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/IntrinsicInst.h"
20 #include "llvm/Support/ConstantRange.h"
21 #include "llvm/Support/GetElementPtrTypeIterator.h"
22 #include "llvm/Support/PatternMatch.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 /// AddOne - Add one to a ConstantInt
32 static Constant *AddOne(Constant *C) {
33 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
35 /// SubOne - Subtract one from a ConstantInt
36 static Constant *SubOne(Constant *C) {
37 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
40 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
41 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
44 static bool HasAddOverflow(ConstantInt *Result,
45 ConstantInt *In1, ConstantInt *In2,
48 return Result->getValue().ult(In1->getValue());
50 if (In2->isNegative())
51 return Result->getValue().sgt(In1->getValue());
52 return Result->getValue().slt(In1->getValue());
55 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
56 /// overflowed for this type.
57 static bool AddWithOverflow(Constant *&Result, Constant *In1,
58 Constant *In2, bool IsSigned = false) {
59 Result = ConstantExpr::getAdd(In1, In2);
61 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
62 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
63 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
64 if (HasAddOverflow(ExtractElement(Result, Idx),
65 ExtractElement(In1, Idx),
66 ExtractElement(In2, Idx),
73 return HasAddOverflow(cast<ConstantInt>(Result),
74 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
78 static bool HasSubOverflow(ConstantInt *Result,
79 ConstantInt *In1, ConstantInt *In2,
82 return Result->getValue().ugt(In1->getValue());
84 if (In2->isNegative())
85 return Result->getValue().slt(In1->getValue());
87 return Result->getValue().sgt(In1->getValue());
90 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
91 /// overflowed for this type.
92 static bool SubWithOverflow(Constant *&Result, Constant *In1,
93 Constant *In2, bool IsSigned = false) {
94 Result = ConstantExpr::getSub(In1, In2);
96 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
97 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
98 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
99 if (HasSubOverflow(ExtractElement(Result, Idx),
100 ExtractElement(In1, Idx),
101 ExtractElement(In2, Idx),
108 return HasSubOverflow(cast<ConstantInt>(Result),
109 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
113 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
114 /// comparison only checks the sign bit. If it only checks the sign bit, set
115 /// TrueIfSigned if the result of the comparison is true when the input value is
117 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
118 bool &TrueIfSigned) {
120 case ICmpInst::ICMP_SLT: // True if LHS s< 0
122 return RHS->isZero();
123 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
125 return RHS->isAllOnesValue();
126 case ICmpInst::ICMP_SGT: // True if LHS s> -1
127 TrueIfSigned = false;
128 return RHS->isAllOnesValue();
129 case ICmpInst::ICMP_UGT:
130 // True if LHS u> RHS and RHS == high-bit-mask - 1
132 return RHS->isMaxValue(true);
133 case ICmpInst::ICMP_UGE:
134 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
136 return RHS->getValue().isSignBit();
142 /// Returns true if the exploded icmp can be expressed as a signed comparison
143 /// to zero and updates the predicate accordingly.
144 /// The signedness of the comparison is preserved.
145 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
146 if (!ICmpInst::isSigned(pred))
150 return ICmpInst::isRelational(pred);
153 if (pred == ICmpInst::ICMP_SLT) {
154 pred = ICmpInst::ICMP_SLE;
157 } else if (RHS->isAllOnesValue()) {
158 if (pred == ICmpInst::ICMP_SGT) {
159 pred = ICmpInst::ICMP_SGE;
167 // isHighOnes - Return true if the constant is of the form 1+0+.
168 // This is the same as lowones(~X).
169 static bool isHighOnes(const ConstantInt *CI) {
170 return (~CI->getValue() + 1).isPowerOf2();
173 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
174 /// set of known zero and one bits, compute the maximum and minimum values that
175 /// could have the specified known zero and known one bits, returning them in
177 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
178 const APInt& KnownOne,
179 APInt& Min, APInt& Max) {
180 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
181 KnownZero.getBitWidth() == Min.getBitWidth() &&
182 KnownZero.getBitWidth() == Max.getBitWidth() &&
183 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
184 APInt UnknownBits = ~(KnownZero|KnownOne);
186 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
187 // bit if it is unknown.
189 Max = KnownOne|UnknownBits;
191 if (UnknownBits.isNegative()) { // Sign bit is unknown
192 Min.setBit(Min.getBitWidth()-1);
193 Max.clearBit(Max.getBitWidth()-1);
197 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
198 // a set of known zero and one bits, compute the maximum and minimum values that
199 // could have the specified known zero and known one bits, returning them in
201 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
202 const APInt &KnownOne,
203 APInt &Min, APInt &Max) {
204 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
205 KnownZero.getBitWidth() == Min.getBitWidth() &&
206 KnownZero.getBitWidth() == Max.getBitWidth() &&
207 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
208 APInt UnknownBits = ~(KnownZero|KnownOne);
210 // The minimum value is when the unknown bits are all zeros.
212 // The maximum value is when the unknown bits are all ones.
213 Max = KnownOne|UnknownBits;
218 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
219 /// cmp pred (load (gep GV, ...)), cmpcst
220 /// where GV is a global variable with a constant initializer. Try to simplify
221 /// this into some simple computation that does not need the load. For example
222 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
224 /// If AndCst is non-null, then the loaded value is masked with that constant
225 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
226 Instruction *InstCombiner::
227 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
228 CmpInst &ICI, ConstantInt *AndCst) {
229 // We need TD information to know the pointer size unless this is inbounds.
230 if (!GEP->isInBounds() && TD == 0) return 0;
232 Constant *Init = GV->getInitializer();
233 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
236 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
237 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays.
239 // There are many forms of this optimization we can handle, for now, just do
240 // the simple index into a single-dimensional array.
242 // Require: GEP GV, 0, i {{, constant indices}}
243 if (GEP->getNumOperands() < 3 ||
244 !isa<ConstantInt>(GEP->getOperand(1)) ||
245 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
246 isa<Constant>(GEP->getOperand(2)))
249 // Check that indices after the variable are constants and in-range for the
250 // type they index. Collect the indices. This is typically for arrays of
252 SmallVector<unsigned, 4> LaterIndices;
254 Type *EltTy = Init->getType()->getArrayElementType();
255 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
256 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
257 if (Idx == 0) return 0; // Variable index.
259 uint64_t IdxVal = Idx->getZExtValue();
260 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
262 if (StructType *STy = dyn_cast<StructType>(EltTy))
263 EltTy = STy->getElementType(IdxVal);
264 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
265 if (IdxVal >= ATy->getNumElements()) return 0;
266 EltTy = ATy->getElementType();
268 return 0; // Unknown type.
271 LaterIndices.push_back(IdxVal);
274 enum { Overdefined = -3, Undefined = -2 };
276 // Variables for our state machines.
278 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
279 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
280 // and 87 is the second (and last) index. FirstTrueElement is -2 when
281 // undefined, otherwise set to the first true element. SecondTrueElement is
282 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
283 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
285 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
286 // form "i != 47 & i != 87". Same state transitions as for true elements.
287 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
289 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
290 /// define a state machine that triggers for ranges of values that the index
291 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
292 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
293 /// index in the range (inclusive). We use -2 for undefined here because we
294 /// use relative comparisons and don't want 0-1 to match -1.
295 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
297 // MagicBitvector - This is a magic bitvector where we set a bit if the
298 // comparison is true for element 'i'. If there are 64 elements or less in
299 // the array, this will fully represent all the comparison results.
300 uint64_t MagicBitvector = 0;
303 // Scan the array and see if one of our patterns matches.
304 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
305 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
306 Constant *Elt = Init->getAggregateElement(i);
307 if (Elt == 0) return 0;
309 // If this is indexing an array of structures, get the structure element.
310 if (!LaterIndices.empty())
311 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
313 // If the element is masked, handle it.
314 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
316 // Find out if the comparison would be true or false for the i'th element.
317 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
318 CompareRHS, TD, TLI);
319 // If the result is undef for this element, ignore it.
320 if (isa<UndefValue>(C)) {
321 // Extend range state machines to cover this element in case there is an
322 // undef in the middle of the range.
323 if (TrueRangeEnd == (int)i-1)
325 if (FalseRangeEnd == (int)i-1)
330 // If we can't compute the result for any of the elements, we have to give
331 // up evaluating the entire conditional.
332 if (!isa<ConstantInt>(C)) return 0;
334 // Otherwise, we know if the comparison is true or false for this element,
335 // update our state machines.
336 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
338 // State machine for single/double/range index comparison.
340 // Update the TrueElement state machine.
341 if (FirstTrueElement == Undefined)
342 FirstTrueElement = TrueRangeEnd = i; // First true element.
344 // Update double-compare state machine.
345 if (SecondTrueElement == Undefined)
346 SecondTrueElement = i;
348 SecondTrueElement = Overdefined;
350 // Update range state machine.
351 if (TrueRangeEnd == (int)i-1)
354 TrueRangeEnd = Overdefined;
357 // Update the FalseElement state machine.
358 if (FirstFalseElement == Undefined)
359 FirstFalseElement = FalseRangeEnd = i; // First false element.
361 // Update double-compare state machine.
362 if (SecondFalseElement == Undefined)
363 SecondFalseElement = i;
365 SecondFalseElement = Overdefined;
367 // Update range state machine.
368 if (FalseRangeEnd == (int)i-1)
371 FalseRangeEnd = Overdefined;
376 // If this element is in range, update our magic bitvector.
377 if (i < 64 && IsTrueForElt)
378 MagicBitvector |= 1ULL << i;
380 // If all of our states become overdefined, bail out early. Since the
381 // predicate is expensive, only check it every 8 elements. This is only
382 // really useful for really huge arrays.
383 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
384 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
385 FalseRangeEnd == Overdefined)
389 // Now that we've scanned the entire array, emit our new comparison(s). We
390 // order the state machines in complexity of the generated code.
391 Value *Idx = GEP->getOperand(2);
393 // If the index is larger than the pointer size of the target, truncate the
394 // index down like the GEP would do implicitly. We don't have to do this for
395 // an inbounds GEP because the index can't be out of range.
