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 if (!GEP->isInBounds())
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 comparison is only true for one or two elements, emit direct
395 if (SecondTrueElement != Overdefined) {
396 // None true -> false.
397 if (FirstTrueElement == Undefined)
398 return ReplaceInstUsesWith(ICI, Builder->getFalse());
400 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
402 // True for one element -> 'i == 47'.
403 if (SecondTrueElement == Undefined)
404 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
406 // True for two elements -> 'i == 47 | i == 72'.
407 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
408 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
409 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
410 return BinaryOperator::CreateOr(C1, C2);
413 // If the comparison is only false for one or two elements, emit direct
415 if (SecondFalseElement != Overdefined) {
416 // None false -> true.
417 if (FirstFalseElement == Undefined)
418 return ReplaceInstUsesWith(ICI, Builder->getTrue());
420 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
422 // False for one element -> 'i != 47'.
423 if (SecondFalseElement == Undefined)
424 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
426 // False for two elements -> 'i != 47 & i != 72'.
427 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
428 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
429 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
430 return BinaryOperator::CreateAnd(C1, C2);
433 // If the comparison can be replaced with a range comparison for the elements
434 // where it is true, emit the range check.
435 if (TrueRangeEnd != Overdefined) {
436 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
438 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
439 if (FirstTrueElement) {
440 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
441 Idx = Builder->CreateAdd(Idx, Offs);
444 Value *End = ConstantInt::get(Idx->getType(),
445 TrueRangeEnd-FirstTrueElement+1);
446 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
449 // False range check.
450 if (FalseRangeEnd != Overdefined) {
451 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
452 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
453 if (FirstFalseElement) {
454 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
455 Idx = Builder->CreateAdd(Idx, Offs);
458 Value *End = ConstantInt::get(Idx->getType(),
459 FalseRangeEnd-FirstFalseElement);
460 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
464 // If a magic bitvector captures the entire comparison state
465 // of this load, replace it with computation that does:
466 // ((magic_cst >> i) & 1) != 0
470 // Look for an appropriate type:
471 // - The type of Idx if the magic fits
472 // - The smallest fitting legal type if we have a DataLayout
474 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
477 Ty = TD->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
478 else if (ArrayElementCount <= 32)
479 Ty = Type::getInt32Ty(Init->getContext());
482 Value *V = Builder->CreateIntCast(Idx, Ty, false);
483 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
484 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
485 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
493 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
494 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
495 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
496 /// be complex, and scales are involved. The above expression would also be
497 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
498 /// This later form is less amenable to optimization though, and we are allowed
499 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
501 /// If we can't emit an optimized form for this expression, this returns null.
503 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
504 DataLayout &TD = *IC.getDataLayout();
505 gep_type_iterator GTI = gep_type_begin(GEP);
507 // Check to see if this gep only has a single variable index. If so, and if
508 // any constant indices are a multiple of its scale, then we can compute this
509 // in terms of the scale of the variable index. For example, if the GEP
510 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
511 // because the expression will cross zero at the same point.
512 unsigned i, e = GEP->getNumOperands();
514 for (i = 1; i != e; ++i, ++GTI) {
515 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
516 // Compute the aggregate offset of constant indices.
517 if (CI->isZero()) continue;
519 // Handle a struct index, which adds its field offset to the pointer.
520 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
521 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
523 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
524 Offset += Size*CI->getSExtValue();
527 // Found our variable index.
532 // If there are no variable indices, we must have a constant offset, just
533 // evaluate it the general way.
534 if (i == e) return 0;
536 Value *VariableIdx = GEP->getOperand(i);
537 // Determine the scale factor of the variable element. For example, this is
538 // 4 if the variable index is into an array of i32.
539 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
541 // Verify that there are no other variable indices. If so, emit the hard way.
542 for (++i, ++GTI; i != e; ++i, ++GTI) {
543 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
546 // Compute the aggregate offset of constant indices.
