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/Analysis/ConstantFolding.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/GetElementPtrTypeIterator.h"
20 #include "llvm/IR/IntrinsicInst.h"
21 #include "llvm/IR/PatternMatch.h"
22 #include "llvm/Support/ConstantRange.h"
23 #include "llvm/Target/TargetLibraryInfo.h"
25 using namespace PatternMatch;
27 static ConstantInt *getOne(Constant *C) {
28 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
31 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
32 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
35 static bool HasAddOverflow(ConstantInt *Result,
36 ConstantInt *In1, ConstantInt *In2,
39 return Result->getValue().ult(In1->getValue());
41 if (In2->isNegative())
42 return Result->getValue().sgt(In1->getValue());
43 return Result->getValue().slt(In1->getValue());
46 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
47 /// overflowed for this type.
48 static bool AddWithOverflow(Constant *&Result, Constant *In1,
49 Constant *In2, bool IsSigned = false) {
50 Result = ConstantExpr::getAdd(In1, In2);
52 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
53 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
54 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
55 if (HasAddOverflow(ExtractElement(Result, Idx),
56 ExtractElement(In1, Idx),
57 ExtractElement(In2, Idx),
64 return HasAddOverflow(cast<ConstantInt>(Result),
65 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
69 static bool HasSubOverflow(ConstantInt *Result,
70 ConstantInt *In1, ConstantInt *In2,
73 return Result->getValue().ugt(In1->getValue());
75 if (In2->isNegative())
76 return Result->getValue().slt(In1->getValue());
78 return Result->getValue().sgt(In1->getValue());
81 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
82 /// overflowed for this type.
83 static bool SubWithOverflow(Constant *&Result, Constant *In1,
84 Constant *In2, bool IsSigned = false) {
85 Result = ConstantExpr::getSub(In1, In2);
87 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
88 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
89 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
90 if (HasSubOverflow(ExtractElement(Result, Idx),
91 ExtractElement(In1, Idx),
92 ExtractElement(In2, Idx),
99 return HasSubOverflow(cast<ConstantInt>(Result),
100 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
104 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
105 /// comparison only checks the sign bit. If it only checks the sign bit, set
106 /// TrueIfSigned if the result of the comparison is true when the input value is
108 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
109 bool &TrueIfSigned) {
111 case ICmpInst::ICMP_SLT: // True if LHS s< 0
113 return RHS->isZero();
114 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
116 return RHS->isAllOnesValue();
117 case ICmpInst::ICMP_SGT: // True if LHS s> -1
118 TrueIfSigned = false;
119 return RHS->isAllOnesValue();
120 case ICmpInst::ICMP_UGT:
121 // True if LHS u> RHS and RHS == high-bit-mask - 1
123 return RHS->isMaxValue(true);
124 case ICmpInst::ICMP_UGE:
125 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
127 return RHS->getValue().isSignBit();
133 /// Returns true if the exploded icmp can be expressed as a signed comparison
134 /// to zero and updates the predicate accordingly.
135 /// The signedness of the comparison is preserved.
136 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
137 if (!ICmpInst::isSigned(pred))
141 return ICmpInst::isRelational(pred);
144 if (pred == ICmpInst::ICMP_SLT) {
145 pred = ICmpInst::ICMP_SLE;
148 } else if (RHS->isAllOnesValue()) {
149 if (pred == ICmpInst::ICMP_SGT) {
150 pred = ICmpInst::ICMP_SGE;
158 // isHighOnes - Return true if the constant is of the form 1+0+.
159 // This is the same as lowones(~X).
160 static bool isHighOnes(const ConstantInt *CI) {
161 return (~CI->getValue() + 1).isPowerOf2();
164 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
165 /// set of known zero and one bits, compute the maximum and minimum values that
166 /// could have the specified known zero and known one bits, returning them in
168 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
169 const APInt& KnownOne,
170 APInt& Min, APInt& Max) {
171 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
172 KnownZero.getBitWidth() == Min.getBitWidth() &&
173 KnownZero.getBitWidth() == Max.getBitWidth() &&
174 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
175 APInt UnknownBits = ~(KnownZero|KnownOne);
177 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
178 // bit if it is unknown.
180 Max = KnownOne|UnknownBits;
182 if (UnknownBits.isNegative()) { // Sign bit is unknown
183 Min.setBit(Min.getBitWidth()-1);
184 Max.clearBit(Max.getBitWidth()-1);
188 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
189 // a set of known zero and one bits, compute the maximum and minimum values that
190 // could have the specified known zero and known one bits, returning them in
192 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
193 const APInt &KnownOne,
194 APInt &Min, APInt &Max) {
195 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
196 KnownZero.getBitWidth() == Min.getBitWidth() &&
197 KnownZero.getBitWidth() == Max.getBitWidth() &&
198 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
199 APInt UnknownBits = ~(KnownZero|KnownOne);
201 // The minimum value is when the unknown bits are all zeros.
203 // The maximum value is when the unknown bits are all ones.
204 Max = KnownOne|UnknownBits;
209 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
210 /// cmp pred (load (gep GV, ...)), cmpcst
211 /// where GV is a global variable with a constant initializer. Try to simplify
212 /// this into some simple computation that does not need the load. For example
213 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
215 /// If AndCst is non-null, then the loaded value is masked with that constant
216 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
217 Instruction *InstCombiner::
218 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
219 CmpInst &ICI, ConstantInt *AndCst) {
220 // We need TD information to know the pointer size unless this is inbounds.
221 if (!GEP->isInBounds() && DL == 0)
224 Constant *Init = GV->getInitializer();
225 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
228 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
229 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays.
231 // There are many forms of this optimization we can handle, for now, just do
232 // the simple index into a single-dimensional array.
234 // Require: GEP GV, 0, i {{, constant indices}}
235 if (GEP->getNumOperands() < 3 ||
236 !isa<ConstantInt>(GEP->getOperand(1)) ||
237 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
238 isa<Constant>(GEP->getOperand(2)))
241 // Check that indices after the variable are constants and in-range for the
242 // type they index. Collect the indices. This is typically for arrays of
244 SmallVector<unsigned, 4> LaterIndices;
246 Type *EltTy = Init->getType()->getArrayElementType();
247 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
248 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
249 if (Idx == 0) return 0; // Variable index.
251 uint64_t IdxVal = Idx->getZExtValue();
252 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
254 if (StructType *STy = dyn_cast<StructType>(EltTy))
255 EltTy = STy->getElementType(IdxVal);
256 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
257 if (IdxVal >= ATy->getNumElements()) return 0;
258 EltTy = ATy->getElementType();
260 return 0; // Unknown type.
263 LaterIndices.push_back(IdxVal);
266 enum { Overdefined = -3, Undefined = -2 };
268 // Variables for our state machines.
270 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
271 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
272 // and 87 is the second (and last) index. FirstTrueElement is -2 when
273 // undefined, otherwise set to the first true element. SecondTrueElement is
274 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
275 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
277 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
278 // form "i != 47 & i != 87". Same state transitions as for true elements.
279 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
281 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
282 /// define a state machine that triggers for ranges of values that the index
283 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
284 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
285 /// index in the range (inclusive). We use -2 for undefined here because we
286 /// use relative comparisons and don't want 0-1 to match -1.
287 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
289 // MagicBitvector - This is a magic bitvector where we set a bit if the
290 // comparison is true for element 'i'. If there are 64 elements or less in
291 // the array, this will fully represent all the comparison results.
292 uint64_t MagicBitvector = 0;
295 // Scan the array and see if one of our patterns matches.
296 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
297 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
298 Constant *Elt = Init->getAggregateElement(i);
299 if (Elt == 0) return 0;
301 // If this is indexing an array of structures, get the structure element.
302 if (!LaterIndices.empty())
303 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
305 // If the element is masked, handle it.
306 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
308 // Find out if the comparison would be true or false for the i'th element.
309 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
310 CompareRHS, DL, TLI);
311 // If the result is undef for this element, ignore it.
312 if (isa<UndefValue>(C)) {
313 // Extend range state machines to cover this element in case there is an
314 // undef in the middle of the range.
315 if (TrueRangeEnd == (int)i-1)
317 if (FalseRangeEnd == (int)i-1)
322 // If we can't compute the result for any of the elements, we have to give
323 // up evaluating the entire conditional.
324 if (!isa<ConstantInt>(C)) return 0;
326 // Otherwise, we know if the comparison is true or false for this element,
327 // update our state machines.
328 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
330 // State machine for single/double/range index comparison.
332 // Update the TrueElement state machine.
333 if (FirstTrueElement == Undefined)
334 FirstTrueElement = TrueRangeEnd = i; // First true element.
336 // Update double-compare state machine.
337 if (SecondTrueElement == Undefined)
338 SecondTrueElement = i;
340 SecondTrueElement = Overdefined;
342 // Update range state machine.
343 if (TrueRangeEnd == (int)i-1)
346 TrueRangeEnd = Overdefined;
349 // Update the FalseElement state machine.
350 if (FirstFalseElement == Undefined)
351 FirstFalseElement = FalseRangeEnd = i; // First false element.
353 // Update double-compare state machine.
354 if (SecondFalseElement == Undefined)
355 SecondFalseElement = i;
357 SecondFalseElement = Overdefined;
359 // Update range state machine.
360 if (FalseRangeEnd == (int)i-1)
363 FalseRangeEnd = Overdefined;
368 // If this element is in range, update our magic bitvector.
369 if (i < 64 && IsTrueForElt)
370 MagicBitvector |= 1ULL << i;
372 // If all of our states become overdefined, bail out early. Since the
373 // predicate is expensive, only check it every 8 elements. This is only
374 // really useful for really huge arrays.
375 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
376 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
377 FalseRangeEnd == Overdefined)
381 // Now that we've scanned the entire array, emit our new comparison(s). We
382 // order the state machines in complexity of the generated code.
383 Value *Idx = GEP->getOperand(2);
385 // If the index is larger than the pointer size of the target, truncate the
386 // index down like the GEP would do implicitly. We don't have to do this for
387 // an inbounds GEP because the index can't be out of range.
388 if (!GEP->isInBounds()) {
389 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
390 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
391 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
392 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
395 // If the comparison is only true for one or two elements, emit direct
397 if (SecondTrueElement != Overdefined) {
398 // None true -> false.
399 if (FirstTrueElement == Undefined)
400 return ReplaceInstUsesWith(ICI, Builder->getFalse());
402 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
404 // True for one element -> 'i == 47'.
405 if (SecondTrueElement == Undefined)
406 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
408 // True for two elements -> 'i == 47 | i == 72'.
409 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
410 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
411 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
412 return BinaryOperator::CreateOr(C1, C2);
415 // If the comparison is only false for one or two elements, emit direct
417 if (SecondFalseElement != Overdefined) {
418 // None false -> true.
419 if (FirstFalseElement == Undefined)
420 return ReplaceInstUsesWith(ICI, Builder->getTrue());
422 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
424 // False for one element -> 'i != 47'.