396 if (!GEP->isInBounds() &&
397 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
398 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
400 // If the comparison is only true for one or two elements, emit direct
402 if (SecondTrueElement != Overdefined) {
403 // None true -> false.
404 if (FirstTrueElement == Undefined)
405 return ReplaceInstUsesWith(ICI, Builder->getFalse());
407 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
409 // True for one element -> 'i == 47'.
410 if (SecondTrueElement == Undefined)
411 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
413 // True for two elements -> 'i == 47 | i == 72'.
414 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
415 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
416 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
417 return BinaryOperator::CreateOr(C1, C2);
420 // If the comparison is only false for one or two elements, emit direct
422 if (SecondFalseElement != Overdefined) {
423 // None false -> true.
424 if (FirstFalseElement == Undefined)
425 return ReplaceInstUsesWith(ICI, Builder->getTrue());
427 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
429 // False for one element -> 'i != 47'.
430 if (SecondFalseElement == Undefined)
431 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
433 // False for two elements -> 'i != 47 & i != 72'.
434 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
435 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
436 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
437 return BinaryOperator::CreateAnd(C1, C2);
440 // If the comparison can be replaced with a range comparison for the elements
441 // where it is true, emit the range check.
442 if (TrueRangeEnd != Overdefined) {
443 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
445 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
446 if (FirstTrueElement) {
447 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
448 Idx = Builder->CreateAdd(Idx, Offs);
451 Value *End = ConstantInt::get(Idx->getType(),
452 TrueRangeEnd-FirstTrueElement+1);
453 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
456 // False range check.
457 if (FalseRangeEnd != Overdefined) {
458 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
459 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
460 if (FirstFalseElement) {
461 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
462 Idx = Builder->CreateAdd(Idx, Offs);
465 Value *End = ConstantInt::get(Idx->getType(),
466 FalseRangeEnd-FirstFalseElement);
467 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
471 // If a magic bitvector captures the entire comparison state
472 // of this load, replace it with computation that does:
473 // ((magic_cst >> i) & 1) != 0
477 // Look for an appropriate type:
478 // - The type of Idx if the magic fits
479 // - The smallest fitting legal type if we have a DataLayout
481 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
484 Ty = TD->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
485 else if (ArrayElementCount <= 32)
486 Ty = Type::getInt32Ty(Init->getContext());
489 Value *V = Builder->CreateIntCast(Idx, Ty, false);
490 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
491 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
492 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
500 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
501 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
502 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
503 /// be complex, and scales are involved. The above expression would also be
504 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
505 /// This later form is less amenable to optimization though, and we are allowed
506 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
508 /// If we can't emit an optimized form for this expression, this returns null.
510 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
511 DataLayout &TD = *IC.getDataLayout();
512 gep_type_iterator GTI = gep_type_begin(GEP);
514 // Check to see if this gep only has a single variable index. If so, and if
515 // any constant indices are a multiple of its scale, then we can compute this
516 // in terms of the scale of the variable index. For example, if the GEP
517 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
518 // because the expression will cross zero at the same point.
519 unsigned i, e = GEP->getNumOperands();
521 for (i = 1; i != e; ++i, ++GTI) {
522 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
523 // Compute the aggregate offset of constant indices.
524 if (CI->isZero()) continue;
526 // Handle a struct index, which adds its field offset to the pointer.
527 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
528 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
530 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
531 Offset += Size*CI->getSExtValue();
534 // Found our variable index.
539 // If there are no variable indices, we must have a constant offset, just
540 // evaluate it the general way.
541 if (i == e) return 0;
543 Value *VariableIdx = GEP->getOperand(i);
544 // Determine the scale factor of the variable element. For example, this is
545 // 4 if the variable index is into an array of i32.
546 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
548 // Verify that there are no other variable indices. If so, emit the hard way.
549 for (++i, ++GTI; i != e; ++i, ++GTI) {
550 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
553 // Compute the aggregate offset of constant indices.
554 if (CI->isZero()) continue;
556 // Handle a struct index, which adds its field offset to the pointer.
557 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
558 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
560 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
561 Offset += Size*CI->getSExtValue();
565 // Okay, we know we have a single variable index, which must be a
566 // pointer/array/vector index. If there is no offset, life is simple, return
568 unsigned IntPtrWidth = TD.getPointerSizeInBits();
570 // Cast to intptrty in case a truncation occurs. If an extension is needed,
571 // we don't need to bother extending: the extension won't affect where the
572 // computation crosses zero.
573 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
574 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
575 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
580 // Otherwise, there is an index. The computation we will do will be modulo
581 // the pointer size, so get it.
582 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
584 Offset &= PtrSizeMask;
585 VariableScale &= PtrSizeMask;
587 // To do this transformation, any constant index must be a multiple of the
588 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
589 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
590 // multiple of the variable scale.
591 int64_t NewOffs = Offset / (int64_t)VariableScale;
592 if (Offset != NewOffs*(int64_t)VariableScale)
595 // Okay, we can do this evaluation. Start by converting the index to intptr.
596 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
597 if (VariableIdx->getType() != IntPtrTy)
598 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
600 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
601 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
604 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
605 /// else. At this point we know that the GEP is on the LHS of the comparison.
606 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
607 ICmpInst::Predicate Cond,
609 // Don't transform signed compares of GEPs into index compares. Even if the
610 // GEP is inbounds, the final add of the base pointer can have signed overflow
611 // and would change the result of the icmp.
612 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
613 // the maximum signed value for the pointer type.
614 if (ICmpInst::isSigned(Cond))
617 // Look through bitcasts.
618 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
619 RHS = BCI->getOperand(0);
621 Value *PtrBase = GEPLHS->getOperand(0);
622 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
623 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
624 // This transformation (ignoring the base and scales) is valid because we
625 // know pointers can't overflow since the gep is inbounds. See if we can
626 // output an optimized form.
627 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
629 // If not, synthesize the offset the hard way.
631 Offset = EmitGEPOffset(GEPLHS);
632 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
633 Constant::getNullValue(Offset->getType()));
634 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
635 // If the base pointers are different, but the indices are the same, just
636 // compare the base pointer.
637 if (PtrBase != GEPRHS->getOperand(0)) {
638 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
639 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
640 GEPRHS->getOperand(0)->getType();
642 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
643 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
644 IndicesTheSame = false;
648 // If all indices are the same, just compare the base pointers.
650 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
651 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
653 // If we're comparing GEPs with two base pointers that only differ in type
654 // and both GEPs have only constant indices or just one use, then fold
655 // the compare with the adjusted indices.
656 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
657 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
658 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
659 PtrBase->stripPointerCasts() ==
660 GEPRHS->getOperand(0)->stripPointerCasts()) {
661 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
662 EmitGEPOffset(GEPLHS),
663 EmitGEPOffset(GEPRHS));
664 return ReplaceInstUsesWith(I, Cmp);
667 // Otherwise, the base pointers are different and the indices are
668 // different, bail out.
672 // If one of the GEPs has all zero indices, recurse.
673 bool AllZeros = true;
674 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
675 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
676 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
681 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
682 ICmpInst::getSwappedPredicate(Cond), I);
684 // If the other GEP has all zero indices, recurse.
686 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
687 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
688 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
693 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
695 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
696 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
697 // If the GEPs only differ by one index, compare it.
698 unsigned NumDifferences = 0; // Keep track of # differences.
699 unsigned DiffOperand = 0; // The operand that differs.
700 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
701 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
702 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
703 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
704 // Irreconcilable differences.
708 if (NumDifferences++) break;
713 if (NumDifferences == 0) // SAME GEP? No comparison is needed here.
714 return ReplaceInstUsesWith(I, Builder->getInt1(Cond));
716 else if (NumDifferences == 1 && GEPsInBounds) {
717 Value *LHSV = GEPLHS->getOperand(DiffOperand);
718 Value *RHSV = GEPRHS->getOperand(DiffOperand);
719 // Make sure we do a signed comparison here.
720 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
724 // Only lower this if the icmp is the only user of the GEP or if we expect
725 // the result to fold to a constant!
728 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
729 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
730 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
731 Value *L = EmitGEPOffset(GEPLHS);
732 Value *R = EmitGEPOffset(GEPRHS);
733 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
739 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
740 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
741 Value *X, ConstantInt *CI,
742 ICmpInst::Predicate Pred,
744 // If we have X+0, exit early (simplifying logic below) and let it get folded
745 // elsewhere. icmp X+0, X -> icmp X, X
747 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
748 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
751 // (X+4) == X -> false.
752 if (Pred == ICmpInst::ICMP_EQ)
753 return ReplaceInstUsesWith(ICI, Builder->getFalse());
755 // (X+4) != X -> true.
756 if (Pred == ICmpInst::ICMP_NE)
757 return ReplaceInstUsesWith(ICI, Builder->getTrue());
759 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
760 // so the values can never be equal. Similarly for all other "or equals"
763 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
764 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
765 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
766 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
768 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
769 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
772 // (X+1) >u X --> X <u (0-1) --> X != 255
773 // (X+2) >u X --> X <u (0-2) --> X <u 254
774 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
775 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
776 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
778 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
779 ConstantInt *SMax = ConstantInt::get(X->getContext(),
780 APInt::getSignedMaxValue(BitWidth));
782 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
783 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
784 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
785 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
786 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
787 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
788 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
789 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
791 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
792 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
793 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
794 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
795 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
796 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
798 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
799 Constant *C = Builder->getInt(CI->getValue()-1);
800 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
803 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
804 /// and CmpRHS are both known to be integer constants.
805 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
806 ConstantInt *DivRHS) {
807 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
808 const APInt &CmpRHSV = CmpRHS->getValue();
810 // FIXME: If the operand types don't match the type of the divide
811 // then don't attempt this transform. The code below doesn't have the
812 // logic to deal with a signed divide and an unsigned compare (and
813 // vice versa). This is because (x /s C1) <s C2 produces different
814 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
815 // (x /u C1) <u C2. Simply casting the operands and result won't
816 // work. :( The if statement below tests that condition and bails
818 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
819 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
821 if (DivRHS->isZero())
822 return 0; // The ProdOV computation fails on divide by zero.