547 if (CI->isZero()) continue;
549 // Handle a struct index, which adds its field offset to the pointer.
550 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
551 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
553 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
554 Offset += Size*CI->getSExtValue();
558 // Okay, we know we have a single variable index, which must be a
559 // pointer/array/vector index. If there is no offset, life is simple, return
561 unsigned IntPtrWidth = TD.getPointerSizeInBits();
563 // Cast to intptrty in case a truncation occurs. If an extension is needed,
564 // we don't need to bother extending: the extension won't affect where the
565 // computation crosses zero.
566 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
567 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
568 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
573 // Otherwise, there is an index. The computation we will do will be modulo
574 // the pointer size, so get it.
575 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
577 Offset &= PtrSizeMask;
578 VariableScale &= PtrSizeMask;
580 // To do this transformation, any constant index must be a multiple of the
581 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
582 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
583 // multiple of the variable scale.
584 int64_t NewOffs = Offset / (int64_t)VariableScale;
585 if (Offset != NewOffs*(int64_t)VariableScale)
588 // Okay, we can do this evaluation. Start by converting the index to intptr.
589 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
590 if (VariableIdx->getType() != IntPtrTy)
591 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
593 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
594 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
597 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
598 /// else. At this point we know that the GEP is on the LHS of the comparison.
599 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
600 ICmpInst::Predicate Cond,
602 // Don't transform signed compares of GEPs into index compares. Even if the
603 // GEP is inbounds, the final add of the base pointer can have signed overflow
604 // and would change the result of the icmp.
605 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
606 // the maximum signed value for the pointer type.
607 if (ICmpInst::isSigned(Cond))
610 // Look through bitcasts.
611 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
612 RHS = BCI->getOperand(0);
614 Value *PtrBase = GEPLHS->getOperand(0);
615 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
616 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
617 // This transformation (ignoring the base and scales) is valid because we
618 // know pointers can't overflow since the gep is inbounds. See if we can
619 // output an optimized form.
620 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
622 // If not, synthesize the offset the hard way.
624 Offset = EmitGEPOffset(GEPLHS);
625 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
626 Constant::getNullValue(Offset->getType()));
627 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
628 // If the base pointers are different, but the indices are the same, just
629 // compare the base pointer.
630 if (PtrBase != GEPRHS->getOperand(0)) {
631 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
632 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
633 GEPRHS->getOperand(0)->getType();
635 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
636 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
637 IndicesTheSame = false;
641 // If all indices are the same, just compare the base pointers.
643 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
645 // If we're comparing GEPs with two base pointers that only differ in type
646 // and both GEPs have only constant indices or just one use, then fold
647 // the compare with the adjusted indices.
648 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
649 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
650 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
651 PtrBase->stripPointerCasts() ==
652 GEPRHS->getOperand(0)->stripPointerCasts()) {
653 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
654 EmitGEPOffset(GEPLHS),
655 EmitGEPOffset(GEPRHS));
656 return ReplaceInstUsesWith(I, Cmp);
659 // Otherwise, the base pointers are different and the indices are
660 // different, bail out.
664 // If one of the GEPs has all zero indices, recurse.
665 bool AllZeros = true;
666 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
667 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
668 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
673 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
674 ICmpInst::getSwappedPredicate(Cond), I);
676 // If the other GEP has all zero indices, recurse.
678 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
679 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
680 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
685 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
687 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
688 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
689 // If the GEPs only differ by one index, compare it.
690 unsigned NumDifferences = 0; // Keep track of # differences.
691 unsigned DiffOperand = 0; // The operand that differs.
692 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
693 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
694 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
695 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
696 // Irreconcilable differences.
700 if (NumDifferences++) break;
705 if (NumDifferences == 0) // SAME GEP?
706 return ReplaceInstUsesWith(I, // No comparison is needed here.
707 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
709 else if (NumDifferences == 1 && GEPsInBounds) {
710 Value *LHSV = GEPLHS->getOperand(DiffOperand);
711 Value *RHSV = GEPRHS->getOperand(DiffOperand);
712 // Make sure we do a signed comparison here.