425 if (SecondFalseElement == Undefined)
426 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
428 // False for two elements -> 'i != 47 & i != 72'.
429 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
430 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
431 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
432 return BinaryOperator::CreateAnd(C1, C2);
435 // If the comparison can be replaced with a range comparison for the elements
436 // where it is true, emit the range check.
437 if (TrueRangeEnd != Overdefined) {
438 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
440 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
441 if (FirstTrueElement) {
442 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
443 Idx = Builder->CreateAdd(Idx, Offs);
446 Value *End = ConstantInt::get(Idx->getType(),
447 TrueRangeEnd-FirstTrueElement+1);
448 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
451 // False range check.
452 if (FalseRangeEnd != Overdefined) {
453 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
454 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
455 if (FirstFalseElement) {
456 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
457 Idx = Builder->CreateAdd(Idx, Offs);
460 Value *End = ConstantInt::get(Idx->getType(),
461 FalseRangeEnd-FirstFalseElement);
462 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
466 // If a magic bitvector captures the entire comparison state
467 // of this load, replace it with computation that does:
468 // ((magic_cst >> i) & 1) != 0
472 // Look for an appropriate type:
473 // - The type of Idx if the magic fits
474 // - The smallest fitting legal type if we have a DataLayout
476 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
479 Ty = DL->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
480 else if (ArrayElementCount <= 32)
481 Ty = Type::getInt32Ty(Init->getContext());
484 Value *V = Builder->CreateIntCast(Idx, Ty, false);
485 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
486 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
487 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
495 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
496 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
497 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
498 /// be complex, and scales are involved. The above expression would also be
499 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
500 /// This later form is less amenable to optimization though, and we are allowed
501 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
503 /// If we can't emit an optimized form for this expression, this returns null.
505 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
506 const DataLayout &DL = *IC.getDataLayout();
507 gep_type_iterator GTI = gep_type_begin(GEP);
509 // Check to see if this gep only has a single variable index. If so, and if
510 // any constant indices are a multiple of its scale, then we can compute this
511 // in terms of the scale of the variable index. For example, if the GEP
512 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
513 // because the expression will cross zero at the same point.
514 unsigned i, e = GEP->getNumOperands();
516 for (i = 1; i != e; ++i, ++GTI) {
517 if (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 += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
525 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
526 Offset += Size*CI->getSExtValue();
529 // Found our variable index.
534 // If there are no variable indices, we must have a constant offset, just
535 // evaluate it the general way.
536 if (i == e) return 0;
538 Value *VariableIdx = GEP->getOperand(i);
539 // Determine the scale factor of the variable element. For example, this is
540 // 4 if the variable index is into an array of i32.
541 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
543 // Verify that there are no other variable indices. If so, emit the hard way.
544 for (++i, ++GTI; i != e; ++i, ++GTI) {
545 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
548 // Compute the aggregate offset of constant indices.
549 if (CI->isZero()) continue;
551 // Handle a struct index, which adds its field offset to the pointer.
552 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
553 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
555 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
556 Offset += Size*CI->getSExtValue();
562 // Okay, we know we have a single variable index, which must be a
563 // pointer/array/vector index. If there is no offset, life is simple, return
565 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
566 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
568 // Cast to intptrty in case a truncation occurs. If an extension is needed,
569 // we don't need to bother extending: the extension won't affect where the
570 // computation crosses zero.
571 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
572 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
577 // Otherwise, there is an index. The computation we will do will be modulo
578 // the pointer size, so get it.
579 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
581 Offset &= PtrSizeMask;
582 VariableScale &= PtrSizeMask;
584 // To do this transformation, any constant index must be a multiple of the
585 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
586 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
587 // multiple of the variable scale.
588 int64_t NewOffs = Offset / (int64_t)VariableScale;
589 if (Offset != NewOffs*(int64_t)VariableScale)
592 // Okay, we can do this evaluation. Start by converting the index to intptr.
593 if (VariableIdx->getType() != IntPtrTy)
594 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
596 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
597 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
600 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
601 /// else. At this point we know that the GEP is on the LHS of the comparison.
602 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
603 ICmpInst::Predicate Cond,
605 // Don't transform signed compares of GEPs into index compares. Even if the
606 // GEP is inbounds, the final add of the base pointer can have signed overflow
607 // and would change the result of the icmp.
608 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
609 // the maximum signed value for the pointer type.
610 if (ICmpInst::isSigned(Cond))
613 // Look through bitcasts.
614 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
615 RHS = BCI->getOperand(0);
617 Value *PtrBase = GEPLHS->getOperand(0);
618 if (DL && PtrBase == RHS && GEPLHS->isInBounds()) {
619 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
620 // This transformation (ignoring the base and scales) is valid because we
621 // know pointers can't overflow since the gep is inbounds. See if we can
622 // output an optimized form.
623 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
625 // If not, synthesize the offset the hard way.
627 Offset = EmitGEPOffset(GEPLHS);
628 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
629 Constant::getNullValue(Offset->getType()));
630 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
631 // If the base pointers are different, but the indices are the same, just
632 // compare the base pointer.
633 if (PtrBase != GEPRHS->getOperand(0)) {
634 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
635 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
636 GEPRHS->getOperand(0)->getType();
638 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
639 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
640 IndicesTheSame = false;
644 // If all indices are the same, just compare the base pointers.
646 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
648 // If we're comparing GEPs with two base pointers that only differ in type
649 // and both GEPs have only constant indices or just one use, then fold
650 // the compare with the adjusted indices.
651 if (DL && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
652 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
653 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
654 PtrBase->stripPointerCasts() ==
655 GEPRHS->getOperand(0)->stripPointerCasts()) {
656 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
657 EmitGEPOffset(GEPLHS),
658 EmitGEPOffset(GEPRHS));
659 return ReplaceInstUsesWith(I, Cmp);
662 // Otherwise, the base pointers are different and the indices are
663 // different, bail out.
667 // If one of the GEPs has all zero indices, recurse.
668 bool AllZeros = true;
669 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
670 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
671 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
676 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
677 ICmpInst::getSwappedPredicate(Cond), I);
679 // If the other GEP has all zero indices, recurse.
681 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
682 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
683 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
688 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
690 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
691 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
692 // If the GEPs only differ by one index, compare it.
693 unsigned NumDifferences = 0; // Keep track of # differences.
694 unsigned DiffOperand = 0; // The operand that differs.
695 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
696 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
697 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
698 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
699 // Irreconcilable differences.
703 if (NumDifferences++) break;
708 if (NumDifferences == 0) // SAME GEP?
709 return ReplaceInstUsesWith(I, // No comparison is needed here.
710 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
712 else if (NumDifferences == 1 && GEPsInBounds) {
713 Value *LHSV = GEPLHS->getOperand(DiffOperand);
714 Value *RHSV = GEPRHS->getOperand(DiffOperand);
715 // Make sure we do a signed comparison here.
716 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
720 // Only lower this if the icmp is the only user of the GEP or if we expect
721 // the result to fold to a constant!
724 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
725 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
726 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
727 Value *L = EmitGEPOffset(GEPLHS);
728 Value *R = EmitGEPOffset(GEPRHS);
729 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
735 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
736 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
737 Value *X, ConstantInt *CI,
738 ICmpInst::Predicate Pred) {
739 // If we have X+0, exit early (simplifying logic below) and let it get folded
740 // elsewhere. icmp X+0, X -> icmp X, X
742 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
743 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
746 // (X+4) == X -> false.
747 if (Pred == ICmpInst::ICMP_EQ)
748 return ReplaceInstUsesWith(ICI, Builder->getFalse());
750 // (X+4) != X -> true.
751 if (Pred == ICmpInst::ICMP_NE)
752 return ReplaceInstUsesWith(ICI, Builder->getTrue());
754 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
755 // so the values can never be equal. Similarly for all other "or equals"
758 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
759 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
760 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
761 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
763 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
764 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
767 // (X+1) >u X --> X <u (0-1) --> X != 255
768 // (X+2) >u X --> X <u (0-2) --> X <u 254
769 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
770 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
771 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
773 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
774 ConstantInt *SMax = ConstantInt::get(X->getContext(),
775 APInt::getSignedMaxValue(BitWidth));
777 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
778 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
779 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
780 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
781 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
782 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
783 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
784 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
786 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
787 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
788 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
789 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
790 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
791 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
793 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
794 Constant *C = Builder->getInt(CI->getValue()-1);
795 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
798 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
799 /// and CmpRHS are both known to be integer constants.
800 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
801 ConstantInt *DivRHS) {
802 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
803 const APInt &CmpRHSV = CmpRHS->getValue();
805 // FIXME: If the operand types don't match the type of the divide
806 // then don't attempt this transform. The code below doesn't have the
807 // logic to deal with a signed divide and an unsigned compare (and
808 // vice versa). This is because (x /s C1) <s C2 produces different
809 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
810 // (x /u C1) <u C2. Simply casting the operands and result won't
811 // work. :( The if statement below tests that condition and bails
813 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
814 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
816 if (DivRHS->isZero())
817 return 0; // The ProdOV computation fails on divide by zero.
818 if (DivIsSigned && DivRHS->isAllOnesValue())
819 return 0; // The overflow computation also screws up here
820 if (DivRHS->isOne()) {
821 // This eliminates some funny cases with INT_MIN.
822 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
826 // Compute Prod = CI * DivRHS. We are essentially solving an equation
827 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
828 // C2 (CI). By solving for X we can turn this into a range check
829 // instead of computing a divide.
830 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
832 // Determine if the product overflows by seeing if the product is
833 // not equal to the divide. Make sure we do the same kind of divide
834 // as in the LHS instruction that we're folding.
835 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
836 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
838 // Get the ICmp opcode
839 ICmpInst::Predicate Pred = ICI.getPredicate();
841 /// If the division is known to be exact, then there is no remainder from the
842 /// divide, so the covered range size is unit, otherwise it is the divisor.
843 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
845 // Figure out the interval that is being checked. For example, a comparison
846 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
847 // Compute this interval based on the constants involved and the signedness of
848 // the compare/divide. This computes a half-open interval, keeping track of
849 // whether either value in the interval overflows. After analysis each
850 // overflow variable is set to 0 if it's corresponding bound variable is valid
851 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
852 int LoOverflow = 0, HiOverflow = 0;
853 Constant *LoBound = 0, *HiBound = 0;
855 if (!DivIsSigned) { // udiv
856 // e.g. X/5 op 3 --> [15, 20)
858 HiOverflow = LoOverflow = ProdOV;
860 // If this is not an exact divide, then many values in the range collapse
861 // to the same result value.
862 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
865 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
866 if (CmpRHSV == 0) { // (X / pos) op 0
867 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
868 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
870 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
871 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
872 HiOverflow = LoOverflow = ProdOV;
874 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
875 } else { // (X / pos) op neg
876 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
877 HiBound = AddOne(Prod);
878 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
880 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
881 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
884 } else if (DivRHS->isNegative()) { // Divisor is < 0.
