522 Constant::getNullValue(A->getType()));
2525 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2526 // A^c1 == C^c2 --> A == C^(c1^c2)
2527 ConstantInt *C1, *C2;
2528 if (match(B, m_ConstantInt(C1)) &&
2529 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2530 Constant *NC = ConstantInt::get(I.getContext(),
2531 C1->getValue() ^ C2->getValue());
2532 Value *Xor = Builder->CreateXor(C, NC);
2533 return new ICmpInst(I.getPredicate(), A, Xor);
2536 // A^B == A^D -> B == D
2537 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2538 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2539 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2540 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2544 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2545 (A == Op0 || B == Op0)) {
2546 // A == (A^B) -> B == 0
2547 Value *OtherVal = A == Op0 ? B : A;
2548 return new ICmpInst(I.getPredicate(), OtherVal,
2549 Constant::getNullValue(A->getType()));
2552 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2553 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2554 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2555 Value *X = 0, *Y = 0, *Z = 0;
2558 X = B; Y = D; Z = A;
2559 } else if (A == D) {
2560 X = B; Y = C; Z = A;
2561 } else if (B == C) {
2562 X = A; Y = D; Z = B;
2563 } else if (B == D) {
2564 X = A; Y = C; Z = B;
2567 if (X) { // Build (X^Y) & Z
2568 Op1 = Builder->CreateXor(X, Y);
2569 Op1 = Builder->CreateAnd(Op1, Z);
2570 I.setOperand(0, Op1);
2571 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2576 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2577 // "icmp (and X, mask), cst"
2580 if (Op0->hasOneUse() &&
2581 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2582 m_ConstantInt(ShAmt))))) &&
2583 match(Op1, m_ConstantInt(Cst1)) &&
2584 // Only do this when A has multiple uses. This is most important to do
2585 // when it exposes other optimizations.
2587 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2589 if (ShAmt < ASize) {
2591 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2594 APInt CmpV = Cst1->getValue().zext(ASize);
2597 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2598 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2604 Value *X; ConstantInt *Cst;
2606 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2607 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2610 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2611 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2613 return Changed ? &I : 0;
2621 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2623 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2626 if (!isa<ConstantFP>(RHSC)) return 0;
2627 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2629 // Get the width of the mantissa. We don't want to hack on conversions that
2630 // might lose information from the integer, e.g. "i64 -> float"
2631 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2632 if (MantissaWidth == -1) return 0; // Unknown.
2634 // Check to see that the input is converted from an integer type that is small
2635 // enough that preserves all bits. TODO: check here for "known" sign bits.
2636 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2637 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2639 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2640 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2644 // If the conversion would lose info, don't hack on this.
2645 if ((int)InputSize > MantissaWidth)
2648 // Otherwise, we can potentially simplify the comparison. We know that it
2649 // will always come through as an integer value and we know the constant is
2650 // not a NAN (it would have been previously simplified).
2651 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2653 ICmpInst::Predicate Pred;
2654 switch (I.getPredicate()) {
2655 default: llvm_unreachable("Unexpected predicate!");
2656 case FCmpInst::FCMP_UEQ:
2657 case FCmpInst::FCMP_OEQ:
2658 Pred = ICmpInst::ICMP_EQ;
2660 case FCmpInst::FCMP_UGT:
2661 case FCmpInst::FCMP_OGT:
2662 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2664 case FCmpInst::FCMP_UGE:
2665 case FCmpInst::FCMP_OGE:
2666 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2668 case FCmpInst::FCMP_ULT:
2669 case FCmpInst::FCMP_OLT:
2670 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2672 case FCmpInst::FCMP_ULE:
2673 case FCmpInst::FCMP_OLE:
2674 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2676 case FCmpInst::FCMP_UNE:
2677 case FCmpInst::FCMP_ONE:
2678 Pred = ICmpInst::ICMP_NE;
2680 case FCmpInst::FCMP_ORD:
2681 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2682 case FCmpInst::FCMP_UNO:
2683 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2686 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2688 // Now we know that the APFloat is a normal number, zero or inf.