823 if (DivIsSigned && DivRHS->isAllOnesValue())
824 return 0; // The overflow computation also screws up here
825 if (DivRHS->isOne()) {
826 // This eliminates some funny cases with INT_MIN.
827 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
831 // Compute Prod = CI * DivRHS. We are essentially solving an equation
832 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
833 // C2 (CI). By solving for X we can turn this into a range check
834 // instead of computing a divide.
835 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
837 // Determine if the product overflows by seeing if the product is
838 // not equal to the divide. Make sure we do the same kind of divide
839 // as in the LHS instruction that we're folding.
840 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
841 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
843 // Get the ICmp opcode
844 ICmpInst::Predicate Pred = ICI.getPredicate();
846 /// If the division is known to be exact, then there is no remainder from the
847 /// divide, so the covered range size is unit, otherwise it is the divisor.
848 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
850 // Figure out the interval that is being checked. For example, a comparison
851 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
852 // Compute this interval based on the constants involved and the signedness of
853 // the compare/divide. This computes a half-open interval, keeping track of
854 // whether either value in the interval overflows. After analysis each
855 // overflow variable is set to 0 if it's corresponding bound variable is valid
856 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
857 int LoOverflow = 0, HiOverflow = 0;
858 Constant *LoBound = 0, *HiBound = 0;
860 if (!DivIsSigned) { // udiv
861 // e.g. X/5 op 3 --> [15, 20)
863 HiOverflow = LoOverflow = ProdOV;
865 // If this is not an exact divide, then many values in the range collapse
866 // to the same result value.
867 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
870 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
871 if (CmpRHSV == 0) { // (X / pos) op 0
872 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
873 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
875 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
876 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
877 HiOverflow = LoOverflow = ProdOV;
879 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
880 } else { // (X / pos) op neg
881 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
882 HiBound = AddOne(Prod);
883 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
885 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
886 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
889 } else if (DivRHS->isNegative()) { // Divisor is < 0.
891 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
892 if (CmpRHSV == 0) { // (X / neg) op 0
893 // e.g. X/-5 op 0 --> [-4, 5)
894 LoBound = AddOne(RangeSize);
895 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
896 if (HiBound == DivRHS) { // -INTMIN = INTMIN
897 HiOverflow = 1; // [INTMIN+1, overflow)
898 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
900 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
901 // e.g. X/-5 op 3 --> [-19, -14)
902 HiBound = AddOne(Prod);
903 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
905 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
906 } else { // (X / neg) op neg
907 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
908 LoOverflow = HiOverflow = ProdOV;
910 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
913 // Dividing by a negative swaps the condition. LT <-> GT
914 Pred = ICmpInst::getSwappedPredicate(Pred);
917 Value *X = DivI->getOperand(0);
919 default: llvm_unreachable("Unhandled icmp opcode!");
920 case ICmpInst::ICMP_EQ:
921 if (LoOverflow && HiOverflow)
922 return ReplaceInstUsesWith(ICI, Builder->getFalse());
924 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
925 ICmpInst::ICMP_UGE, X, LoBound);
927 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
928 ICmpInst::ICMP_ULT, X, HiBound);
929 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
931 case ICmpInst::ICMP_NE:
932 if (LoOverflow && HiOverflow)
933 return ReplaceInstUsesWith(ICI, Builder->getTrue());
935 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
936 ICmpInst::ICMP_ULT, X, LoBound);
938 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
939 ICmpInst::ICMP_UGE, X, HiBound);
940 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
941 DivIsSigned, false));
942 case ICmpInst::ICMP_ULT:
943 case ICmpInst::ICMP_SLT:
944 if (LoOverflow == +1) // Low bound is greater than input range.
945 return ReplaceInstUsesWith(ICI, Builder->getTrue());
946 if (LoOverflow == -1) // Low bound is less than input range.
947 return ReplaceInstUsesWith(ICI, Builder->getFalse());
948 return new ICmpInst(Pred, X, LoBound);
949 case ICmpInst::ICMP_UGT:
950 case ICmpInst::ICMP_SGT:
951 if (HiOverflow == +1) // High bound greater than input range.
952 return ReplaceInstUsesWith(ICI, Builder->getFalse());
953 if (HiOverflow == -1) // High bound less than input range.
954 return ReplaceInstUsesWith(ICI, Builder->getTrue());
955 if (Pred == ICmpInst::ICMP_UGT)
956 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
957 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
961 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
962 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
963 ConstantInt *ShAmt) {
964 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
966 // Check that the shift amount is in range. If not, don't perform
967 // undefined shifts. When the shift is visited it will be
969 uint32_t TypeBits = CmpRHSV.getBitWidth();
970 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
971 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
974 if (!ICI.isEquality()) {
975 // If we have an unsigned comparison and an ashr, we can't simplify this.
976 // Similarly for signed comparisons with lshr.
977 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
980 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
981 // by a power of 2. Since we already have logic to simplify these,
982 // transform to div and then simplify the resultant comparison.
983 if (Shr->getOpcode() == Instruction::AShr &&
984 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
987 // Revisit the shift (to delete it).
991 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
994 Shr->getOpcode() == Instruction::AShr ?
995 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
996 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
998 ICI.setOperand(0, Tmp);
1000 // If the builder folded the binop, just return it.
1001 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
1005 // Otherwise, fold this div/compare.
1006 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1007 TheDiv->getOpcode() == Instruction::UDiv);
1009 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1010 assert(Res && "This div/cst should have folded!");
1015 // If we are comparing against bits always shifted out, the
1016 // comparison cannot succeed.
1017 APInt Comp = CmpRHSV << ShAmtVal;
1018 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1019 if (Shr->getOpcode() == Instruction::LShr)
1020 Comp = Comp.lshr(ShAmtVal);
1022 Comp = Comp.ashr(ShAmtVal);
1024 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1025 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1026 Constant *Cst = Builder->getInt1(IsICMP_NE);
1027 return ReplaceInstUsesWith(ICI, Cst);
1030 // Otherwise, check to see if the bits shifted out are known to be zero.
1031 // If so, we can compare against the unshifted value:
1032 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1033 if (Shr->hasOneUse() && Shr->isExact())
1034 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1036 if (Shr->hasOneUse()) {
1037 // Otherwise strength reduce the shift into an and.
1038 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1039 Constant *Mask = Builder->getInt(Val);
1041 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1042 Mask, Shr->getName()+".mask");
1043 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1049 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1051 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1054 const APInt &RHSV = RHS->getValue();
1056 switch (LHSI->getOpcode()) {
1057 case Instruction::Trunc:
1058 if (ICI.isEquality() && LHSI->hasOneUse()) {
1059 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1060 // of the high bits truncated out of x are known.
1061 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1062 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1063 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1064 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1066 // If all the high bits are known, we can do this xform.
1067 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1068 // Pull in the high bits from known-ones set.
1069 APInt NewRHS = RHS->getValue().zext(SrcBits);
1070 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1071 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1072 Builder->getInt(NewRHS));
1077 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1078 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1079 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1081 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1082 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1083 Value *CompareVal = LHSI->getOperand(0);
1085 // If the sign bit of the XorCST is not set, there is no change to
1086 // the operation, just stop using the Xor.
1087 if (!XorCST->isNegative()) {
1088 ICI.setOperand(0, CompareVal);
1093 // Was the old condition true if the operand is positive?
1094 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1096 // If so, the new one isn't.
1097 isTrueIfPositive ^= true;
1099 if (isTrueIfPositive)
1100 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1103 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1107 if (LHSI->hasOneUse()) {
1108 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1109 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1110 const APInt &SignBit = XorCST->getValue();
1111 ICmpInst::Predicate Pred = ICI.isSigned()
1112 ? ICI.getUnsignedPredicate()
1113 : ICI.getSignedPredicate();
1114 return new ICmpInst(Pred, LHSI->getOperand(0),
1115 Builder->getInt(RHSV ^ SignBit));
1118 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1119 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1120 const APInt &NotSignBit = XorCST->getValue();
1121 ICmpInst::Predicate Pred = ICI.isSigned()
1122 ? ICI.getUnsignedPredicate()
1123 : ICI.getSignedPredicate();
1124 Pred = ICI.getSwappedPredicate(Pred);
1125 return new ICmpInst(Pred, LHSI->getOperand(0),
1126 Builder->getInt(RHSV ^ NotSignBit));
1131 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1132 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1133 LHSI->getOperand(0)->hasOneUse()) {
1134 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1136 // If the LHS is an AND of a truncating cast, we can widen the
1137 // and/compare to be the input width without changing the value
1138 // produced, eliminating a cast.
1139 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1140 // We can do this transformation if either the AND constant does not
1141 // have its sign bit set or if it is an equality comparison.
1142 // Extending a relational comparison when we're checking the sign
1143 // bit would not work.
1144 if (ICI.isEquality() ||
1145 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1147 Builder->CreateAnd(Cast->getOperand(0),
1148 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1149 NewAnd->takeName(LHSI);
1150 return new ICmpInst(ICI.getPredicate(), NewAnd,
1151 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1155 // If the LHS is an AND of a zext, and we have an equality compare, we can
1156 // shrink the and/compare to the smaller type, eliminating the cast.
1157 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1158 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1159 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1160 // should fold the icmp to true/false in that case.
1161 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1163 Builder->CreateAnd(Cast->getOperand(0),
1164 ConstantExpr::getTrunc(AndCST, Ty));
1165 NewAnd->takeName(LHSI);
1166 return new ICmpInst(ICI.getPredicate(), NewAnd,
1167 ConstantExpr::getTrunc(RHS, Ty));
1171 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1172 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1173 // happens a LOT in code produced by the C front-end, for bitfield
1175 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1176 if (Shift && !Shift->isShift())
1180 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1181 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1182 Type *AndTy = AndCST->getType(); // Type of the and.
1184 // We can fold this as long as we can't shift unknown bits
1185 // into the mask. This can only happen with signed shift
1186 // rights, as they sign-extend.
1188 bool CanFold = Shift->isLogicalShift();
1190 // To test for the bad case of the signed shr, see if any
1191 // of the bits shifted in could be tested after the mask.