886 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
887 if (CmpRHSV == 0) { // (X / neg) op 0
888 // e.g. X/-5 op 0 --> [-4, 5)
889 LoBound = AddOne(RangeSize);
890 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
891 if (HiBound == DivRHS) { // -INTMIN = INTMIN
892 HiOverflow = 1; // [INTMIN+1, overflow)
893 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
895 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
896 // e.g. X/-5 op 3 --> [-19, -14)
897 HiBound = AddOne(Prod);
898 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
900 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
901 } else { // (X / neg) op neg
902 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
903 LoOverflow = HiOverflow = ProdOV;
905 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
908 // Dividing by a negative swaps the condition. LT <-> GT
909 Pred = ICmpInst::getSwappedPredicate(Pred);
912 Value *X = DivI->getOperand(0);
914 default: llvm_unreachable("Unhandled icmp opcode!");
915 case ICmpInst::ICMP_EQ:
916 if (LoOverflow && HiOverflow)
917 return ReplaceInstUsesWith(ICI, Builder->getFalse());
919 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
920 ICmpInst::ICMP_UGE, X, LoBound);
922 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
923 ICmpInst::ICMP_ULT, X, HiBound);
924 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
926 case ICmpInst::ICMP_NE:
927 if (LoOverflow && HiOverflow)
928 return ReplaceInstUsesWith(ICI, Builder->getTrue());
930 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
931 ICmpInst::ICMP_ULT, X, LoBound);
933 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
934 ICmpInst::ICMP_UGE, X, HiBound);
935 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
936 DivIsSigned, false));
937 case ICmpInst::ICMP_ULT:
938 case ICmpInst::ICMP_SLT:
939 if (LoOverflow == +1) // Low bound is greater than input range.
940 return ReplaceInstUsesWith(ICI, Builder->getTrue());
941 if (LoOverflow == -1) // Low bound is less than input range.
942 return ReplaceInstUsesWith(ICI, Builder->getFalse());
943 return new ICmpInst(Pred, X, LoBound);
944 case ICmpInst::ICMP_UGT:
945 case ICmpInst::ICMP_SGT:
946 if (HiOverflow == +1) // High bound greater than input range.
947 return ReplaceInstUsesWith(ICI, Builder->getFalse());
948 if (HiOverflow == -1) // High bound less than input range.
949 return ReplaceInstUsesWith(ICI, Builder->getTrue());
950 if (Pred == ICmpInst::ICMP_UGT)
951 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
952 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
956 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
957 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
958 ConstantInt *ShAmt) {
959 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
961 // Check that the shift amount is in range. If not, don't perform
962 // undefined shifts. When the shift is visited it will be
964 uint32_t TypeBits = CmpRHSV.getBitWidth();
965 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
966 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
969 if (!ICI.isEquality()) {
970 // If we have an unsigned comparison and an ashr, we can't simplify this.
971 // Similarly for signed comparisons with lshr.
972 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
975 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
976 // by a power of 2. Since we already have logic to simplify these,
977 // transform to div and then simplify the resultant comparison.
978 if (Shr->getOpcode() == Instruction::AShr &&
979 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
982 // Revisit the shift (to delete it).
986 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
989 Shr->getOpcode() == Instruction::AShr ?
990 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
991 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
993 ICI.setOperand(0, Tmp);
995 // If the builder folded the binop, just return it.
996 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
1000 // Otherwise, fold this div/compare.
1001 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1002 TheDiv->getOpcode() == Instruction::UDiv);
1004 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1005 assert(Res && "This div/cst should have folded!");
1010 // If we are comparing against bits always shifted out, the
1011 // comparison cannot succeed.
1012 APInt Comp = CmpRHSV << ShAmtVal;
1013 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1014 if (Shr->getOpcode() == Instruction::LShr)
1015 Comp = Comp.lshr(ShAmtVal);
1017 Comp = Comp.ashr(ShAmtVal);
1019 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1020 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1021 Constant *Cst = Builder->getInt1(IsICMP_NE);
1022 return ReplaceInstUsesWith(ICI, Cst);
1025 // Otherwise, check to see if the bits shifted out are known to be zero.
1026 // If so, we can compare against the unshifted value:
1027 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1028 if (Shr->hasOneUse() && Shr->isExact())
1029 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1031 if (Shr->hasOneUse()) {
1032 // Otherwise strength reduce the shift into an and.
1033 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1034 Constant *Mask = Builder->getInt(Val);
1036 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1037 Mask, Shr->getName()+".mask");
1038 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1044 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1046 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1049 const APInt &RHSV = RHS->getValue();
1051 switch (LHSI->getOpcode()) {
1052 case Instruction::Trunc:
1053 if (ICI.isEquality() && LHSI->hasOneUse()) {
1054 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1055 // of the high bits truncated out of x are known.
1056 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1057 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1058 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1059 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1061 // If all the high bits are known, we can do this xform.
1062 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1063 // Pull in the high bits from known-ones set.
1064 APInt NewRHS = RHS->getValue().zext(SrcBits);
1065 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1066 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1067 Builder->getInt(NewRHS));
1072 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1073 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1074 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1076 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1077 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1078 Value *CompareVal = LHSI->getOperand(0);
1080 // If the sign bit of the XorCst is not set, there is no change to
1081 // the operation, just stop using the Xor.
1082 if (!XorCst->isNegative()) {
1083 ICI.setOperand(0, CompareVal);
1088 // Was the old condition true if the operand is positive?
1089 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1091 // If so, the new one isn't.
1092 isTrueIfPositive ^= true;
1094 if (isTrueIfPositive)
1095 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1098 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1102 if (LHSI->hasOneUse()) {
1103 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1104 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1105 const APInt &SignBit = XorCst->getValue();
1106 ICmpInst::Predicate Pred = ICI.isSigned()
1107 ? ICI.getUnsignedPredicate()
1108 : ICI.getSignedPredicate();
1109 return new ICmpInst(Pred, LHSI->getOperand(0),
1110 Builder->getInt(RHSV ^ SignBit));
1113 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1114 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1115 const APInt &NotSignBit = XorCst->getValue();
1116 ICmpInst::Predicate Pred = ICI.isSigned()
1117 ? ICI.getUnsignedPredicate()
1118 : ICI.getSignedPredicate();
1119 Pred = ICI.getSwappedPredicate(Pred);
1120 return new ICmpInst(Pred, LHSI->getOperand(0),
1121 Builder->getInt(RHSV ^ NotSignBit));
1125 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1126 // iff -C is a power of 2
1127 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1128 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1129 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1131 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1132 // iff -C is a power of 2
1133 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1134 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1135 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1138 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1139 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1140 LHSI->getOperand(0)->hasOneUse()) {
1141 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1143 // If the LHS is an AND of a truncating cast, we can widen the
1144 // and/compare to be the input width without changing the value
1145 // produced, eliminating a cast.
1146 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1147 // We can do this transformation if either the AND constant does not
1148 // have its sign bit set or if it is an equality comparison.
1149 // Extending a relational comparison when we're checking the sign
1150 // bit would not work.
1151 if (ICI.isEquality() ||
1152 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1154 Builder->CreateAnd(Cast->getOperand(0),
1155 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1156 NewAnd->takeName(LHSI);
1157 return new ICmpInst(ICI.getPredicate(), NewAnd,
1158 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1162 // If the LHS is an AND of a zext, and we have an equality compare, we can
1163 // shrink the and/compare to the smaller type, eliminating the cast.
1164 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1165 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1166 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1167 // should fold the icmp to true/false in that case.
1168 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1170 Builder->CreateAnd(Cast->getOperand(0),
1171 ConstantExpr::getTrunc(AndCst, Ty));
1172 NewAnd->takeName(LHSI);
1173 return new ICmpInst(ICI.getPredicate(), NewAnd,
1174 ConstantExpr::getTrunc(RHS, Ty));
1178 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1179 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1180 // happens a LOT in code produced by the C front-end, for bitfield
1182 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1183 if (Shift && !Shift->isShift())
1187 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1189 // This seemingly simple opportunity to fold away a shift turns out to
1190 // be rather complicated. See PR17827
1191 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1193 bool CanFold = false;
1194 unsigned ShiftOpcode = Shift->getOpcode();
1195 if (ShiftOpcode == Instruction::AShr) {
1196 // There may be some constraints that make this possible,
1197 // but nothing simple has been discovered yet.
1199 } else if (ShiftOpcode == Instruction::Shl) {
1200 // For a left shift, we can fold if the comparison is not signed.
1201 // We can also fold a signed comparison if the mask value and
1202 // comparison value are not negative. These constraints may not be
1203 // obvious, but we can prove that they are correct using an SMT
1205 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1207 } else if (ShiftOpcode == Instruction::LShr) {
1208 // For a logical right shift, we can fold if the comparison is not
1209 // signed. We can also fold a signed comparison if the shifted mask
1210 // value and the shifted comparison value are not negative.
1211 // These constraints may not be obvious, but we can prove that they
1212 // are correct using an SMT solver.
1213 if (!ICI.isSigned())
1216 ConstantInt *ShiftedAndCst =
1217 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1218 ConstantInt *ShiftedRHSCst =
1219 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1221 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1228 if (ShiftOpcode == Instruction::Shl)
1229 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1231 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1233 // Check to see if we are shifting out any of the bits being
1235 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1236 // If we shifted bits out, the fold is not going to work out.
1237 // As a special case, check to see if this means that the
1238 // result is always true or false now.
1239 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1240 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1241 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1242 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1244 ICI.setOperand(1, NewCst);
1245 Constant *NewAndCst;
1246 if (ShiftOpcode == Instruction::Shl)
1247 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1249 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1250 LHSI->setOperand(1, NewAndCst);
1251 LHSI->setOperand(0, Shift->getOperand(0));
1252 Worklist.Add(Shift); // Shift is dead.
1258 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1259 // preferable because it allows the C<<Y expression to be hoisted out
1260 // of a loop if Y is invariant and X is not.
1261 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1262 ICI.isEquality() && !Shift->isArithmeticShift() &&
1263 !isa<Constant>(Shift->getOperand(0))) {
1266 if (Shift->getOpcode() == Instruction::LShr) {
1267 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1269 // Insert a logical shift.
1270 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1273 // Compute X & (C << Y).
1275 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1277 ICI.setOperand(0, NewAnd);
1281 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1282 // bit set in (X & AndCst) will produce a result greater than RHSV.
1283 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1284 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1285 if ((NTZ < AndCst->getBitWidth()) &&
1286 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1287 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1288 Constant::getNullValue(RHS->getType()));
1292 // Try to optimize things like "A[i]&42 == 0" to index computations.
1293 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1294 if (GetElementPtrInst *GEP =
1295 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1296 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1297 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1298 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1299 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1300 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1305 // X & -C == -C -> X > u ~C
1306 // X & -C != -C -> X <= u ~C
1307 // iff C is a power of 2
1308 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1309 return new ICmpInst(
1310 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1311 : ICmpInst::ICMP_ULE,
1312 LHSI->getOperand(0), SubOne(RHS));
1315 case Instruction::Or: {
1316 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1319 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1320 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1321 // -> and (icmp eq P, null), (icmp eq Q, null).