713 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
717 // Only lower this if the icmp is the only user of the GEP or if we expect
718 // the result to fold to a constant!
721 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
722 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
723 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
724 Value *L = EmitGEPOffset(GEPLHS);
725 Value *R = EmitGEPOffset(GEPRHS);
726 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
732 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
733 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
734 Value *X, ConstantInt *CI,
735 ICmpInst::Predicate Pred,
737 // If we have X+0, exit early (simplifying logic below) and let it get folded
738 // elsewhere. icmp X+0, X -> icmp X, X
740 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
741 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
744 // (X+4) == X -> false.
745 if (Pred == ICmpInst::ICMP_EQ)
746 return ReplaceInstUsesWith(ICI, Builder->getFalse());
748 // (X+4) != X -> true.
749 if (Pred == ICmpInst::ICMP_NE)
750 return ReplaceInstUsesWith(ICI, Builder->getTrue());
752 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
753 // so the values can never be equal. Similarly for all other "or equals"
756 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
757 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
758 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
759 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
761 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
762 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
765 // (X+1) >u X --> X <u (0-1) --> X != 255
766 // (X+2) >u X --> X <u (0-2) --> X <u 254
767 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
768 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
769 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
771 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
772 ConstantInt *SMax = ConstantInt::get(X->getContext(),
773 APInt::getSignedMaxValue(BitWidth));
775 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
776 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
777 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
778 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
779 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
780 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
781 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
782 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
784 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
785 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
786 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
787 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
788 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
789 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
791 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
792 Constant *C = Builder->getInt(CI->getValue()-1);
793 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
796 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
797 /// and CmpRHS are both known to be integer constants.
798 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
799 ConstantInt *DivRHS) {
800 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
801 const APInt &CmpRHSV = CmpRHS->getValue();
803 // FIXME: If the operand types don't match the type of the divide
804 // then don't attempt this transform. The code below doesn't have the
805 // logic to deal with a signed divide and an unsigned compare (and
806 // vice versa). This is because (x /s C1) <s C2 produces different
807 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
808 // (x /u C1) <u C2. Simply casting the operands and result won't
809 // work. :( The if statement below tests that condition and bails
811 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
812 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
814 if (DivRHS->isZero())
815 return 0; // The ProdOV computation fails on divide by zero.
816 if (DivIsSigned && DivRHS->isAllOnesValue())
817 return 0; // The overflow computation also screws up here
818 if (DivRHS->isOne()) {
819 // This eliminates some funny cases with INT_MIN.
820 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
824 // Compute Prod = CI * DivRHS. We are essentially solving an equation
825 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
826 // C2 (CI). By solving for X we can turn this into a range check
827 // instead of computing a divide.
828 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
830 // Determine if the product overflows by seeing if the product is
831 // not equal to the divide. Make sure we do the same kind of divide
832 // as in the LHS instruction that we're folding.
833 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
834 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
836 // Get the ICmp opcode
837 ICmpInst::Predicate Pred = ICI.getPredicate();
839 /// If the division is known to be exact, then there is no remainder from the
840 /// divide, so the covered range size is unit, otherwise it is the divisor.
841 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
843 // Figure out the interval that is being checked. For example, a comparison
844 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
845 // Compute this interval based on the constants involved and the signedness of
846 // the compare/divide. This computes a half-open interval, keeping track of
847 // whether either value in the interval overflows. After analysis each
848 // overflow variable is set to 0 if it's corresponding bound variable is valid
849 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
850 int LoOverflow = 0, HiOverflow = 0;
851 Constant *LoBound = 0, *HiBound = 0;
853 if (!DivIsSigned) { // udiv
854 // e.g. X/5 op 3 --> [15, 20)
856 HiOverflow = LoOverflow = ProdOV;
858 // If this is not an exact divide, then many values in the range collapse
859 // to the same result value.