2690 // See if the FP constant is too large for the integer. For example,
2691 // comparing an i8 to 300.0.
2692 unsigned IntWidth = IntTy->getScalarSizeInBits();
2695 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2696 // and large values.
2697 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2698 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2699 APFloat::rmNearestTiesToEven);
2700 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2701 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2702 Pred == ICmpInst::ICMP_SLE)
2703 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2704 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2707 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2708 // +INF and large values.
2709 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2710 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2711 APFloat::rmNearestTiesToEven);
2712 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2713 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2714 Pred == ICmpInst::ICMP_ULE)
2715 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2716 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2721 // See if the RHS value is < SignedMin.
2722 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2723 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2724 APFloat::rmNearestTiesToEven);
2725 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2726 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2727 Pred == ICmpInst::ICMP_SGE)
2728 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2729 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2732 // See if the RHS value is < UnsignedMin.
2733 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2734 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
2735 APFloat::rmNearestTiesToEven);
2736 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
2737 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
2738 Pred == ICmpInst::ICMP_UGE)
2739 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2740 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2744 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2745 // [0, UMAX], but it may still be fractional. See if it is fractional by
2746 // casting the FP value to the integer value and back, checking for equality.
2747 // Don't do this for zero, because -0.0 is not fractional.
2748 Constant *RHSInt = LHSUnsigned
2749 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2750 : ConstantExpr::getFPToSI(RHSC, IntTy);
2751 if (!RHS.isZero()) {
2752 bool Equal = LHSUnsigned
2753 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2754 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2756 // If we had a comparison against a fractional value, we have to adjust
2757 // the compare predicate and sometimes the value. RHSC is rounded towards
2758 // zero at this point.
2760 default: llvm_unreachable("Unexpected integer comparison!");
2761 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2762 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2763 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2764 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2765 case ICmpInst::ICMP_ULE:
2766 // (float)int <= 4.4 --> int <= 4
2767 // (float)int <= -4.4 --> false
2768 if (RHS.isNegative())
2769 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2771 case ICmpInst::ICMP_SLE:
2772 // (float)int <= 4.4 --> int <= 4
2773 // (float)int <= -4.4 --> int < -4
2774 if (RHS.isNegative())
2775 Pred = ICmpInst::ICMP_SLT;
2777 case ICmpInst::ICMP_ULT:
2778 // (float)int < -4.4 --> false
2779 // (float)int < 4.4 --> int <= 4
2780 if (RHS.isNegative())
2781 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2782 Pred = ICmpInst::ICMP_ULE;
2784 case ICmpInst::ICMP_SLT:
2785 // (float)int < -4.4 --> int < -4
2786 // (float)int < 4.4 --> int <= 4
2787 if (!RHS.isNegative())
2788 Pred = ICmpInst::ICMP_SLE;
2790 case ICmpInst::ICMP_UGT:
2791 // (float)int > 4.4 --> int > 4
2792 // (float)int > -4.4 --> true
2793 if (RHS.isNegative())
2794 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2796 case ICmpInst::ICMP_SGT:
2797 // (float)int > 4.4 --> int > 4
2798 // (float)int > -4.4 --> int >= -4
2799 if (RHS.isNegative())
2800 Pred = ICmpInst::ICMP_SGE;
2802 case ICmpInst::ICMP_UGE:
2803 // (float)int >= -4.4 --> true
2804 // (float)int >= 4.4 --> int > 4
2805 if (!RHS.isNegative())
2806 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2807 Pred = ICmpInst::ICMP_UGT;
2809 case ICmpInst::ICMP_SGE:
2810 // (float)int >= -4.4 --> int >= -4
2811 // (float)int >= 4.4 --> int > 4
2812 if (!RHS.isNegative())
2813 Pred = ICmpInst::ICMP_SGT;
2819 // Lower this FP comparison into an appropriate integer version of the
2821 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2824 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2825 bool Changed = false;
2827 /// Orders the operands of the compare so that they are listed from most
2828 /// complex to least complex. This puts constants before unary operators,
2829 /// before binary operators.