1322 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1323 Constant::getNullValue(P->getType()));
1324 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1325 Constant::getNullValue(Q->getType()));
1327 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1328 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1330 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1336 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1337 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1340 // If this is a signed comparison to 0 and the mul is sign preserving,
1341 // use the mul LHS operand instead.
1342 ICmpInst::Predicate pred = ICI.getPredicate();
1343 if (isSignTest(pred, RHS) && !Val->isZero() &&
1344 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1345 return new ICmpInst(Val->isNegative() ?
1346 ICmpInst::getSwappedPredicate(pred) : pred,
1347 LHSI->getOperand(0),
1348 Constant::getNullValue(RHS->getType()));
1353 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1354 uint32_t TypeBits = RHSV.getBitWidth();
1355 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1358 // (1 << X) pred P2 -> X pred Log2(P2)
1359 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1360 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1361 ICmpInst::Predicate Pred = ICI.getPredicate();
1362 if (ICI.isUnsigned()) {
1363 if (!RHSVIsPowerOf2) {
1364 // (1 << X) < 30 -> X <= 4
1365 // (1 << X) <= 30 -> X <= 4
1366 // (1 << X) >= 30 -> X > 4
1367 // (1 << X) > 30 -> X > 4
1368 if (Pred == ICmpInst::ICMP_ULT)
1369 Pred = ICmpInst::ICMP_ULE;
1370 else if (Pred == ICmpInst::ICMP_UGE)
1371 Pred = ICmpInst::ICMP_UGT;
1373 unsigned RHSLog2 = RHSV.logBase2();
1375 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1376 // (1 << X) > 2147483648 -> X > 31 -> false
1377 // (1 << X) <= 2147483648 -> X <= 31 -> true
1378 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1379 if (RHSLog2 == TypeBits-1) {
1380 if (Pred == ICmpInst::ICMP_UGE)
1381 Pred = ICmpInst::ICMP_EQ;
1382 else if (Pred == ICmpInst::ICMP_UGT)
1383 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1384 else if (Pred == ICmpInst::ICMP_ULE)
1385 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1386 else if (Pred == ICmpInst::ICMP_ULT)
1387 Pred = ICmpInst::ICMP_NE;
1390 return new ICmpInst(Pred, X,
1391 ConstantInt::get(RHS->getType(), RHSLog2));
1392 } else if (ICI.isSigned()) {
1393 if (RHSV.isAllOnesValue()) {
1394 // (1 << X) <= -1 -> X == 31
1395 if (Pred == ICmpInst::ICMP_SLE)
1396 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1397 ConstantInt::get(RHS->getType(), TypeBits-1));
1399 // (1 << X) > -1 -> X != 31
1400 if (Pred == ICmpInst::ICMP_SGT)
1401 return new ICmpInst(ICmpInst::ICMP_NE, X,
1402 ConstantInt::get(RHS->getType(), TypeBits-1));
1404 // (1 << X) < 0 -> X == 31
1405 // (1 << X) <= 0 -> X == 31
1406 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1407 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1408 ConstantInt::get(RHS->getType(), TypeBits-1));
1410 // (1 << X) >= 0 -> X != 31
1411 // (1 << X) > 0 -> X != 31
1412 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1413 return new ICmpInst(ICmpInst::ICMP_NE, X,
1414 ConstantInt::get(RHS->getType(), TypeBits-1));
1416 } else if (ICI.isEquality()) {
1418 return new ICmpInst(
1419 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1421 return ReplaceInstUsesWith(
1422 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse()
1423 : Builder->getTrue());
1429 // Check that the shift amount is in range. If not, don't perform
1430 // undefined shifts. When the shift is visited it will be
1432 if (ShAmt->uge(TypeBits))
1435 if (ICI.isEquality()) {
1436 // If we are comparing against bits always shifted out, the
1437 // comparison cannot succeed.
1439 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1441 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1442 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1443 Constant *Cst = Builder->getInt1(IsICMP_NE);
1444 return ReplaceInstUsesWith(ICI, Cst);
1447 // If the shift is NUW, then it is just shifting out zeros, no need for an
1449 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1450 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1451 ConstantExpr::getLShr(RHS, ShAmt));
1453 // If the shift is NSW and we compare to 0, then it is just shifting out
1454 // sign bits, no need for an AND either.
1455 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1456 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1457 ConstantExpr::getLShr(RHS, ShAmt));
1459 if (LHSI->hasOneUse()) {
1460 // Otherwise strength reduce the shift into an and.
1461 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1462 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1463 TypeBits - ShAmtVal));
1466 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1467 return new ICmpInst(ICI.getPredicate(), And,
1468 ConstantExpr::getLShr(RHS, ShAmt));
1472 // If this is a signed comparison to 0 and the shift is sign preserving,
1473 // use the shift LHS operand instead.
1474 ICmpInst::Predicate pred = ICI.getPredicate();
1475 if (isSignTest(pred, RHS) &&
1476 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1477 return new ICmpInst(pred,
1478 LHSI->getOperand(0),
1479 Constant::getNullValue(RHS->getType()));
1481 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1482 bool TrueIfSigned = false;
1483 if (LHSI->hasOneUse() &&
1484 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1485 // (X << 31) <s 0 --> (X&1) != 0
1486 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1487 APInt::getOneBitSet(TypeBits,
1488 TypeBits-ShAmt->getZExtValue()-1));
1490 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1491 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1492 And, Constant::getNullValue(And->getType()));
1495 // Transform (icmp pred iM (shl iM %v, N), CI)
1496 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1497 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1498 // This enables to get rid of the shift in favor of a trunc which can be
1499 // free on the target. It has the additional benefit of comparing to a
1500 // smaller constant, which will be target friendly.
1501 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1502 if (LHSI->hasOneUse() &&
1503 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1504 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1505 Constant *NCI = ConstantExpr::getTrunc(
1506 ConstantExpr::getAShr(RHS,
1507 ConstantInt::get(RHS->getType(), Amt)),
1509 return new ICmpInst(ICI.getPredicate(),
1510 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1517 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1518 case Instruction::AShr: {
1519 // Handle equality comparisons of shift-by-constant.
1520 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1521 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1522 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1526 // Handle exact shr's.
1527 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1528 if (RHSV.isMinValue())
1529 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1534 case Instruction::SDiv:
1535 case Instruction::UDiv:
1536 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1537 // Fold this div into the comparison, producing a range check.
1538 // Determine, based on the divide type, what the range is being
1539 // checked. If there is an overflow on the low or high side, remember
1540 // it, otherwise compute the range [low, hi) bounding the new value.
1541 // See: InsertRangeTest above for the kinds of replacements possible.
1542 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1543 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1548 case Instruction::Sub: {
1549 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1551 const APInt &LHSV = LHSC->getValue();
1553 // C1-X <u C2 -> (X|(C2-1)) == C1
1554 // iff C1 & (C2-1) == C2-1
1555 // C2 is a power of 2
1556 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1557 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1558 return new ICmpInst(ICmpInst::ICMP_EQ,
1559 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1562 // C1-X >u C2 -> (X|C2) != C1
1563 // iff C1 & C2 == C2
1564 // C2+1 is a power of 2
1565 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1566 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1567 return new ICmpInst(ICmpInst::ICMP_NE,
1568 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1572 case Instruction::Add:
1573 // Fold: icmp pred (add X, C1), C2
1574 if (!ICI.isEquality()) {
1575 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1577 const APInt &LHSV = LHSC->getValue();
1579 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1582 if (ICI.isSigned()) {
1583 if (CR.getLower().isSignBit()) {
1584 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1585 Builder->getInt(CR.getUpper()));
1586 } else if (CR.getUpper().isSignBit()) {
1587 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1588 Builder->getInt(CR.getLower()));
1591 if (CR.getLower().isMinValue()) {
1592 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1593 Builder->getInt(CR.getUpper()));
1594 } else if (CR.getUpper().isMinValue()) {
1595 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1596 Builder->getInt(CR.getLower()));
1600 // X-C1 <u C2 -> (X & -C2) == C1
1601 // iff C1 & (C2-1) == 0
1602 // C2 is a power of 2
1603 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1604 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1605 return new ICmpInst(ICmpInst::ICMP_EQ,
1606 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1607 ConstantExpr::getNeg(LHSC));
1609 // X-C1 >u C2 -> (X & ~C2) != C1
1611 // C2+1 is a power of 2
1612 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1613 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1614 return new ICmpInst(ICmpInst::ICMP_NE,
1615 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1616 ConstantExpr::getNeg(LHSC));
1621 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1622 if (ICI.isEquality()) {
1623 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1625 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1626 // the second operand is a constant, simplify a bit.
1627 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1628 switch (BO->getOpcode()) {
1629 case Instruction::SRem:
1630 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1631 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1632 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1633 if (V.sgt(1) && V.isPowerOf2()) {
1635 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1637 return new ICmpInst(ICI.getPredicate(), NewRem,
1638 Constant::getNullValue(BO->getType()));
1642 case Instruction::Add:
1643 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1644 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1645 if (BO->hasOneUse())
1646 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1647 ConstantExpr::getSub(RHS, BOp1C));
1648 } else if (RHSV == 0) {
1649 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1650 // efficiently invertible, or if the add has just this one use.
1651 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1653 if (Value *NegVal = dyn_castNegVal(BOp1))
1654 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1655 if (Value *NegVal = dyn_castNegVal(BOp0))
1656 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1657 if (BO->hasOneUse()) {
1658 Value *Neg = Builder->CreateNeg(BOp1);
1660 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1664 case Instruction::Xor:
1665 // For the xor case, we can xor two constants together, eliminating
1666 // the explicit xor.
1667 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1668 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1669 ConstantExpr::getXor(RHS, BOC));
1670 } else if (RHSV == 0) {
1671 // Replace ((xor A, B) != 0) with (A != B)
1672 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1676 case Instruction::Sub:
1677 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1678 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1679 if (BO->hasOneUse())
1680 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1681 ConstantExpr::getSub(BOp0C, RHS));
1682 } else if (RHSV == 0) {
1683 // Replace ((sub A, B) != 0) with (A != B)
1684 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1688 case Instruction::Or:
1689 // If bits are being or'd in that are not present in the constant we
1690 // are comparing against, then the comparison could never succeed!
1691 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1692 Constant *NotCI = ConstantExpr::getNot(RHS);
1693 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1694 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1698 case Instruction::And:
1699 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1700 // If bits are being compared against that are and'd out, then the
1701 // comparison can never succeed!