1192 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1193 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1195 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1196 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1197 AndCST->getValue()) == 0)
1203 if (Shift->getOpcode() == Instruction::Shl)
1204 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1206 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1208 // Check to see if we are shifting out any of the bits being
1210 if (ConstantExpr::get(Shift->getOpcode(),
1211 NewCst, ShAmt) != RHS) {
1212 // If we shifted bits out, the fold is not going to work out.
1213 // As a special case, check to see if this means that the
1214 // result is always true or false now.
1215 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1216 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1217 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1218 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1220 ICI.setOperand(1, NewCst);
1221 Constant *NewAndCST;
1222 if (Shift->getOpcode() == Instruction::Shl)
1223 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1225 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1226 LHSI->setOperand(1, NewAndCST);
1227 LHSI->setOperand(0, Shift->getOperand(0));
1228 Worklist.Add(Shift); // Shift is dead.
1234 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1235 // preferable because it allows the C<<Y expression to be hoisted out
1236 // of a loop if Y is invariant and X is not.
1237 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1238 ICI.isEquality() && !Shift->isArithmeticShift() &&
1239 !isa<Constant>(Shift->getOperand(0))) {
1242 if (Shift->getOpcode() == Instruction::LShr) {
1243 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1245 // Insert a logical shift.
1246 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1249 // Compute X & (C << Y).
1251 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1253 ICI.setOperand(0, NewAnd);
1257 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
1258 // bit set in (X & AndCST) will produce a result greater than RHSV.
1259 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1260 unsigned NTZ = AndCST->getValue().countTrailingZeros();
1261 if ((NTZ < AndCST->getBitWidth()) &&
1262 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
1263 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1264 Constant::getNullValue(RHS->getType()));
1268 // Try to optimize things like "A[i]&42 == 0" to index computations.
1269 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1270 if (GetElementPtrInst *GEP =
1271 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1272 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1273 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1274 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1275 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1276 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1282 case Instruction::Or: {
1283 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1286 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1287 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1288 // -> and (icmp eq P, null), (icmp eq Q, null).
1289 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1290 Constant::getNullValue(P->getType()));
1291 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1292 Constant::getNullValue(Q->getType()));
1294 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1295 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1297 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1303 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1304 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1307 // If this is a signed comparison to 0 and the mul is sign preserving,
1308 // use the mul LHS operand instead.
1309 ICmpInst::Predicate pred = ICI.getPredicate();
1310 if (isSignTest(pred, RHS) && !Val->isZero() &&
1311 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1312 return new ICmpInst(Val->isNegative() ?
1313 ICmpInst::getSwappedPredicate(pred) : pred,
1314 LHSI->getOperand(0),
1315 Constant::getNullValue(RHS->getType()));
1320 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1321 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1324 uint32_t TypeBits = RHSV.getBitWidth();
1326 // Check that the shift amount is in range. If not, don't perform
1327 // undefined shifts. When the shift is visited it will be
1329 if (ShAmt->uge(TypeBits))
1332 if (ICI.isEquality()) {
1333 // If we are comparing against bits always shifted out, the
1334 // comparison cannot succeed.
1336 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1338 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1339 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1340 Constant *Cst = Builder->getInt1(IsICMP_NE);
1341 return ReplaceInstUsesWith(ICI, Cst);
1344 // If the shift is NUW, then it is just shifting out zeros, no need for an
1346 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1347 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1348 ConstantExpr::getLShr(RHS, ShAmt));
1350 // If the shift is NSW and we compare to 0, then it is just shifting out
1351 // sign bits, no need for an AND either.
1352 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1353 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1354 ConstantExpr::getLShr(RHS, ShAmt));
1356 if (LHSI->hasOneUse()) {
1357 // Otherwise strength reduce the shift into an and.
1358 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1359 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1360 TypeBits - ShAmtVal));
1363 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1364 return new ICmpInst(ICI.getPredicate(), And,
1365 ConstantExpr::getLShr(RHS, ShAmt));
1369 // If this is a signed comparison to 0 and the shift is sign preserving,
1370 // use the shift LHS operand instead.
1371 ICmpInst::Predicate pred = ICI.getPredicate();
1372 if (isSignTest(pred, RHS) &&
1373 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1374 return new ICmpInst(pred,
1375 LHSI->getOperand(0),
1376 Constant::getNullValue(RHS->getType()));
1378 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1379 bool TrueIfSigned = false;
1380 if (LHSI->hasOneUse() &&
1381 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1382 // (X << 31) <s 0 --> (X&1) != 0
1383 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1384 APInt::getOneBitSet(TypeBits,
1385 TypeBits-ShAmt->getZExtValue()-1));
1387 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1388 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1389 And, Constant::getNullValue(And->getType()));
1392 // Transform (icmp pred iM (shl iM %v, N), CI)
1393 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1394 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1395 // This enables to get rid of the shift in favor of a trunc which can be
1396 // free on the target. It has the additional benefit of comparing to a
1397 // smaller constant, which will be target friendly.
1398 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1399 if (LHSI->hasOneUse() &&
1400 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1401 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1402 Constant *NCI = ConstantExpr::getTrunc(
1403 ConstantExpr::getAShr(RHS,
1404 ConstantInt::get(RHS->getType(), Amt)),
1406 return new ICmpInst(ICI.getPredicate(),
1407 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1414 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1415 case Instruction::AShr: {
1416 // Handle equality comparisons of shift-by-constant.
1417 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1418 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1419 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1423 // Handle exact shr's.
1424 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1425 if (RHSV.isMinValue())
1426 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1431 case Instruction::SDiv:
1432 case Instruction::UDiv:
1433 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1434 // Fold this div into the comparison, producing a range check.
1435 // Determine, based on the divide type, what the range is being
1436 // checked. If there is an overflow on the low or high side, remember
1437 // it, otherwise compute the range [low, hi) bounding the new value.
1438 // See: InsertRangeTest above for the kinds of replacements possible.
1439 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1440 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1445 case Instruction::Add:
1446 // Fold: icmp pred (add X, C1), C2
1447 if (!ICI.isEquality()) {
1448 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1450 const APInt &LHSV = LHSC->getValue();
1452 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1455 if (ICI.isSigned()) {
1456 if (CR.getLower().isSignBit()) {
1457 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1458 Builder->getInt(CR.getUpper()));
1459 } else if (CR.getUpper().isSignBit()) {
1460 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1461 Builder->getInt(CR.getLower()));
1464 if (CR.getLower().isMinValue()) {
1465 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1466 Builder->getInt(CR.getUpper()));
1467 } else if (CR.getUpper().isMinValue()) {
1468 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1469 Builder->getInt(CR.getLower()));
1476 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1477 if (ICI.isEquality()) {
1478 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1480 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1481 // the second operand is a constant, simplify a bit.
1482 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1483 switch (BO->getOpcode()) {
1484 case Instruction::SRem:
1485 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1486 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1487 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1488 if (V.sgt(1) && V.isPowerOf2()) {
1490 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1492 return new ICmpInst(ICI.getPredicate(), NewRem,
1493 Constant::getNullValue(BO->getType()));
1497 case Instruction::Add:
1498 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1499 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1500 if (BO->hasOneUse())
1501 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1502 ConstantExpr::getSub(RHS, BOp1C));
1503 } else if (RHSV == 0) {
1504 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1505 // efficiently invertible, or if the add has just this one use.
1506 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1508 if (Value *NegVal = dyn_castNegVal(BOp1))
1509 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1510 if (Value *NegVal = dyn_castNegVal(BOp0))
1511 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1512 if (BO->hasOneUse()) {
1513 Value *Neg = Builder->CreateNeg(BOp1);
1515 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1519 case Instruction::Xor:
1520 // For the xor case, we can xor two constants together, eliminating
1521 // the explicit xor.
1522 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1523 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1524 ConstantExpr::getXor(RHS, BOC));
1525 } else if (RHSV == 0) {
1526 // Replace ((xor A, B) != 0) with (A != B)
1527 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1531 case Instruction::Sub:
1532 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1533 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1534 if (BO->hasOneUse())
1535 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1536 ConstantExpr::getSub(BOp0C, RHS));
1537 } else if (RHSV == 0) {
1538 // Replace ((sub A, B) != 0) with (A != B)
1539 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1543 case Instruction::Or:
1544 // If bits are being or'd in that are not present in the constant we
1545 // are comparing against, then the comparison could never succeed!
1546 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1547 Constant *NotCI = ConstantExpr::getNot(RHS);
1548 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1549 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1553 case Instruction::And:
1554 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1555 // If bits are being compared against that are and'd out, then the
1556 // comparison can never succeed!
1557 if ((RHSV & ~BOC->getValue()) != 0)
1558 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1560 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1561 if (RHS == BOC && RHSV.isPowerOf2())
1562 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1563 ICmpInst::ICMP_NE, LHSI,
1564 Constant::getNullValue(RHS->getType()));
1566 // Don't perform the following transforms if the AND has multiple uses
1567 if (!BO->hasOneUse())
1570 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1571 if (BOC->getValue().isSignBit()) {
1572 Value *X = BO->getOperand(0);
1573 Constant *Zero = Constant::getNullValue(X->getType());
1574 ICmpInst::Predicate pred = isICMP_NE ?
1575 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1576 return new ICmpInst(pred, X, Zero);
1579 // ((X & ~7) == 0) --> X < 8
1580 if (RHSV == 0 && isHighOnes(BOC)) {
1581 Value *X = BO->getOperand(0);
1582 Constant *NegX = ConstantExpr::getNeg(BOC);
1583 ICmpInst::Predicate pred = isICMP_NE ?
1584 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1585 return new ICmpInst(pred, X, NegX);
1589 case Instruction::Mul:
1590 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1591 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1592 // The trivial case (mul X, 0) is handled by InstSimplify
1593 // General case : (mul X, C) != 0 iff X != 0
1594 // (mul X, C) == 0 iff X == 0
1596 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1597 Constant::getNullValue(RHS->getType()));
1603 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1604 // Handle icmp {eq|ne} <intrinsic>, intcst.