860 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
863 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
864 if (CmpRHSV == 0) { // (X / pos) op 0
865 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
866 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
868 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
869 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
870 HiOverflow = LoOverflow = ProdOV;
872 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
873 } else { // (X / pos) op neg
874 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
875 HiBound = AddOne(Prod);
876 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
878 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
879 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
882 } else if (DivRHS->isNegative()) { // Divisor is < 0.
884 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
885 if (CmpRHSV == 0) { // (X / neg) op 0
886 // e.g. X/-5 op 0 --> [-4, 5)
887 LoBound = AddOne(RangeSize);
888 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
889 if (HiBound == DivRHS) { // -INTMIN = INTMIN
890 HiOverflow = 1; // [INTMIN+1, overflow)
891 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
893 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
894 // e.g. X/-5 op 3 --> [-19, -14)
895 HiBound = AddOne(Prod);
896 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
898 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
899 } else { // (X / neg) op neg
900 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
901 LoOverflow = HiOverflow = ProdOV;
903 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
906 // Dividing by a negative swaps the condition. LT <-> GT
907 Pred = ICmpInst::getSwappedPredicate(Pred);
910 Value *X = DivI->getOperand(0);
912 default: llvm_unreachable("Unhandled icmp opcode!");
913 case ICmpInst::ICMP_EQ:
914 if (LoOverflow && HiOverflow)
915 return ReplaceInstUsesWith(ICI, Builder->getFalse());
917 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
918 ICmpInst::ICMP_UGE, X, LoBound);
920 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
921 ICmpInst::ICMP_ULT, X, HiBound);
922 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
924 case ICmpInst::ICMP_NE:
925 if (LoOverflow && HiOverflow)
926 return ReplaceInstUsesWith(ICI, Builder->getTrue());
928 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
929 ICmpInst::ICMP_ULT, X, LoBound);
931 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
932 ICmpInst::ICMP_UGE, X, HiBound);
933 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
934 DivIsSigned, false));
935 case ICmpInst::ICMP_ULT:
936 case ICmpInst::ICMP_SLT:
937 if (LoOverflow == +1) // Low bound is greater than input range.
938 return ReplaceInstUsesWith(ICI, Builder->getTrue());
939 if (LoOverflow == -1) // Low bound is less than input range.
940 return ReplaceInstUsesWith(ICI, Builder->getFalse());
941 return new ICmpInst(Pred, X, LoBound);
942 case ICmpInst::ICMP_UGT:
943 case ICmpInst::ICMP_SGT:
944 if (HiOverflow == +1) // High bound greater than input range.
945 return ReplaceInstUsesWith(ICI, Builder->getFalse());
946 if (HiOverflow == -1) // High bound less than input range.
947 return ReplaceInstUsesWith(ICI, Builder->getTrue());
948 if (Pred == ICmpInst::ICMP_UGT)
949 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
950 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
954 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
955 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
956 ConstantInt *ShAmt) {
957 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
959 // Check that the shift amount is in range. If not, don't perform
960 // undefined shifts. When the shift is visited it will be
962 uint32_t TypeBits = CmpRHSV.getBitWidth();
963 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
964 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
967 if (!ICI.isEquality()) {
968 // If we have an unsigned comparison and an ashr, we can't simplify this.
969 // Similarly for signed comparisons with lshr.
970 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
973 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
974 // by a power of 2. Since we already have logic to simplify these,
975 // transform to div and then simplify the resultant comparison.
976 if (Shr->getOpcode() == Instruction::AShr &&
977 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
980 // Revisit the shift (to delete it).
984 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
987 Shr->getOpcode() == Instruction::AShr ?
988 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
989 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
991 ICI.setOperand(0, Tmp);
993 // If the builder folded the binop, just return it.
994 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
998 // Otherwise, fold this div/compare.
999 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1000 TheDiv->getOpcode() == Instruction::UDiv);
1002 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1003 assert(Res && "This div/cst should have folded!");
1008 // If we are comparing against bits always shifted out, the
1009 // comparison cannot succeed.