1702 if ((RHSV & ~BOC->getValue()) != 0)
1703 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1705 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1706 if (RHS == BOC && RHSV.isPowerOf2())
1707 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1708 ICmpInst::ICMP_NE, LHSI,
1709 Constant::getNullValue(RHS->getType()));
1711 // Don't perform the following transforms if the AND has multiple uses
1712 if (!BO->hasOneUse())
1715 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1716 if (BOC->getValue().isSignBit()) {
1717 Value *X = BO->getOperand(0);
1718 Constant *Zero = Constant::getNullValue(X->getType());
1719 ICmpInst::Predicate pred = isICMP_NE ?
1720 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1721 return new ICmpInst(pred, X, Zero);
1724 // ((X & ~7) == 0) --> X < 8
1725 if (RHSV == 0 && isHighOnes(BOC)) {
1726 Value *X = BO->getOperand(0);
1727 Constant *NegX = ConstantExpr::getNeg(BOC);
1728 ICmpInst::Predicate pred = isICMP_NE ?
1729 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1730 return new ICmpInst(pred, X, NegX);
1734 case Instruction::Mul:
1735 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1736 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1737 // The trivial case (mul X, 0) is handled by InstSimplify
1738 // General case : (mul X, C) != 0 iff X != 0
1739 // (mul X, C) == 0 iff X == 0
1741 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1742 Constant::getNullValue(RHS->getType()));
1748 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1749 // Handle icmp {eq|ne} <intrinsic>, intcst.
1750 switch (II->getIntrinsicID()) {
1751 case Intrinsic::bswap:
1753 ICI.setOperand(0, II->getArgOperand(0));
1754 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1756 case Intrinsic::ctlz:
1757 case Intrinsic::cttz:
1758 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1759 if (RHSV == RHS->getType()->getBitWidth()) {
1761 ICI.setOperand(0, II->getArgOperand(0));
1762 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1766 case Intrinsic::ctpop:
1767 // popcount(A) == 0 -> A == 0 and likewise for !=
1768 if (RHS->isZero()) {
1770 ICI.setOperand(0, II->getArgOperand(0));
1771 ICI.setOperand(1, RHS);
1783 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1784 /// We only handle extending casts so far.
1786 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1787 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1788 Value *LHSCIOp = LHSCI->getOperand(0);
1789 Type *SrcTy = LHSCIOp->getType();
1790 Type *DestTy = LHSCI->getType();
1793 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1794 // integer type is the same size as the pointer type.
1795 if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
1796 DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1798 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1799 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1800 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1801 RHSOp = RHSC->getOperand(0);
1802 // If the pointer types don't match, insert a bitcast.
1803 if (LHSCIOp->getType() != RHSOp->getType())
1804 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1808 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1811 // The code below only handles extension cast instructions, so far.
1813 if (LHSCI->getOpcode() != Instruction::ZExt &&
1814 LHSCI->getOpcode() != Instruction::SExt)
1817 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1818 bool isSignedCmp = ICI.isSigned();
1820 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1821 // Not an extension from the same type?
1822 RHSCIOp = CI->getOperand(0);
1823 if (RHSCIOp->getType() != LHSCIOp->getType())
1826 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1827 // and the other is a zext), then we can't handle this.
1828 if (CI->getOpcode() != LHSCI->getOpcode())
1831 // Deal with equality cases early.
1832 if (ICI.isEquality())
1833 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1835 // A signed comparison of sign extended values simplifies into a
1836 // signed comparison.
1837 if (isSignedCmp && isSignedExt)
1838 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1840 // The other three cases all fold into an unsigned comparison.
1841 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1844 // If we aren't dealing with a constant on the RHS, exit early
1845 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1849 // Compute the constant that would happen if we truncated to SrcTy then
1850 // reextended to DestTy.
1851 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1852 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1855 // If the re-extended constant didn't change...
1857 // Deal with equality cases early.
1858 if (ICI.isEquality())
1859 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1861 // A signed comparison of sign extended values simplifies into a
1862 // signed comparison.
1863 if (isSignedExt && isSignedCmp)
1864 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1866 // The other three cases all fold into an unsigned comparison.
1867 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1870 // The re-extended constant changed so the constant cannot be represented
1871 // in the shorter type. Consequently, we cannot emit a simple comparison.
1872 // All the cases that fold to true or false will have already been handled
1873 // by SimplifyICmpInst, so only deal with the tricky case.
1875 if (isSignedCmp || !isSignedExt)
1878 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1879 // should have been folded away previously and not enter in here.
1881 // We're performing an unsigned comp with a sign extended value.
1882 // This is true if the input is >= 0. [aka >s -1]
1883 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1884 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1886 // Finally, return the value computed.
1887 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1888 return ReplaceInstUsesWith(ICI, Result);
1890 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1891 return BinaryOperator::CreateNot(Result);
1894 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1895 /// I = icmp ugt (add (add A, B), CI2), CI1
1896 /// If this is of the form:
1898 /// if (sum+128 >u 255)
1899 /// Then replace it with llvm.sadd.with.overflow.i8.
1901 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1902 ConstantInt *CI2, ConstantInt *CI1,
1904 // The transformation we're trying to do here is to transform this into an
1905 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1906 // with a narrower add, and discard the add-with-constant that is part of the
1907 // range check (if we can't eliminate it, this isn't profitable).
1909 // In order to eliminate the add-with-constant, the compare can be its only
1911 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1912 if (!AddWithCst->hasOneUse()) return 0;
1914 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1915 if (!CI2->getValue().isPowerOf2()) return 0;
1916 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1917 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1919 // The width of the new add formed is 1 more than the bias.
1922 // Check to see that CI1 is an all-ones value with NewWidth bits.
1923 if (CI1->getBitWidth() == NewWidth ||
1924 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1927 // This is only really a signed overflow check if the inputs have been
1928 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1929 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1930 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1931 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1932 IC.ComputeNumSignBits(B) < NeededSignBits)
1935 // In order to replace the original add with a narrower
1936 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1937 // and truncates that discard the high bits of the add. Verify that this is
1939 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1940 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1942 if (*UI == AddWithCst) continue;
1944 // Only accept truncates for now. We would really like a nice recursive
1945 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1946 // chain to see which bits of a value are actually demanded. If the
1947 // original add had another add which was then immediately truncated, we
1948 // could still do the transformation.
1949 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1951 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1954 // If the pattern matches, truncate the inputs to the narrower type and
1955 // use the sadd_with_overflow intrinsic to efficiently compute both the
1956 // result and the overflow bit.
1957 Module *M = I.getParent()->getParent()->getParent();
1959 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1960 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1963 InstCombiner::BuilderTy *Builder = IC.Builder;
1965 // Put the new code above the original add, in case there are any uses of the
1966 // add between the add and the compare.
1967 Builder->SetInsertPoint(OrigAdd);
1969 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1970 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1971 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1972 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1973 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1975 // The inner add was the result of the narrow add, zero extended to the
1976 // wider type. Replace it with the result computed by the intrinsic.
1977 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1979 // The original icmp gets replaced with the overflow value.
1980 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1983 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1985 // Don't bother doing this transformation for pointers, don't do it for
1987 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1989 // If the add is a constant expr, then we don't bother transforming it.
1990 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1991 if (OrigAdd == 0) return 0;
1993 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1995 // Put the new code above the original add, in case there are any uses of the
1996 // add between the add and the compare.
1997 InstCombiner::BuilderTy *Builder = IC.Builder;
1998 Builder->SetInsertPoint(OrigAdd);
2000 Module *M = I.getParent()->getParent()->getParent();
2001 Type *Ty = LHS->getType();
2002 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
2003 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
2004 Value *Add = Builder->CreateExtractValue(Call, 0);
2006 IC.ReplaceInstUsesWith(*OrigAdd, Add);
2008 // The original icmp gets replaced with the overflow value.
2009 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2012 // DemandedBitsLHSMask - When performing a comparison against a constant,
2013 // it is possible that not all the bits in the LHS are demanded. This helper
2014 // method computes the mask that IS demanded.
2015 static APInt DemandedBitsLHSMask(ICmpInst &I,
2016 unsigned BitWidth, bool isSignCheck) {
2018 return APInt::getSignBit(BitWidth);
2020 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2021 if (!CI) return APInt::getAllOnesValue(BitWidth);
2022 const APInt &RHS = CI->getValue();
2024 switch (I.getPredicate()) {
2025 // For a UGT comparison, we don't care about any bits that
2026 // correspond to the trailing ones of the comparand. The value of these
2027 // bits doesn't impact the outcome of the comparison, because any value
2028 // greater than the RHS must differ in a bit higher than these due to carry.
2029 case ICmpInst::ICMP_UGT: {
2030 unsigned trailingOnes = RHS.countTrailingOnes();
2031 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2035 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2036 // Any value less than the RHS must differ in a higher bit because of carries.
2037 case ICmpInst::ICMP_ULT: {
2038 unsigned trailingZeros = RHS.countTrailingZeros();
2039 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2044 return APInt::getAllOnesValue(BitWidth);
2049 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2050 /// should be swapped.
2051 /// The decision is based on how many times these two operands are reused
2052 /// as subtract operands and their positions in those instructions.
2053 /// The rational is that several architectures use the same instruction for
2054 /// both subtract and cmp, thus it is better if the order of those operands
2056 /// \return true if Op0 and Op1 should be swapped.
2057 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2058 const Value * Op1) {
2059 // Filter out pointer value as those cannot appears directly in subtract.
2060 // FIXME: we may want to go through inttoptrs or bitcasts.
2061 if (Op0->getType()->isPointerTy())
2063 // Count every uses of both Op0 and Op1 in a subtract.
2064 // Each time Op0 is the first operand, count -1: swapping is bad, the
2065 // subtract has already the same layout as the compare.
2066 // Each time Op0 is the second operand, count +1: swapping is good, the
2067 // subtract has a different layout as the compare.
2068 // At the end, if the benefit is greater than 0, Op0 should come second to
2069 // expose more CSE opportunities.
2070 int GlobalSwapBenefits = 0;
2071 for (Value::const_use_iterator UI = Op0->use_begin(), UIEnd = Op0->use_end(); UI != UIEnd; ++UI) {
2072 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(*UI);
2073 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2075 // If Op0 is the first argument, this is not beneficial to swap the
2077 int LocalSwapBenefits = -1;
2078 unsigned Op1Idx = 1;
2079 if (BinOp->getOperand(Op1Idx) == Op0) {
2081 LocalSwapBenefits = 1;
2083 if (BinOp->getOperand(Op1Idx) != Op1)
2085 GlobalSwapBenefits += LocalSwapBenefits;
2087 return GlobalSwapBenefits > 0;
2090 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2091 bool Changed = false;
2092 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2093 unsigned Op0Cplxity = getComplexity(Op0);
2094 unsigned Op1Cplxity = getComplexity(Op1);
2096 /// Orders the operands of the compare so that they are listed from most
2097 /// complex to least complex. This puts constants before unary operators,
2098 /// before binary operators.