1605 switch (II->getIntrinsicID()) {
1606 case Intrinsic::bswap:
1608 ICI.setOperand(0, II->getArgOperand(0));
1609 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1611 case Intrinsic::ctlz:
1612 case Intrinsic::cttz:
1613 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1614 if (RHSV == RHS->getType()->getBitWidth()) {
1616 ICI.setOperand(0, II->getArgOperand(0));
1617 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1621 case Intrinsic::ctpop:
1622 // popcount(A) == 0 -> A == 0 and likewise for !=
1623 if (RHS->isZero()) {
1625 ICI.setOperand(0, II->getArgOperand(0));
1626 ICI.setOperand(1, RHS);
1638 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1639 /// We only handle extending casts so far.
1641 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1642 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1643 Value *LHSCIOp = LHSCI->getOperand(0);
1644 Type *SrcTy = LHSCIOp->getType();
1645 Type *DestTy = LHSCI->getType();
1648 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1649 // integer type is the same size as the pointer type.
1650 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1651 TD->getPointerSizeInBits() ==
1652 cast<IntegerType>(DestTy)->getBitWidth()) {
1654 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1655 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1656 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1657 RHSOp = RHSC->getOperand(0);
1658 // If the pointer types don't match, insert a bitcast.
1659 if (LHSCIOp->getType() != RHSOp->getType())
1660 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1664 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1667 // The code below only handles extension cast instructions, so far.
1669 if (LHSCI->getOpcode() != Instruction::ZExt &&
1670 LHSCI->getOpcode() != Instruction::SExt)
1673 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1674 bool isSignedCmp = ICI.isSigned();
1676 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1677 // Not an extension from the same type?
1678 RHSCIOp = CI->getOperand(0);
1679 if (RHSCIOp->getType() != LHSCIOp->getType())
1682 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1683 // and the other is a zext), then we can't handle this.
1684 if (CI->getOpcode() != LHSCI->getOpcode())
1687 // Deal with equality cases early.
1688 if (ICI.isEquality())
1689 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1691 // A signed comparison of sign extended values simplifies into a
1692 // signed comparison.
1693 if (isSignedCmp && isSignedExt)
1694 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1696 // The other three cases all fold into an unsigned comparison.
1697 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1700 // If we aren't dealing with a constant on the RHS, exit early
1701 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1705 // Compute the constant that would happen if we truncated to SrcTy then
1706 // reextended to DestTy.
1707 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1708 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1711 // If the re-extended constant didn't change...
1713 // Deal with equality cases early.
1714 if (ICI.isEquality())
1715 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1717 // A signed comparison of sign extended values simplifies into a
1718 // signed comparison.
1719 if (isSignedExt && isSignedCmp)
1720 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1722 // The other three cases all fold into an unsigned comparison.
1723 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1726 // The re-extended constant changed so the constant cannot be represented
1727 // in the shorter type. Consequently, we cannot emit a simple comparison.
1728 // All the cases that fold to true or false will have already been handled
1729 // by SimplifyICmpInst, so only deal with the tricky case.
1731 if (isSignedCmp || !isSignedExt)
1734 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1735 // should have been folded away previously and not enter in here.
1737 // We're performing an unsigned comp with a sign extended value.
1738 // This is true if the input is >= 0. [aka >s -1]
1739 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1740 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1742 // Finally, return the value computed.
1743 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1744 return ReplaceInstUsesWith(ICI, Result);
1746 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1747 return BinaryOperator::CreateNot(Result);
1750 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1751 /// I = icmp ugt (add (add A, B), CI2), CI1
1752 /// If this is of the form:
1754 /// if (sum+128 >u 255)
1755 /// Then replace it with llvm.sadd.with.overflow.i8.
1757 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1758 ConstantInt *CI2, ConstantInt *CI1,
1760 // The transformation we're trying to do here is to transform this into an
1761 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1762 // with a narrower add, and discard the add-with-constant that is part of the
1763 // range check (if we can't eliminate it, this isn't profitable).
1765 // In order to eliminate the add-with-constant, the compare can be its only
1767 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1768 if (!AddWithCst->hasOneUse()) return 0;
1770 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1771 if (!CI2->getValue().isPowerOf2()) return 0;
1772 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1773 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1775 // The width of the new add formed is 1 more than the bias.
1778 // Check to see that CI1 is an all-ones value with NewWidth bits.
1779 if (CI1->getBitWidth() == NewWidth ||
1780 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1783 // This is only really a signed overflow check if the inputs have been
1784 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1785 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1786 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1787 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1788 IC.ComputeNumSignBits(B) < NeededSignBits)
1791 // In order to replace the original add with a narrower
1792 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1793 // and truncates that discard the high bits of the add. Verify that this is
1795 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1796 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1798 if (*UI == AddWithCst) continue;
1800 // Only accept truncates for now. We would really like a nice recursive
1801 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1802 // chain to see which bits of a value are actually demanded. If the
1803 // original add had another add which was then immediately truncated, we
1804 // could still do the transformation.
1805 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1807 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1810 // If the pattern matches, truncate the inputs to the narrower type and
1811 // use the sadd_with_overflow intrinsic to efficiently compute both the
1812 // result and the overflow bit.
1813 Module *M = I.getParent()->getParent()->getParent();
1815 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1816 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1819 InstCombiner::BuilderTy *Builder = IC.Builder;
1821 // Put the new code above the original add, in case there are any uses of the
1822 // add between the add and the compare.
1823 Builder->SetInsertPoint(OrigAdd);
1825 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1826 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1827 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1828 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1829 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1831 // The inner add was the result of the narrow add, zero extended to the
1832 // wider type. Replace it with the result computed by the intrinsic.
1833 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1835 // The original icmp gets replaced with the overflow value.
1836 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1839 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1841 // Don't bother doing this transformation for pointers, don't do it for
1843 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1845 // If the add is a constant expr, then we don't bother transforming it.
1846 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1847 if (OrigAdd == 0) return 0;
1849 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1851 // Put the new code above the original add, in case there are any uses of the
1852 // add between the add and the compare.
1853 InstCombiner::BuilderTy *Builder = IC.Builder;
1854 Builder->SetInsertPoint(OrigAdd);
1856 Module *M = I.getParent()->getParent()->getParent();
1857 Type *Ty = LHS->getType();
1858 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1859 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1860 Value *Add = Builder->CreateExtractValue(Call, 0);
1862 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1864 // The original icmp gets replaced with the overflow value.
1865 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1868 // DemandedBitsLHSMask - When performing a comparison against a constant,
1869 // it is possible that not all the bits in the LHS are demanded. This helper
1870 // method computes the mask that IS demanded.
1871 static APInt DemandedBitsLHSMask(ICmpInst &I,
1872 unsigned BitWidth, bool isSignCheck) {
1874 return APInt::getSignBit(BitWidth);
1876 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1877 if (!CI) return APInt::getAllOnesValue(BitWidth);
1878 const APInt &RHS = CI->getValue();
1880 switch (I.getPredicate()) {
1881 // For a UGT comparison, we don't care about any bits that
1882 // correspond to the trailing ones of the comparand. The value of these
1883 // bits doesn't impact the outcome of the comparison, because any value
1884 // greater than the RHS must differ in a bit higher than these due to carry.
1885 case ICmpInst::ICMP_UGT: {
1886 unsigned trailingOnes = RHS.countTrailingOnes();
1887 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1891 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1892 // Any value less than the RHS must differ in a higher bit because of carries.
1893 case ICmpInst::ICMP_ULT: {
1894 unsigned trailingZeros = RHS.countTrailingZeros();
1895 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1900 return APInt::getAllOnesValue(BitWidth);
1905 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1906 bool Changed = false;
1907 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1909 /// Orders the operands of the compare so that they are listed from most
1910 /// complex to least complex. This puts constants before unary operators,
1911 /// before binary operators.
1912 if (getComplexity(Op0) < getComplexity(Op1)) {
1914 std::swap(Op0, Op1);
1918 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1919 return ReplaceInstUsesWith(I, V);
1921 // comparing -val or val with non-zero is the same as just comparing val
1922 // ie, abs(val) != 0 -> val != 0
1923 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
1925 Value *Cond, *SelectTrue, *SelectFalse;
1926 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
1927 m_Value(SelectFalse)))) {
1928 if (Value *V = dyn_castNegVal(SelectTrue)) {
1929 if (V == SelectFalse)
1930 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1932 else if (Value *V = dyn_castNegVal(SelectFalse)) {
1933 if (V == SelectTrue)
1934 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1939 Type *Ty = Op0->getType();
1941 // icmp's with boolean values can always be turned into bitwise operations
1942 if (Ty->isIntegerTy(1)) {
1943 switch (I.getPredicate()) {
1944 default: llvm_unreachable("Invalid icmp instruction!");
1945 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1946 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1947 return BinaryOperator::CreateNot(Xor);
1949 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1950 return BinaryOperator::CreateXor(Op0, Op1);
1952 case ICmpInst::ICMP_UGT:
1953 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1955 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1956 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1957 return BinaryOperator::CreateAnd(Not, Op1);
1959 case ICmpInst::ICMP_SGT:
1960 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1962 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1963 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1964 return BinaryOperator::CreateAnd(Not, Op0);
1966 case ICmpInst::ICMP_UGE:
1967 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1969 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1970 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1971 return BinaryOperator::CreateOr(Not, Op1);
1973 case ICmpInst::ICMP_SGE:
1974 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1976 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1977 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1978 return BinaryOperator::CreateOr(Not, Op0);
1983 unsigned BitWidth = 0;
1984 if (Ty->isIntOrIntVectorTy())
1985 BitWidth = Ty->getScalarSizeInBits();
1986 else if (TD) // Pointers require TD info to get their size.
1987 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1989 bool isSignBit = false;
1991 // See if we are doing a comparison with a constant.
1992 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1993 Value *A = 0, *B = 0;
1995 // Match the following pattern, which is a common idiom when writing
1996 // overflow-safe integer arithmetic function. The source performs an
1997 // addition in wider type, and explicitly checks for overflow using
1998 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1999 // sadd_with_overflow intrinsic.
2001 // TODO: This could probably be generalized to handle other overflow-safe
2002 // operations if we worked out the formulas to compute the appropriate
2006 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2008 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2009 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2010 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2011 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2015 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2016 if (I.isEquality() && CI->isZero() &&
2017 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2018 // (icmp cond A B) if cond is equality
2019 return new ICmpInst(I.getPredicate(), A, B);
2022 // If we have an icmp le or icmp ge instruction, turn it into the
2023 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2024 // them being folded in the code below. The SimplifyICmpInst code has
2025 // already handled the edge cases for us, so we just assert on them.