2830 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2835 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2837 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2838 return ReplaceInstUsesWith(I, V);
2840 // Simplify 'fcmp pred X, X'
2842 switch (I.getPredicate()) {
2843 default: llvm_unreachable("Unknown predicate!");
2844 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2845 case FCmpInst::FCMP_ULT: // True if unordered or less than
2846 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2847 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2848 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2849 I.setPredicate(FCmpInst::FCMP_UNO);
2850 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2853 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2854 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2855 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2856 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2857 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2858 I.setPredicate(FCmpInst::FCMP_ORD);
2859 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2864 // Handle fcmp with constant RHS
2865 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2866 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2867 switch (LHSI->getOpcode()) {
2868 case Instruction::FPExt: {
2869 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2870 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2871 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2875 // We can't convert a PPC double double.
2876 if (RHSF->getType()->isPPC_FP128Ty())
2879 const fltSemantics *Sem;
2880 // FIXME: This shouldn't be here.
2881 if (LHSExt->getSrcTy()->isHalfTy())
2882 Sem = &APFloat::IEEEhalf;
2883 else if (LHSExt->getSrcTy()->isFloatTy())
2884 Sem = &APFloat::IEEEsingle;
2885 else if (LHSExt->getSrcTy()->isDoubleTy())
2886 Sem = &APFloat::IEEEdouble;
2887 else if (LHSExt->getSrcTy()->isFP128Ty())
2888 Sem = &APFloat::IEEEquad;
2889 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2890 Sem = &APFloat::x87DoubleExtended;
2895 APFloat F = RHSF->getValueAPF();
2896 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2898 // Avoid lossy conversions and denormals. Zero is a special case
2899 // that's OK to convert.
2903 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
2904 APFloat::cmpLessThan) || Fabs.isZero()))
2906 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2907 ConstantFP::get(RHSC->getContext(), F));
2910 case Instruction::PHI:
2911 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2912 // block. If in the same block, we're encouraging jump threading. If
2913 // not, we are just pessimizing the code by making an i1 phi.
2914 if (LHSI->getParent() == I.getParent())
2915 if (Instruction *NV = FoldOpIntoPhi(I))
2918 case Instruction::SIToFP:
2919 case Instruction::UIToFP:
2920 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2923 case Instruction::Select: {
2924 // If either operand of the select is a constant, we can fold the
2925 // comparison into the select arms, which will cause one to be
2926 // constant folded and the select turned into a bitwise or.
2927 Value *Op1 = 0, *Op2 = 0;
2928 if (LHSI->hasOneUse()) {
2929 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2930 // Fold the known value into the constant operand.
2931 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2932 // Insert a new FCmp of the other select operand.
2933 Op2 = Builder->CreateFCmp(I.getPredicate(),
2934 LHSI->getOperand(2), RHSC, I.getName());
2935 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2936 // Fold the known value into the constant operand.
2937 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2938 // Insert a new FCmp of the other select operand.
2939 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2945 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2948 case Instruction::FSub: {
2949 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2951 if (match(LHSI, m_FNeg(m_Value(Op))))
2952 return new FCmpInst(I.getSwappedPredicate(), Op,
2953 ConstantExpr::getFNeg(RHSC));
2956 case Instruction::Load:
2957 if (GetElementPtrInst *GEP =
2958 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2959 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2960 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2961 !cast<LoadInst>(LHSI)->isVolatile())
2962 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2969 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
2971 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
2972 return new FCmpInst(I.getSwappedPredicate(), X, Y);
2974 // fcmp (fpext x), (fpext y) -> fcmp x, y
2975 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
2976 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
2977 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
2978 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2979 RHSExt->getOperand(0));
2981 return Changed ? &I : 0;