2099 if (Op0Cplxity < Op1Cplxity ||
2100 (Op0Cplxity == Op1Cplxity &&
2101 swapMayExposeCSEOpportunities(Op0, Op1))) {
2103 std::swap(Op0, Op1);
2107 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL))
2108 return ReplaceInstUsesWith(I, V);
2110 // comparing -val or val with non-zero is the same as just comparing val
2111 // ie, abs(val) != 0 -> val != 0
2112 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2114 Value *Cond, *SelectTrue, *SelectFalse;
2115 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2116 m_Value(SelectFalse)))) {
2117 if (Value *V = dyn_castNegVal(SelectTrue)) {
2118 if (V == SelectFalse)
2119 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2121 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2122 if (V == SelectTrue)
2123 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2128 Type *Ty = Op0->getType();
2130 // icmp's with boolean values can always be turned into bitwise operations
2131 if (Ty->isIntegerTy(1)) {
2132 switch (I.getPredicate()) {
2133 default: llvm_unreachable("Invalid icmp instruction!");
2134 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2135 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2136 return BinaryOperator::CreateNot(Xor);
2138 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2139 return BinaryOperator::CreateXor(Op0, Op1);
2141 case ICmpInst::ICMP_UGT:
2142 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2144 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2145 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2146 return BinaryOperator::CreateAnd(Not, Op1);
2148 case ICmpInst::ICMP_SGT:
2149 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2151 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2152 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2153 return BinaryOperator::CreateAnd(Not, Op0);
2155 case ICmpInst::ICMP_UGE:
2156 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2158 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2159 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2160 return BinaryOperator::CreateOr(Not, Op1);
2162 case ICmpInst::ICMP_SGE:
2163 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2165 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2166 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2167 return BinaryOperator::CreateOr(Not, Op0);
2172 unsigned BitWidth = 0;
2173 if (Ty->isIntOrIntVectorTy())
2174 BitWidth = Ty->getScalarSizeInBits();
2175 else if (DL) // Pointers require DL info to get their size.
2176 BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
2178 bool isSignBit = false;
2180 // See if we are doing a comparison with a constant.
2181 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2182 Value *A = 0, *B = 0;
2184 // Match the following pattern, which is a common idiom when writing
2185 // overflow-safe integer arithmetic function. The source performs an
2186 // addition in wider type, and explicitly checks for overflow using
2187 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2188 // sadd_with_overflow intrinsic.
2190 // TODO: This could probably be generalized to handle other overflow-safe
2191 // operations if we worked out the formulas to compute the appropriate
2195 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2197 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2198 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2199 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2200 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2204 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2205 if (I.isEquality() && CI->isZero() &&
2206 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2207 // (icmp cond A B) if cond is equality
2208 return new ICmpInst(I.getPredicate(), A, B);
2211 // If we have an icmp le or icmp ge instruction, turn it into the
2212 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2213 // them being folded in the code below. The SimplifyICmpInst code has
2214 // already handled the edge cases for us, so we just assert on them.
2215 switch (I.getPredicate()) {
2217 case ICmpInst::ICMP_ULE:
2218 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2219 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2220 Builder->getInt(CI->getValue()+1));
2221 case ICmpInst::ICMP_SLE:
2222 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2223 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2224 Builder->getInt(CI->getValue()+1));
2225 case ICmpInst::ICMP_UGE:
2226 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2227 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2228 Builder->getInt(CI->getValue()-1));
2229 case ICmpInst::ICMP_SGE:
2230 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2231 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2232 Builder->getInt(CI->getValue()-1));
2235 // If this comparison is a normal comparison, it demands all
2236 // bits, if it is a sign bit comparison, it only demands the sign bit.
2238 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2241 // See if we can fold the comparison based on range information we can get
2242 // by checking whether bits are known to be zero or one in the input.
2243 if (BitWidth != 0) {
2244 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2245 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2247 if (SimplifyDemandedBits(I.getOperandUse(0),
2248 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2249 Op0KnownZero, Op0KnownOne, 0))
2251 if (SimplifyDemandedBits(I.getOperandUse(1),
2252 APInt::getAllOnesValue(BitWidth),
2253 Op1KnownZero, Op1KnownOne, 0))
2256 // Given the known and unknown bits, compute a range that the LHS could be
2257 // in. Compute the Min, Max and RHS values based on the known bits. For the
2258 // EQ and NE we use unsigned values.
2259 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2260 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2262 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2264 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2267 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2269 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2273 // If Min and Max are known to be the same, then SimplifyDemandedBits
2274 // figured out that the LHS is a constant. Just constant fold this now so
2275 // that code below can assume that Min != Max.
2276 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2277 return new ICmpInst(I.getPredicate(),
2278 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2279 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2280 return new ICmpInst(I.getPredicate(), Op0,
2281 ConstantInt::get(Op1->getType(), Op1Min));
2283 // Based on the range information we know about the LHS, see if we can
2284 // simplify this comparison. For example, (x&4) < 8 is always true.
2285 switch (I.getPredicate()) {
2286 default: llvm_unreachable("Unknown icmp opcode!");
2287 case ICmpInst::ICMP_EQ: {
2288 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2289 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2291 // If all bits are known zero except for one, then we know at most one
2292 // bit is set. If the comparison is against zero, then this is a check
2293 // to see if *that* bit is set.
2294 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2295 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2296 // If the LHS is an AND with the same constant, look through it.
2298 ConstantInt *LHSC = 0;
2299 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2300 LHSC->getValue() != Op0KnownZeroInverted)
2303 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2304 // then turn "((1 << x)&8) == 0" into "x != 3".
2306 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2307 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2308 return new ICmpInst(ICmpInst::ICMP_NE, X,
2309 ConstantInt::get(X->getType(), CmpVal));
2312 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2313 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2315 if (Op0KnownZeroInverted == 1 &&
2316 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2317 return new ICmpInst(ICmpInst::ICMP_NE, X,
2318 ConstantInt::get(X->getType(),
2319 CI->countTrailingZeros()));
2324 case ICmpInst::ICMP_NE: {
2325 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2326 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2328 // If all bits are known zero except for one, then we know at most one
2329 // bit is set. If the comparison is against zero, then this is a check
2330 // to see if *that* bit is set.
2331 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2332 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2333 // If the LHS is an AND with the same constant, look through it.
2335 ConstantInt *LHSC = 0;
2336 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2337 LHSC->getValue() != Op0KnownZeroInverted)
2340 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2341 // then turn "((1 << x)&8) != 0" into "x == 3".
2343 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2344 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2345 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2346 ConstantInt::get(X->getType(), CmpVal));
2349 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2350 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2352 if (Op0KnownZeroInverted == 1 &&
2353 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2354 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2355 ConstantInt::get(X->getType(),
2356 CI->countTrailingZeros()));
2361 case ICmpInst::ICMP_ULT:
2362 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2363 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2364 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2365 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2366 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2367 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2368 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2369 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2370 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2371 Builder->getInt(CI->getValue()-1));
2373 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2374 if (CI->isMinValue(true))
2375 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2376 Constant::getAllOnesValue(Op0->getType()));
2379 case ICmpInst::ICMP_UGT:
2380 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2381 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2382 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2383 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2385 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2386 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2387 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2388 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2389 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2390 Builder->getInt(CI->getValue()+1));
2392 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2393 if (CI->isMaxValue(true))
2394 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2395 Constant::getNullValue(Op0->getType()));
2398 case ICmpInst::ICMP_SLT:
2399 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2400 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2401 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2402 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2403 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2404 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2405 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2406 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2407 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2408 Builder->getInt(CI->getValue()-1));
2411 case ICmpInst::ICMP_SGT:
2412 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2413 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2414 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2415 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2417 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2418 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2419 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2420 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2421 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2422 Builder->getInt(CI->getValue()+1));
2425 case ICmpInst::ICMP_SGE:
2426 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2427 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2428 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2429 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2430 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2432 case ICmpInst::ICMP_SLE:
2433 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2434 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2435 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2436 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2437 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2439 case ICmpInst::ICMP_UGE:
2440 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2441 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2442 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2443 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2444 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2446 case ICmpInst::ICMP_ULE:
2447 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2448 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2449 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2450 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2451 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2455 // Turn a signed comparison into an unsigned one if both operands
2456 // are known to have the same sign.
2458 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2459 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2460 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2463 // Test if the ICmpInst instruction is used exclusively by a select as
2464 // part of a minimum or maximum operation. If so, refrain from doing
2465 // any other folding. This helps out other analyses which understand
2466 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2467 // and CodeGen. And in this case, at least one of the comparison
2468 // operands has at least one user besides the compare (the select),
2469 // which would often largely negate the benefit of folding anyway.
2471 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2472 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2473 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2476 // See if we are doing a comparison between a constant and an instruction that
2477 // can be folded into the comparison.
2478 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2479 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2480 // instruction, see if that instruction also has constants so that the
2481 // instruction can be folded into the icmp
2482 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2483 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2487 // Handle icmp with constant (but not simple integer constant) RHS
2488 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2489 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2490 switch (LHSI->getOpcode()) {
2491 case Instruction::GetElementPtr:
2492 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2493 if (RHSC->isNullValue() &&
2494 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2495 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2496 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2498 case Instruction::PHI:
2499 // Only fold icmp into the PHI if the phi and icmp are in the same
2500 // block. If in the same block, we're encouraging jump threading. If
2501 // not, we are just pessimizing the code by making an i1 phi.
2502 if (LHSI->getParent() == I.getParent())
2503 if (Instruction *NV = FoldOpIntoPhi(I))
2506 case Instruction::Select: {
2507 // If either operand of the select is a constant, we can fold the
2508 // comparison into the select arms, which will cause one to be
2509 // constant folded and the select turned into a bitwise or.
2510 Value *Op1 = 0, *Op2 = 0;
2511 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2512 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2513 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2514 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2516 // We only want to perform this transformation if it will not lead to
2517 // additional code. This is true if either both sides of the select
2518 // fold to a constant (in which case the icmp is replaced with a select
2519 // which will usually simplify) or this is the only user of the
2520 // select (in which case we are trading a select+icmp for a simpler
2522 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2524 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2527 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2529 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2533 case Instruction::IntToPtr:
2534 // icmp pred inttoptr(X), null -> icmp pred X, 0
2535 if (RHSC->isNullValue() && DL &&
2536 DL->getIntPtrType(RHSC->getType()) ==
2537 LHSI->getOperand(0)->getType())
2538 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2539 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2542 case Instruction::Load:
2543 // Try to optimize things like "A[i] > 4" to index computations.
2544 if (GetElementPtrInst *GEP =
2545 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2546 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2547 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2548 !cast<LoadInst>(LHSI)->isVolatile())
2549 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2556 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2557 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2558 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2560 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2561 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2562 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2565 // Test to see if the operands of the icmp are casted versions of other
2566 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2568 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2569 if (Op0->getType()->isPointerTy() &&
2570 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2571 // We keep moving the cast from the left operand over to the right
2572 // operand, where it can often be eliminated completely.
2573 Op0 = CI->getOperand(0);
2575 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2576 // so eliminate it as well.