2026 switch (I.getPredicate()) {
2028 case ICmpInst::ICMP_ULE:
2029 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2030 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2031 Builder->getInt(CI->getValue()+1));
2032 case ICmpInst::ICMP_SLE:
2033 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2034 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2035 Builder->getInt(CI->getValue()+1));
2036 case ICmpInst::ICMP_UGE:
2037 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2038 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2039 Builder->getInt(CI->getValue()-1));
2040 case ICmpInst::ICMP_SGE:
2041 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2042 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2043 Builder->getInt(CI->getValue()-1));
2046 // If this comparison is a normal comparison, it demands all
2047 // bits, if it is a sign bit comparison, it only demands the sign bit.
2049 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2052 // See if we can fold the comparison based on range information we can get
2053 // by checking whether bits are known to be zero or one in the input.
2054 if (BitWidth != 0) {
2055 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2056 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2058 if (SimplifyDemandedBits(I.getOperandUse(0),
2059 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2060 Op0KnownZero, Op0KnownOne, 0))
2062 if (SimplifyDemandedBits(I.getOperandUse(1),
2063 APInt::getAllOnesValue(BitWidth),
2064 Op1KnownZero, Op1KnownOne, 0))
2067 // Given the known and unknown bits, compute a range that the LHS could be
2068 // in. Compute the Min, Max and RHS values based on the known bits. For the
2069 // EQ and NE we use unsigned values.
2070 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2071 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2073 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2075 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2078 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2080 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2084 // If Min and Max are known to be the same, then SimplifyDemandedBits
2085 // figured out that the LHS is a constant. Just constant fold this now so
2086 // that code below can assume that Min != Max.
2087 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2088 return new ICmpInst(I.getPredicate(),
2089 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2090 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2091 return new ICmpInst(I.getPredicate(), Op0,
2092 ConstantInt::get(Op1->getType(), Op1Min));
2094 // Based on the range information we know about the LHS, see if we can
2095 // simplify this comparison. For example, (x&4) < 8 is always true.
2096 switch (I.getPredicate()) {
2097 default: llvm_unreachable("Unknown icmp opcode!");
2098 case ICmpInst::ICMP_EQ: {
2099 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2100 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2102 // If all bits are known zero except for one, then we know at most one
2103 // bit is set. If the comparison is against zero, then this is a check
2104 // to see if *that* bit is set.
2105 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2106 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2107 // If the LHS is an AND with the same constant, look through it.
2109 ConstantInt *LHSC = 0;
2110 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2111 LHSC->getValue() != Op0KnownZeroInverted)
2114 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2115 // then turn "((1 << x)&8) == 0" into "x != 3".
2117 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2118 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2119 return new ICmpInst(ICmpInst::ICMP_NE, X,
2120 ConstantInt::get(X->getType(), CmpVal));
2123 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2124 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2126 if (Op0KnownZeroInverted == 1 &&
2127 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2128 return new ICmpInst(ICmpInst::ICMP_NE, X,
2129 ConstantInt::get(X->getType(),
2130 CI->countTrailingZeros()));
2135 case ICmpInst::ICMP_NE: {
2136 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2137 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2139 // If all bits are known zero except for one, then we know at most one
2140 // bit is set. If the comparison is against zero, then this is a check
2141 // to see if *that* bit is set.
2142 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2143 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2144 // If the LHS is an AND with the same constant, look through it.
2146 ConstantInt *LHSC = 0;
2147 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2148 LHSC->getValue() != Op0KnownZeroInverted)
2151 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2152 // then turn "((1 << x)&8) != 0" into "x == 3".
2154 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2155 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2156 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2157 ConstantInt::get(X->getType(), CmpVal));
2160 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2161 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2163 if (Op0KnownZeroInverted == 1 &&
2164 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2165 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2166 ConstantInt::get(X->getType(),
2167 CI->countTrailingZeros()));
2172 case ICmpInst::ICMP_ULT:
2173 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2174 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2175 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2176 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2177 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2178 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2179 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2180 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2181 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2182 Builder->getInt(CI->getValue()-1));
2184 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2185 if (CI->isMinValue(true))
2186 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2187 Constant::getAllOnesValue(Op0->getType()));
2190 case ICmpInst::ICMP_UGT:
2191 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2192 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2193 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2194 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2196 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2197 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2198 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2199 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2200 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2201 Builder->getInt(CI->getValue()+1));
2203 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2204 if (CI->isMaxValue(true))
2205 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2206 Constant::getNullValue(Op0->getType()));
2209 case ICmpInst::ICMP_SLT:
2210 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2211 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2212 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2213 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2214 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2215 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2216 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2217 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2218 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2219 Builder->getInt(CI->getValue()-1));
2222 case ICmpInst::ICMP_SGT:
2223 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2224 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2225 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2226 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2228 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2229 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2230 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2231 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2232 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2233 Builder->getInt(CI->getValue()+1));
2236 case ICmpInst::ICMP_SGE:
2237 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2238 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2239 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2240 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2241 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2243 case ICmpInst::ICMP_SLE:
2244 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2245 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2246 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2247 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2248 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2250 case ICmpInst::ICMP_UGE:
2251 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2252 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2253 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2254 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2255 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2257 case ICmpInst::ICMP_ULE:
2258 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2259 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2260 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2261 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2262 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2266 // Turn a signed comparison into an unsigned one if both operands
2267 // are known to have the same sign.
2269 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2270 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2271 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2274 // Test if the ICmpInst instruction is used exclusively by a select as
2275 // part of a minimum or maximum operation. If so, refrain from doing
2276 // any other folding. This helps out other analyses which understand
2277 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2278 // and CodeGen. And in this case, at least one of the comparison
2279 // operands has at least one user besides the compare (the select),
2280 // which would often largely negate the benefit of folding anyway.
2282 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2283 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2284 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2287 // See if we are doing a comparison between a constant and an instruction that
2288 // can be folded into the comparison.
2289 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2290 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2291 // instruction, see if that instruction also has constants so that the
2292 // instruction can be folded into the icmp
2293 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2294 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2298 // Handle icmp with constant (but not simple integer constant) RHS
2299 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2300 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2301 switch (LHSI->getOpcode()) {
2302 case Instruction::GetElementPtr:
2303 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2304 if (RHSC->isNullValue() &&
2305 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2306 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2307 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2309 case Instruction::PHI:
2310 // Only fold icmp into the PHI if the phi and icmp are in the same
2311 // block. If in the same block, we're encouraging jump threading. If
2312 // not, we are just pessimizing the code by making an i1 phi.
2313 if (LHSI->getParent() == I.getParent())
2314 if (Instruction *NV = FoldOpIntoPhi(I))
2317 case Instruction::Select: {
2318 // If either operand of the select is a constant, we can fold the
2319 // comparison into the select arms, which will cause one to be
2320 // constant folded and the select turned into a bitwise or.
2321 Value *Op1 = 0, *Op2 = 0;
2322 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2323 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2324 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2325 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2327 // We only want to perform this transformation if it will not lead to
2328 // additional code. This is true if either both sides of the select
2329 // fold to a constant (in which case the icmp is replaced with a select
2330 // which will usually simplify) or this is the only user of the
2331 // select (in which case we are trading a select+icmp for a simpler
2333 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2335 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2338 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2340 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2344 case Instruction::IntToPtr:
2345 // icmp pred inttoptr(X), null -> icmp pred X, 0
2346 if (RHSC->isNullValue() && TD &&
2347 TD->getIntPtrType(RHSC->getContext()) ==
2348 LHSI->getOperand(0)->getType())
2349 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2350 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2353 case Instruction::Load:
2354 // Try to optimize things like "A[i] > 4" to index computations.
2355 if (GetElementPtrInst *GEP =
2356 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2357 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2358 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2359 !cast<LoadInst>(LHSI)->isVolatile())
2360 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2367 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2368 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2369 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2371 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2372 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2373 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2376 // Test to see if the operands of the icmp are casted versions of other
2377 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2379 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2380 if (Op0->getType()->isPointerTy() &&
2381 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2382 // We keep moving the cast from the left operand over to the right
2383 // operand, where it can often be eliminated completely.
2384 Op0 = CI->getOperand(0);
2386 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2387 // so eliminate it as well.
2388 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2389 Op1 = CI2->getOperand(0);
2391 // If Op1 is a constant, we can fold the cast into the constant.
2392 if (Op0->getType() != Op1->getType()) {
2393 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2394 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2396 // Otherwise, cast the RHS right before the icmp
2397 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2400 return new ICmpInst(I.getPredicate(), Op0, Op1);
2404 if (isa<CastInst>(Op0)) {
2405 // Handle the special case of: icmp (cast bool to X), <cst>
2406 // This comes up when you have code like
2409 // For generality, we handle any zero-extension of any operand comparison
2410 // with a constant or another cast from the same type.
2411 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2412 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2416 // Special logic for binary operators.
2417 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2418 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2420 CmpInst::Predicate Pred = I.getPredicate();
2421 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2422 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2423 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2424 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2425 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2426 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2427 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2428 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2429 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2431 // Analyze the case when either Op0 or Op1 is an add instruction.
2432 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2433 Value *A = 0, *B = 0, *C = 0, *D = 0;
2434 if (BO0 && BO0->getOpcode() == Instruction::Add)
2435 A = BO0->getOperand(0), B = BO0->getOperand(1);
2436 if (BO1 && BO1->getOpcode() == Instruction::Add)
2437 C = BO1->getOperand(0), D = BO1->getOperand(1);
2439 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2440 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2441 return new ICmpInst(Pred, A == Op1 ? B : A,
2442 Constant::getNullValue(Op1->getType()));
2444 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2445 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2446 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2449 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2450 if (A && C && (A == C || A == D || B == C || B == D) &&
2451 NoOp0WrapProblem && NoOp1WrapProblem &&
2452 // Try not to increase register pressure.