2577 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2578 Op1 = CI2->getOperand(0);
2580 // If Op1 is a constant, we can fold the cast into the constant.
2581 if (Op0->getType() != Op1->getType()) {
2582 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2583 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2585 // Otherwise, cast the RHS right before the icmp
2586 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2589 return new ICmpInst(I.getPredicate(), Op0, Op1);
2593 if (isa<CastInst>(Op0)) {
2594 // Handle the special case of: icmp (cast bool to X), <cst>
2595 // This comes up when you have code like
2598 // For generality, we handle any zero-extension of any operand comparison
2599 // with a constant or another cast from the same type.
2600 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2601 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2605 // Special logic for binary operators.
2606 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2607 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2609 CmpInst::Predicate Pred = I.getPredicate();
2610 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2611 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2612 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2613 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2614 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2615 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2616 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2617 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2618 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2620 // Analyze the case when either Op0 or Op1 is an add instruction.
2621 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2622 Value *A = 0, *B = 0, *C = 0, *D = 0;
2623 if (BO0 && BO0->getOpcode() == Instruction::Add)
2624 A = BO0->getOperand(0), B = BO0->getOperand(1);
2625 if (BO1 && BO1->getOpcode() == Instruction::Add)
2626 C = BO1->getOperand(0), D = BO1->getOperand(1);
2628 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2629 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2630 return new ICmpInst(Pred, A == Op1 ? B : A,
2631 Constant::getNullValue(Op1->getType()));
2633 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2634 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2635 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2638 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2639 if (A && C && (A == C || A == D || B == C || B == D) &&
2640 NoOp0WrapProblem && NoOp1WrapProblem &&
2641 // Try not to increase register pressure.
2642 BO0->hasOneUse() && BO1->hasOneUse()) {
2643 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2646 // C + B == C + D -> B == D
2649 } else if (A == D) {
2650 // D + B == C + D -> B == C
2653 } else if (B == C) {
2654 // A + C == C + D -> A == D
2659 // A + D == C + D -> A == C
2663 return new ICmpInst(Pred, Y, Z);
2666 // icmp slt (X + -1), Y -> icmp sle X, Y
2667 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2668 match(B, m_AllOnes()))
2669 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2671 // icmp sge (X + -1), Y -> icmp sgt X, Y
2672 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2673 match(B, m_AllOnes()))
2674 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2676 // icmp sle (X + 1), Y -> icmp slt X, Y
2677 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
2679 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2681 // icmp sgt (X + 1), Y -> icmp sge X, Y
2682 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
2684 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2686 // if C1 has greater magnitude than C2:
2687 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2688 // s.t. C3 = C1 - C2
2690 // if C2 has greater magnitude than C1:
2691 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2692 // s.t. C3 = C2 - C1
2693 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2694 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2695 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2696 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2697 const APInt &AP1 = C1->getValue();
2698 const APInt &AP2 = C2->getValue();
2699 if (AP1.isNegative() == AP2.isNegative()) {
2700 APInt AP1Abs = C1->getValue().abs();
2701 APInt AP2Abs = C2->getValue().abs();
2702 if (AP1Abs.uge(AP2Abs)) {
2703 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2704 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2705 return new ICmpInst(Pred, NewAdd, C);
2707 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2708 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2709 return new ICmpInst(Pred, A, NewAdd);
2715 // Analyze the case when either Op0 or Op1 is a sub instruction.
2716 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2717 A = 0; B = 0; C = 0; D = 0;
2718 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2719 A = BO0->getOperand(0), B = BO0->getOperand(1);
2720 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2721 C = BO1->getOperand(0), D = BO1->getOperand(1);
2723 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2724 if (A == Op1 && NoOp0WrapProblem)
2725 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2727 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2728 if (C == Op0 && NoOp1WrapProblem)
2729 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2731 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2732 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2733 // Try not to increase register pressure.
2734 BO0->hasOneUse() && BO1->hasOneUse())
2735 return new ICmpInst(Pred, A, C);
2737 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2738 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2739 // Try not to increase register pressure.
2740 BO0->hasOneUse() && BO1->hasOneUse())
2741 return new ICmpInst(Pred, D, B);
2743 BinaryOperator *SRem = NULL;
2744 // icmp (srem X, Y), Y
2745 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2746 Op1 == BO0->getOperand(1))
2748 // icmp Y, (srem X, Y)
2749 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2750 Op0 == BO1->getOperand(1))
2753 // We don't check hasOneUse to avoid increasing register pressure because
2754 // the value we use is the same value this instruction was already using.
2755 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2757 case ICmpInst::ICMP_EQ:
2758 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2759 case ICmpInst::ICMP_NE:
2760 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2761 case ICmpInst::ICMP_SGT:
2762 case ICmpInst::ICMP_SGE:
2763 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2764 Constant::getAllOnesValue(SRem->getType()));
2765 case ICmpInst::ICMP_SLT:
2766 case ICmpInst::ICMP_SLE:
2767 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2768 Constant::getNullValue(SRem->getType()));
2772 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2773 BO0->hasOneUse() && BO1->hasOneUse() &&
2774 BO0->getOperand(1) == BO1->getOperand(1)) {
2775 switch (BO0->getOpcode()) {
2777 case Instruction::Add:
2778 case Instruction::Sub:
2779 case Instruction::Xor:
2780 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2781 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2782 BO1->getOperand(0));
2783 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2784 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2785 if (CI->getValue().isSignBit()) {
2786 ICmpInst::Predicate Pred = I.isSigned()
2787 ? I.getUnsignedPredicate()
2788 : I.getSignedPredicate();
2789 return new ICmpInst(Pred, BO0->getOperand(0),
2790 BO1->getOperand(0));
2793 if (CI->isMaxValue(true)) {
2794 ICmpInst::Predicate Pred = I.isSigned()
2795 ? I.getUnsignedPredicate()
2796 : I.getSignedPredicate();
2797 Pred = I.getSwappedPredicate(Pred);
2798 return new ICmpInst(Pred, BO0->getOperand(0),
2799 BO1->getOperand(0));
2803 case Instruction::Mul:
2804 if (!I.isEquality())
2807 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2808 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2809 // Mask = -1 >> count-trailing-zeros(Cst).
2810 if (!CI->isZero() && !CI->isOne()) {
2811 const APInt &AP = CI->getValue();
2812 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2813 APInt::getLowBitsSet(AP.getBitWidth(),
2815 AP.countTrailingZeros()));
2816 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2817 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2818 return new ICmpInst(I.getPredicate(), And1, And2);
2822 case Instruction::UDiv:
2823 case Instruction::LShr:
2827 case Instruction::SDiv:
2828 case Instruction::AShr:
2829 if (!BO0->isExact() || !BO1->isExact())
2831 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2832 BO1->getOperand(0));
2833 case Instruction::Shl: {
2834 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2835 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2838 if (!NSW && I.isSigned())
2840 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2841 BO1->getOperand(0));
2848 // Transform (A & ~B) == 0 --> (A & B) != 0
2849 // and (A & ~B) != 0 --> (A & B) == 0
2850 // if A is a power of 2.
2851 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2852 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
2853 return new ICmpInst(I.getInversePredicate(),
2854 Builder->CreateAnd(A, B),
2857 // ~x < ~y --> y < x
2858 // ~x < cst --> ~cst < x
2859 if (match(Op0, m_Not(m_Value(A)))) {
2860 if (match(Op1, m_Not(m_Value(B))))
2861 return new ICmpInst(I.getPredicate(), B, A);
2862 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2863 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2866 // (a+b) <u a --> llvm.uadd.with.overflow.
2867 // (a+b) <u b --> llvm.uadd.with.overflow.
2868 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2869 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2870 (Op1 == A || Op1 == B))
2871 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2874 // a >u (a+b) --> llvm.uadd.with.overflow.
2875 // b >u (a+b) --> llvm.uadd.with.overflow.
2876 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2877 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2878 (Op0 == A || Op0 == B))
2879 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2883 if (I.isEquality()) {
2884 Value *A, *B, *C, *D;
2886 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2887 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2888 Value *OtherVal = A == Op1 ? B : A;
2889 return new ICmpInst(I.getPredicate(), OtherVal,
2890 Constant::getNullValue(A->getType()));
2893 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2894 // A^c1 == C^c2 --> A == C^(c1^c2)
2895 ConstantInt *C1, *C2;
2896 if (match(B, m_ConstantInt(C1)) &&
2897 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2898 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
2899 Value *Xor = Builder->CreateXor(C, NC);
2900 return new ICmpInst(I.getPredicate(), A, Xor);
2903 // A^B == A^D -> B == D
2904 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2905 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2906 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2907 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2911 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2912 (A == Op0 || B == Op0)) {
2913 // A == (A^B) -> B == 0
2914 Value *OtherVal = A == Op0 ? B : A;
2915 return new ICmpInst(I.getPredicate(), OtherVal,
2916 Constant::getNullValue(A->getType()));
2919 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2920 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2921 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2922 Value *X = 0, *Y = 0, *Z = 0;
2925 X = B; Y = D; Z = A;
2926 } else if (A == D) {
2927 X = B; Y = C; Z = A;
2928 } else if (B == C) {
2929 X = A; Y = D; Z = B;
2930 } else if (B == D) {
2931 X = A; Y = C; Z = B;
2934 if (X) { // Build (X^Y) & Z
2935 Op1 = Builder->CreateXor(X, Y);
2936 Op1 = Builder->CreateAnd(Op1, Z);
2937 I.setOperand(0, Op1);
2938 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2943 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2944 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2946 if ((Op0->hasOneUse() &&
2947 match(Op0, m_ZExt(m_Value(A))) &&
2948 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2949 (Op1->hasOneUse() &&
2950 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2951 match(Op1, m_ZExt(m_Value(A))))) {
2952 APInt Pow2 = Cst1->getValue() + 1;
2953 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2954 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2955 return new ICmpInst(I.getPredicate(), A,
2956 Builder->CreateTrunc(B, A->getType()));
2959 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
2960 // For lshr and ashr pairs.
2961 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
2962 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
2963 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
2964 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
2965 unsigned TypeBits = Cst1->getBitWidth();
2966 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
2967 if (ShAmt < TypeBits && ShAmt != 0) {
2968 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
2969 ? ICmpInst::ICMP_UGE
2970 : ICmpInst::ICMP_ULT;
2971 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
2972 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
2973 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
2977 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2978 // "icmp (and X, mask), cst"
2980 if (Op0->hasOneUse() &&
2981 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2982 m_ConstantInt(ShAmt))))) &&
2983 match(Op1, m_ConstantInt(Cst1)) &&
2984 // Only do this when A has multiple uses. This is most important to do
2985 // when it exposes other optimizations.