2453 BO0->hasOneUse() && BO1->hasOneUse()) {
2454 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2457 // C + B == C + D -> B == D
2460 } else if (A == D) {
2461 // D + B == C + D -> B == C
2464 } else if (B == C) {
2465 // A + C == C + D -> A == D
2470 // A + D == C + D -> A == C
2474 return new ICmpInst(Pred, Y, Z);
2477 // icmp slt (X + -1), Y -> icmp sle X, Y
2478 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2479 match(B, m_AllOnes()))
2480 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2482 // icmp sge (X + -1), Y -> icmp sgt X, Y
2483 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2484 match(B, m_AllOnes()))
2485 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2487 // icmp sle (X + 1), Y -> icmp slt X, Y
2488 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
2490 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2492 // icmp sgt (X + 1), Y -> icmp sge X, Y
2493 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
2495 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2497 // if C1 has greater magnitude than C2:
2498 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2499 // s.t. C3 = C1 - C2
2501 // if C2 has greater magnitude than C1:
2502 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2503 // s.t. C3 = C2 - C1
2504 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2505 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2506 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2507 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2508 const APInt &AP1 = C1->getValue();
2509 const APInt &AP2 = C2->getValue();
2510 if (AP1.isNegative() == AP2.isNegative()) {
2511 APInt AP1Abs = C1->getValue().abs();
2512 APInt AP2Abs = C2->getValue().abs();
2513 if (AP1Abs.uge(AP2Abs)) {
2514 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2515 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2516 return new ICmpInst(Pred, NewAdd, C);
2518 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2519 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2520 return new ICmpInst(Pred, A, NewAdd);
2526 // Analyze the case when either Op0 or Op1 is a sub instruction.
2527 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2528 A = 0; B = 0; C = 0; D = 0;
2529 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2530 A = BO0->getOperand(0), B = BO0->getOperand(1);
2531 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2532 C = BO1->getOperand(0), D = BO1->getOperand(1);
2534 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2535 if (A == Op1 && NoOp0WrapProblem)
2536 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2538 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2539 if (C == Op0 && NoOp1WrapProblem)
2540 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2542 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2543 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2544 // Try not to increase register pressure.
2545 BO0->hasOneUse() && BO1->hasOneUse())
2546 return new ICmpInst(Pred, A, C);
2548 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2549 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2550 // Try not to increase register pressure.
2551 BO0->hasOneUse() && BO1->hasOneUse())
2552 return new ICmpInst(Pred, D, B);
2554 BinaryOperator *SRem = NULL;
2555 // icmp (srem X, Y), Y
2556 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2557 Op1 == BO0->getOperand(1))
2559 // icmp Y, (srem X, Y)
2560 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2561 Op0 == BO1->getOperand(1))
2564 // We don't check hasOneUse to avoid increasing register pressure because
2565 // the value we use is the same value this instruction was already using.
2566 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2568 case ICmpInst::ICMP_EQ:
2569 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2570 case ICmpInst::ICMP_NE:
2571 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2572 case ICmpInst::ICMP_SGT:
2573 case ICmpInst::ICMP_SGE:
2574 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2575 Constant::getAllOnesValue(SRem->getType()));
2576 case ICmpInst::ICMP_SLT:
2577 case ICmpInst::ICMP_SLE:
2578 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2579 Constant::getNullValue(SRem->getType()));
2583 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2584 BO0->hasOneUse() && BO1->hasOneUse() &&
2585 BO0->getOperand(1) == BO1->getOperand(1)) {
2586 switch (BO0->getOpcode()) {
2588 case Instruction::Add:
2589 case Instruction::Sub:
2590 case Instruction::Xor:
2591 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2592 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2593 BO1->getOperand(0));
2594 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2595 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2596 if (CI->getValue().isSignBit()) {
2597 ICmpInst::Predicate Pred = I.isSigned()
2598 ? I.getUnsignedPredicate()
2599 : I.getSignedPredicate();
2600 return new ICmpInst(Pred, BO0->getOperand(0),
2601 BO1->getOperand(0));
2604 if (CI->isMaxValue(true)) {
2605 ICmpInst::Predicate Pred = I.isSigned()
2606 ? I.getUnsignedPredicate()
2607 : I.getSignedPredicate();
2608 Pred = I.getSwappedPredicate(Pred);
2609 return new ICmpInst(Pred, BO0->getOperand(0),
2610 BO1->getOperand(0));
2614 case Instruction::Mul:
2615 if (!I.isEquality())
2618 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2619 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2620 // Mask = -1 >> count-trailing-zeros(Cst).
2621 if (!CI->isZero() && !CI->isOne()) {
2622 const APInt &AP = CI->getValue();
2623 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2624 APInt::getLowBitsSet(AP.getBitWidth(),
2626 AP.countTrailingZeros()));
2627 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2628 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2629 return new ICmpInst(I.getPredicate(), And1, And2);
2633 case Instruction::UDiv:
2634 case Instruction::LShr:
2638 case Instruction::SDiv:
2639 case Instruction::AShr:
2640 if (!BO0->isExact() || !BO1->isExact())
2642 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2643 BO1->getOperand(0));
2644 case Instruction::Shl: {
2645 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2646 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2649 if (!NSW && I.isSigned())
2651 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2652 BO1->getOperand(0));
2659 // Transform (A & ~B) == 0 --> (A & B) != 0
2660 // and (A & ~B) != 0 --> (A & B) == 0
2661 // if A is a power of 2.
2662 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2663 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
2664 return new ICmpInst(I.getInversePredicate(),
2665 Builder->CreateAnd(A, B),
2668 // ~x < ~y --> y < x
2669 // ~x < cst --> ~cst < x
2670 if (match(Op0, m_Not(m_Value(A)))) {
2671 if (match(Op1, m_Not(m_Value(B))))
2672 return new ICmpInst(I.getPredicate(), B, A);
2673 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2674 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2677 // (a+b) <u a --> llvm.uadd.with.overflow.
2678 // (a+b) <u b --> llvm.uadd.with.overflow.
2679 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2680 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2681 (Op1 == A || Op1 == B))
2682 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2685 // a >u (a+b) --> llvm.uadd.with.overflow.
2686 // b >u (a+b) --> llvm.uadd.with.overflow.
2687 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2688 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2689 (Op0 == A || Op0 == B))
2690 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2694 if (I.isEquality()) {
2695 Value *A, *B, *C, *D;
2697 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2698 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2699 Value *OtherVal = A == Op1 ? B : A;
2700 return new ICmpInst(I.getPredicate(), OtherVal,
2701 Constant::getNullValue(A->getType()));
2704 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2705 // A^c1 == C^c2 --> A == C^(c1^c2)
2706 ConstantInt *C1, *C2;
2707 if (match(B, m_ConstantInt(C1)) &&
2708 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2709 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
2710 Value *Xor = Builder->CreateXor(C, NC);
2711 return new ICmpInst(I.getPredicate(), A, Xor);
2714 // A^B == A^D -> B == D
2715 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2716 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2717 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2718 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2722 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2723 (A == Op0 || B == Op0)) {
2724 // A == (A^B) -> B == 0
2725 Value *OtherVal = A == Op0 ? B : A;
2726 return new ICmpInst(I.getPredicate(), OtherVal,
2727 Constant::getNullValue(A->getType()));
2730 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2731 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2732 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2733 Value *X = 0, *Y = 0, *Z = 0;
2736 X = B; Y = D; Z = A;
2737 } else if (A == D) {
2738 X = B; Y = C; Z = A;
2739 } else if (B == C) {
2740 X = A; Y = D; Z = B;
2741 } else if (B == D) {
2742 X = A; Y = C; Z = B;
2745 if (X) { // Build (X^Y) & Z
2746 Op1 = Builder->CreateXor(X, Y);
2747 Op1 = Builder->CreateAnd(Op1, Z);
2748 I.setOperand(0, Op1);
2749 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2754 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2755 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2757 if ((Op0->hasOneUse() &&
2758 match(Op0, m_ZExt(m_Value(A))) &&
2759 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2760 (Op1->hasOneUse() &&
2761 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2762 match(Op1, m_ZExt(m_Value(A))))) {
2763 APInt Pow2 = Cst1->getValue() + 1;
2764 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2765 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2766 return new ICmpInst(I.getPredicate(), A,
2767 Builder->CreateTrunc(B, A->getType()));
2770 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2771 // "icmp (and X, mask), cst"
2773 if (Op0->hasOneUse() &&
2774 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2775 m_ConstantInt(ShAmt))))) &&
2776 match(Op1, m_ConstantInt(Cst1)) &&
2777 // Only do this when A has multiple uses. This is most important to do
2778 // when it exposes other optimizations.
2780 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2782 if (ShAmt < ASize) {
2784 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2787 APInt CmpV = Cst1->getValue().zext(ASize);
2790 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2791 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2797 Value *X; ConstantInt *Cst;
2799 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2800 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2803 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2804 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2806 return Changed ? &I : 0;
2814 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2816 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2819 if (!isa<ConstantFP>(RHSC)) return 0;
2820 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2822 // Get the width of the mantissa. We don't want to hack on conversions that
2823 // might lose information from the integer, e.g. "i64 -> float"
2824 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2825 if (MantissaWidth == -1) return 0; // Unknown.
2827 // Check to see that the input is converted from an integer type that is small
2828 // enough that preserves all bits. TODO: check here for "known" sign bits.
2829 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2830 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2832 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2833 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2837 // If the conversion would lose info, don't hack on this.
2838 if ((int)InputSize > MantissaWidth)
2841 // Otherwise, we can potentially simplify the comparison. We know that it
2842 // will always come through as an integer value and we know the constant is
2843 // not a NAN (it would have been previously simplified).
2844 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2846 ICmpInst::Predicate Pred;
2847 switch (I.getPredicate()) {
2848 default: llvm_unreachable("Unexpected predicate!");
2849 case FCmpInst::FCMP_UEQ:
2850 case FCmpInst::FCMP_OEQ:
2851 Pred = ICmpInst::ICMP_EQ;
2853 case FCmpInst::FCMP_UGT:
2854 case FCmpInst::FCMP_OGT:
2855 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2857 case FCmpInst::FCMP_UGE:
2858 case FCmpInst::FCMP_OGE:
2859 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2861 case FCmpInst::FCMP_ULT:
2862 case FCmpInst::FCMP_OLT:
2863 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2865 case FCmpInst::FCMP_ULE:
2866 case FCmpInst::FCMP_OLE:
2867 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2869 case FCmpInst::FCMP_UNE:
2870 case FCmpInst::FCMP_ONE:
2871 Pred = ICmpInst::ICMP_NE;
2873 case FCmpInst::FCMP_ORD:
2874 return ReplaceInstUsesWith(I, Builder->getTrue());
2875 case FCmpInst::FCMP_UNO:
2876 return ReplaceInstUsesWith(I, Builder->getTrue());
2879 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2881 // Now we know that the APFloat is a normal number, zero or inf.
2883 // See if the FP constant is too large for the integer. For example,
2884 // comparing an i8 to 300.0.
2885 unsigned IntWidth = IntTy->getScalarSizeInBits();
2888 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2889 // and large values.
2890 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2891 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2892 APFloat::rmNearestTiesToEven);
2893 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2894 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2895 Pred == ICmpInst::ICMP_SLE)
2896 return ReplaceInstUsesWith(I, Builder->getTrue());
2897 return ReplaceInstUsesWith(I, Builder->getFalse());
2900 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2901 // +INF and large values.
2902 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2903 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2904 APFloat::rmNearestTiesToEven);
2905 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2906 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2907 Pred == ICmpInst::ICMP_ULE)
2908 return ReplaceInstUsesWith(I, Builder->getTrue());
2909 return ReplaceInstUsesWith(I, Builder->getFalse());
2914 // See if the RHS value is < SignedMin.
2915 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2916 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2917 APFloat::rmNearestTiesToEven);
2918 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2919 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2920 Pred == ICmpInst::ICMP_SGE)
2921 return ReplaceInstUsesWith(I, Builder->getTrue());
2922 return ReplaceInstUsesWith(I, Builder->getFalse());
2925 // See if the RHS value is < UnsignedMin.
2926 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2927 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
2928 APFloat::rmNearestTiesToEven);
2929 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
2930 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
2931 Pred == ICmpInst::ICMP_UGE)
2932 return ReplaceInstUsesWith(I, Builder->getTrue());
2933 return ReplaceInstUsesWith(I, Builder->getFalse());
2937 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2938 // [0, UMAX], but it may still be fractional. See if it is fractional by
2939 // casting the FP value to the integer value and back, checking for equality.
2940 // Don't do this for zero, because -0.0 is not fractional.
2941 Constant *RHSInt = LHSUnsigned
2942 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2943 : ConstantExpr::getFPToSI(RHSC, IntTy);
2944 if (!RHS.isZero()) {
2945 bool Equal = LHSUnsigned
2946 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2947 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2949 // If we had a comparison against a fractional value, we have to adjust
2950 // the compare predicate and sometimes the value. RHSC is rounded towards
2951 // zero at this point.
2953 default: llvm_unreachable("Unexpected integer comparison!");
2954 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2955 return ReplaceInstUsesWith(I, Builder->getTrue());
2956 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2957 return ReplaceInstUsesWith(I, Builder->getFalse());
2958 case ICmpInst::ICMP_ULE:
2959 // (float)int <= 4.4 --> int <= 4
2960 // (float)int <= -4.4 --> false
2961 if (RHS.isNegative())
2962 return ReplaceInstUsesWith(I, Builder->getFalse());
2964 case ICmpInst::ICMP_SLE:
2965 // (float)int <= 4.4 --> int <= 4
2966 // (float)int <= -4.4 --> int < -4
2967 if (RHS.isNegative())
2968 Pred = ICmpInst::ICMP_SLT;
2970 case ICmpInst::ICMP_ULT:
2971 // (float)int < -4.4 --> false
2972 // (float)int < 4.4 --> int <= 4
2973 if (RHS.isNegative())
2974 return ReplaceInstUsesWith(I, Builder->getFalse());
2975 Pred = ICmpInst::ICMP_ULE;
2977 case ICmpInst::ICMP_SLT:
2978 // (float)int < -4.4 --> int < -4
2979 // (float)int < 4.4 --> int <= 4
2980 if (!RHS.isNegative())
2981 Pred = ICmpInst::ICMP_SLE;
2983 case ICmpInst::ICMP_UGT:
2984 // (float)int > 4.4 --> int > 4
2985 // (float)int > -4.4 --> true
2986 if (RHS.isNegative())
2987 return ReplaceInstUsesWith(I, Builder->getTrue());
2989 case ICmpInst::ICMP_SGT:
2990 // (float)int > 4.4 --> int > 4
2991 // (float)int > -4.4 --> int >= -4
2992 if (RHS.isNegative())
2993 Pred = ICmpInst::ICMP_SGE;
2995 case ICmpInst::ICMP_UGE:
2996 // (float)int >= -4.4 --> true
2997 // (float)int >= 4.4 --> int > 4
2998 if (RHS.isNegative())
2999 return ReplaceInstUsesWith(I, Builder->getTrue());
3000 Pred = ICmpInst::ICMP_UGT;
3002 case ICmpInst::ICMP_SGE:
3003 // (float)int >= -4.4 --> int >= -4
3004 // (float)int >= 4.4 --> int > 4
3005 if (!RHS.isNegative())
3006 Pred = ICmpInst::ICMP_SGT;
3012 // Lower this FP comparison into an appropriate integer version of the
3014 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3017 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3018 bool Changed = false;
3020 /// Orders the operands of the compare so that they are listed from most
3021 /// complex to least complex. This puts constants before unary operators,
3022 /// before binary operators.
3023 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3028 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3030 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
3031 return ReplaceInstUsesWith(I, V);
3033 // Simplify 'fcmp pred X, X'
3035 switch (I.getPredicate()) {
3036 default: llvm_unreachable("Unknown predicate!");
3037 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3038 case FCmpInst::FCMP_ULT: // True if unordered or less than
3039 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3040 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3041 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3042 I.setPredicate(FCmpInst::FCMP_UNO);
3043 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3046 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3047 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3048 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3049 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3050 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3051 I.setPredicate(FCmpInst::FCMP_ORD);
3052 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3057 // Handle fcmp with constant RHS
3058 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3059 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3060 switch (LHSI->getOpcode()) {
3061 case Instruction::FPExt: {
3062 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3063 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3064 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3068 const fltSemantics *Sem;
3069 // FIXME: This shouldn't be here.
3070 if (LHSExt->getSrcTy()->isHalfTy())
3071 Sem = &APFloat::IEEEhalf;
3072 else if (LHSExt->getSrcTy()->isFloatTy())
3073 Sem = &APFloat::IEEEsingle;
3074 else if (LHSExt->getSrcTy()->isDoubleTy())
3075 Sem = &APFloat::IEEEdouble;
3076 else if (LHSExt->getSrcTy()->isFP128Ty())
3077 Sem = &APFloat::IEEEquad;
3078 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3079 Sem = &APFloat::x87DoubleExtended;
3080 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3081 Sem = &APFloat::PPCDoubleDouble;
3086 APFloat F = RHSF->getValueAPF();
3087 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3089 // Avoid lossy conversions and denormals. Zero is a special case
3090 // that's OK to convert.
3094 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3095 APFloat::cmpLessThan) || Fabs.isZero()))
3097 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3098 ConstantFP::get(RHSC->getContext(), F));
3101 case Instruction::PHI:
3102 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3103 // block. If in the same block, we're encouraging jump threading. If
3104 // not, we are just pessimizing the code by making an i1 phi.
3105 if (LHSI->getParent() == I.getParent())
3106 if (Instruction *NV = FoldOpIntoPhi(I))
3109 case Instruction::SIToFP:
3110 case Instruction::UIToFP:
3111 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3114 case Instruction::Select: {
3115 // If either operand of the select is a constant, we can fold the
3116 // comparison into the select arms, which will cause one to be
3117 // constant folded and the select turned into a bitwise or.
3118 Value *Op1 = 0, *Op2 = 0;
3119 if (LHSI->hasOneUse()) {
3120 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3121 // Fold the known value into the constant operand.
3122 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3123 // Insert a new FCmp of the other select operand.
3124 Op2 = Builder->CreateFCmp(I.getPredicate(),
3125 LHSI->getOperand(2), RHSC, I.getName());
3126 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3127 // Fold the known value into the constant operand.
3128 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3129 // Insert a new FCmp of the other select operand.
3130 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
3136 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3139 case Instruction::FSub: {
3140 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3142 if (match(LHSI, m_FNeg(m_Value(Op))))
3143 return new FCmpInst(I.getSwappedPredicate(), Op,
3144 ConstantExpr::getFNeg(RHSC));
3147 case Instruction::Load:
3148 if (GetElementPtrInst *GEP =
3149 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3150 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3151 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3152 !cast<LoadInst>(LHSI)->isVolatile())
3153 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3157 case Instruction::Call: {
3158 CallInst *CI = cast<CallInst>(LHSI);
3160 // Various optimization for fabs compared with zero.
3161 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3162 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3164 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3165 Func == LibFunc::fabsl) {
3166 switch (I.getPredicate()) {
3168 // fabs(x) < 0 --> false
3169 case FCmpInst::FCMP_OLT:
3170 return ReplaceInstUsesWith(I, Builder->getFalse());
3171 // fabs(x) > 0 --> x != 0
3172 case FCmpInst::FCMP_OGT:
3173 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3175 // fabs(x) <= 0 --> x == 0
3176 case FCmpInst::FCMP_OLE:
3177 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3179 // fabs(x) >= 0 --> !isnan(x)
3180 case FCmpInst::FCMP_OGE:
3181 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3183 // fabs(x) == 0 --> x == 0
3184 // fabs(x) != 0 --> x != 0
3185 case FCmpInst::FCMP_OEQ:
3186 case FCmpInst::FCMP_UEQ:
3187 case FCmpInst::FCMP_ONE:
3188 case FCmpInst::FCMP_UNE:
3189 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3198 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3200 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3201 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3203 // fcmp (fpext x), (fpext y) -> fcmp x, y
3204 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3205 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3206 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3207 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3208 RHSExt->getOperand(0));
3210 return Changed ? &I : 0;