2987 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2989 if (ShAmt < ASize) {
2991 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2994 APInt CmpV = Cst1->getValue().zext(ASize);
2997 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2998 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3004 Value *X; ConstantInt *Cst;
3006 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3007 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3010 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3011 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3013 return Changed ? &I : 0;
3016 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3018 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3021 if (!isa<ConstantFP>(RHSC)) return 0;
3022 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3024 // Get the width of the mantissa. We don't want to hack on conversions that
3025 // might lose information from the integer, e.g. "i64 -> float"
3026 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3027 if (MantissaWidth == -1) return 0; // Unknown.
3029 // Check to see that the input is converted from an integer type that is small
3030 // enough that preserves all bits. TODO: check here for "known" sign bits.
3031 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3032 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
3034 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3035 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3039 // If the conversion would lose info, don't hack on this.
3040 if ((int)InputSize > MantissaWidth)
3043 // Otherwise, we can potentially simplify the comparison. We know that it
3044 // will always come through as an integer value and we know the constant is
3045 // not a NAN (it would have been previously simplified).
3046 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3048 ICmpInst::Predicate Pred;
3049 switch (I.getPredicate()) {
3050 default: llvm_unreachable("Unexpected predicate!");
3051 case FCmpInst::FCMP_UEQ:
3052 case FCmpInst::FCMP_OEQ:
3053 Pred = ICmpInst::ICMP_EQ;
3055 case FCmpInst::FCMP_UGT:
3056 case FCmpInst::FCMP_OGT:
3057 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3059 case FCmpInst::FCMP_UGE:
3060 case FCmpInst::FCMP_OGE:
3061 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3063 case FCmpInst::FCMP_ULT:
3064 case FCmpInst::FCMP_OLT:
3065 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3067 case FCmpInst::FCMP_ULE:
3068 case FCmpInst::FCMP_OLE:
3069 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3071 case FCmpInst::FCMP_UNE:
3072 case FCmpInst::FCMP_ONE:
3073 Pred = ICmpInst::ICMP_NE;
3075 case FCmpInst::FCMP_ORD:
3076 return ReplaceInstUsesWith(I, Builder->getTrue());
3077 case FCmpInst::FCMP_UNO:
3078 return ReplaceInstUsesWith(I, Builder->getFalse());
3081 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3083 // Now we know that the APFloat is a normal number, zero or inf.
3085 // See if the FP constant is too large for the integer. For example,
3086 // comparing an i8 to 300.0.
3087 unsigned IntWidth = IntTy->getScalarSizeInBits();
3090 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3091 // and large values.
3092 APFloat SMax(RHS.getSemantics());
3093 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3094 APFloat::rmNearestTiesToEven);
3095 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3096 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3097 Pred == ICmpInst::ICMP_SLE)
3098 return ReplaceInstUsesWith(I, Builder->getTrue());
3099 return ReplaceInstUsesWith(I, Builder->getFalse());
3102 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3103 // +INF and large values.
3104 APFloat UMax(RHS.getSemantics());
3105 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3106 APFloat::rmNearestTiesToEven);
3107 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3108 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3109 Pred == ICmpInst::ICMP_ULE)
3110 return ReplaceInstUsesWith(I, Builder->getTrue());
3111 return ReplaceInstUsesWith(I, Builder->getFalse());
3116 // See if the RHS value is < SignedMin.
3117 APFloat SMin(RHS.getSemantics());
3118 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3119 APFloat::rmNearestTiesToEven);
3120 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3121 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3122 Pred == ICmpInst::ICMP_SGE)
3123 return ReplaceInstUsesWith(I, Builder->getTrue());
3124 return ReplaceInstUsesWith(I, Builder->getFalse());
3127 // See if the RHS value is < UnsignedMin.
3128 APFloat SMin(RHS.getSemantics());
3129 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3130 APFloat::rmNearestTiesToEven);
3131 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3132 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3133 Pred == ICmpInst::ICMP_UGE)
3134 return ReplaceInstUsesWith(I, Builder->getTrue());
3135 return ReplaceInstUsesWith(I, Builder->getFalse());
3139 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3140 // [0, UMAX], but it may still be fractional. See if it is fractional by
3141 // casting the FP value to the integer value and back, checking for equality.
3142 // Don't do this for zero, because -0.0 is not fractional.
3143 Constant *RHSInt = LHSUnsigned
3144 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3145 : ConstantExpr::getFPToSI(RHSC, IntTy);
3146 if (!RHS.isZero()) {
3147 bool Equal = LHSUnsigned
3148 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3149 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3151 // If we had a comparison against a fractional value, we have to adjust
3152 // the compare predicate and sometimes the value. RHSC is rounded towards
3153 // zero at this point.
3155 default: llvm_unreachable("Unexpected integer comparison!");
3156 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3157 return ReplaceInstUsesWith(I, Builder->getTrue());
3158 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3159 return ReplaceInstUsesWith(I, Builder->getFalse());
3160 case ICmpInst::ICMP_ULE:
3161 // (float)int <= 4.4 --> int <= 4
3162 // (float)int <= -4.4 --> false
3163 if (RHS.isNegative())
3164 return ReplaceInstUsesWith(I, Builder->getFalse());
3166 case ICmpInst::ICMP_SLE:
3167 // (float)int <= 4.4 --> int <= 4
3168 // (float)int <= -4.4 --> int < -4
3169 if (RHS.isNegative())
3170 Pred = ICmpInst::ICMP_SLT;
3172 case ICmpInst::ICMP_ULT:
3173 // (float)int < -4.4 --> false
3174 // (float)int < 4.4 --> int <= 4
3175 if (RHS.isNegative())
3176 return ReplaceInstUsesWith(I, Builder->getFalse());
3177 Pred = ICmpInst::ICMP_ULE;
3179 case ICmpInst::ICMP_SLT:
3180 // (float)int < -4.4 --> int < -4
3181 // (float)int < 4.4 --> int <= 4
3182 if (!RHS.isNegative())
3183 Pred = ICmpInst::ICMP_SLE;
3185 case ICmpInst::ICMP_UGT:
3186 // (float)int > 4.4 --> int > 4
3187 // (float)int > -4.4 --> true
3188 if (RHS.isNegative())
3189 return ReplaceInstUsesWith(I, Builder->getTrue());
3191 case ICmpInst::ICMP_SGT:
3192 // (float)int > 4.4 --> int > 4
3193 // (float)int > -4.4 --> int >= -4
3194 if (RHS.isNegative())
3195 Pred = ICmpInst::ICMP_SGE;
3197 case ICmpInst::ICMP_UGE:
3198 // (float)int >= -4.4 --> true
3199 // (float)int >= 4.4 --> int > 4
3200 if (RHS.isNegative())
3201 return ReplaceInstUsesWith(I, Builder->getTrue());
3202 Pred = ICmpInst::ICMP_UGT;
3204 case ICmpInst::ICMP_SGE:
3205 // (float)int >= -4.4 --> int >= -4
3206 // (float)int >= 4.4 --> int > 4
3207 if (!RHS.isNegative())
3208 Pred = ICmpInst::ICMP_SGT;
3214 // Lower this FP comparison into an appropriate integer version of the
3216 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3219 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3220 bool Changed = false;
3222 /// Orders the operands of the compare so that they are listed from most
3223 /// complex to least complex. This puts constants before unary operators,
3224 /// before binary operators.
3225 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3230 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3232 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL))
3233 return ReplaceInstUsesWith(I, V);
3235 // Simplify 'fcmp pred X, X'
3237 switch (I.getPredicate()) {
3238 default: llvm_unreachable("Unknown predicate!");
3239 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3240 case FCmpInst::FCMP_ULT: // True if unordered or less than
3241 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3242 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3243 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3244 I.setPredicate(FCmpInst::FCMP_UNO);
3245 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3248 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3249 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3250 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3251 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3252 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3253 I.setPredicate(FCmpInst::FCMP_ORD);
3254 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3259 // Handle fcmp with constant RHS
3260 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3261 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3262 switch (LHSI->getOpcode()) {
3263 case Instruction::FPExt: {
3264 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3265 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3266 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3270 const fltSemantics *Sem;
3271 // FIXME: This shouldn't be here.
3272 if (LHSExt->getSrcTy()->isHalfTy())
3273 Sem = &APFloat::IEEEhalf;
3274 else if (LHSExt->getSrcTy()->isFloatTy())
3275 Sem = &APFloat::IEEEsingle;
3276 else if (LHSExt->getSrcTy()->isDoubleTy())
3277 Sem = &APFloat::IEEEdouble;
3278 else if (LHSExt->getSrcTy()->isFP128Ty())
3279 Sem = &APFloat::IEEEquad;
3280 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3281 Sem = &APFloat::x87DoubleExtended;
3282 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3283 Sem = &APFloat::PPCDoubleDouble;
3288 APFloat F = RHSF->getValueAPF();
3289 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3291 // Avoid lossy conversions and denormals. Zero is a special case
3292 // that's OK to convert.
3296 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3297 APFloat::cmpLessThan) || Fabs.isZero()))
3299 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3300 ConstantFP::get(RHSC->getContext(), F));
3303 case Instruction::PHI:
3304 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3305 // block. If in the same block, we're encouraging jump threading. If
3306 // not, we are just pessimizing the code by making an i1 phi.
3307 if (LHSI->getParent() == I.getParent())
3308 if (Instruction *NV = FoldOpIntoPhi(I))
3311 case Instruction::SIToFP:
3312 case Instruction::UIToFP:
3313 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3316 case Instruction::FSub: {
3317 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3319 if (match(LHSI, m_FNeg(m_Value(Op))))
3320 return new FCmpInst(I.getSwappedPredicate(), Op,
3321 ConstantExpr::getFNeg(RHSC));
3324 case Instruction::Load:
3325 if (GetElementPtrInst *GEP =
3326 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3327 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3328 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3329 !cast<LoadInst>(LHSI)->isVolatile())
3330 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3334 case Instruction::Call: {
3335 CallInst *CI = cast<CallInst>(LHSI);
3337 // Various optimization for fabs compared with zero.
3338 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3339 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3341 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3342 Func == LibFunc::fabsl) {
3343 switch (I.getPredicate()) {
3345 // fabs(x) < 0 --> false
3346 case FCmpInst::FCMP_OLT:
3347 return ReplaceInstUsesWith(I, Builder->getFalse());
3348 // fabs(x) > 0 --> x != 0
3349 case FCmpInst::FCMP_OGT:
3350 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3352 // fabs(x) <= 0 --> x == 0
3353 case FCmpInst::FCMP_OLE:
3354 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3356 // fabs(x) >= 0 --> !isnan(x)
3357 case FCmpInst::FCMP_OGE:
3358 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3360 // fabs(x) == 0 --> x == 0
3361 // fabs(x) != 0 --> x != 0
3362 case FCmpInst::FCMP_OEQ:
3363 case FCmpInst::FCMP_UEQ:
3364 case FCmpInst::FCMP_ONE:
3365 case FCmpInst::FCMP_UNE:
3366 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3375 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3377 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3378 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3380 // fcmp (fpext x), (fpext y) -> fcmp x, y
3381 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3382 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3383 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3384 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3385 RHSExt->getOperand(0));
3387 return Changed ? &I : 0;