1010 APInt Comp = CmpRHSV << ShAmtVal;
1011 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1012 if (Shr->getOpcode() == Instruction::LShr)
1013 Comp = Comp.lshr(ShAmtVal);
1015 Comp = Comp.ashr(ShAmtVal);
1017 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1018 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1019 Constant *Cst = Builder->getInt1(IsICMP_NE);
1020 return ReplaceInstUsesWith(ICI, Cst);
1023 // Otherwise, check to see if the bits shifted out are known to be zero.
1024 // If so, we can compare against the unshifted value:
1025 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1026 if (Shr->hasOneUse() && Shr->isExact())
1027 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1029 if (Shr->hasOneUse()) {
1030 // Otherwise strength reduce the shift into an and.
1031 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1032 Constant *Mask = Builder->getInt(Val);
1034 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1035 Mask, Shr->getName()+".mask");
1036 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1042 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1044 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1047 const APInt &RHSV = RHS->getValue();
1049 switch (LHSI->getOpcode()) {
1050 case Instruction::Trunc:
1051 if (ICI.isEquality() && LHSI->hasOneUse()) {
1052 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1053 // of the high bits truncated out of x are known.
1054 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1055 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1056 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1057 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1059 // If all the high bits are known, we can do this xform.
1060 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1061 // Pull in the high bits from known-ones set.
1062 APInt NewRHS = RHS->getValue().zext(SrcBits);
1063 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1064 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1065 Builder->getInt(NewRHS));
1070 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1071 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1072 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1074 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1075 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1076 Value *CompareVal = LHSI->getOperand(0);
1078 // If the sign bit of the XorCST is not set, there is no change to
1079 // the operation, just stop using the Xor.
1080 if (!XorCST->isNegative()) {
1081 ICI.setOperand(0, CompareVal);
1086 // Was the old condition true if the operand is positive?
1087 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1089 // If so, the new one isn't.
1090 isTrueIfPositive ^= true;
1092 if (isTrueIfPositive)
1093 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1096 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1100 if (LHSI->hasOneUse()) {
1101 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1102 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1103 const APInt &SignBit = XorCST->getValue();
1104 ICmpInst::Predicate Pred = ICI.isSigned()
1105 ? ICI.getUnsignedPredicate()
1106 : ICI.getSignedPredicate();
1107 return new ICmpInst(Pred, LHSI->getOperand(0),
1108 Builder->getInt(RHSV ^ SignBit));
1111 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1112 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1113 const APInt &NotSignBit = XorCST->getValue();
1114 ICmpInst::Predicate Pred = ICI.isSigned()
1115 ? ICI.getUnsignedPredicate()
1116 : ICI.getSignedPredicate();
1117 Pred = ICI.getSwappedPredicate(Pred);
1118 return new ICmpInst(Pred, LHSI->getOperand(0),
1119 Builder->getInt(RHSV ^ NotSignBit));
1123 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1124 // iff -C is a power of 2
1125 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1126 XorCST->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1127 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCST);
1129 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1130 // iff -C is a power of 2
1131 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1132 XorCST->getValue() == -RHSV && RHSV.isPowerOf2())
1133 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCST);
1136 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1137 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1138 LHSI->getOperand(0)->hasOneUse()) {
1139 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1141 // If the LHS is an AND of a truncating cast, we can widen the
1142 // and/compare to be the input width without changing the value
1143 // produced, eliminating a cast.
1144 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1145 // We can do this transformation if either the AND constant does not
1146 // have its sign bit set or if it is an equality comparison.
1147 // Extending a relational comparison when we're checking the sign
1148 // bit would not work.
1149 if (ICI.isEquality() ||
1150 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1152 Builder->CreateAnd(Cast->getOperand(0),
1153 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1154 NewAnd->takeName(LHSI);
1155 return new ICmpInst(ICI.getPredicate(), NewAnd,
1156 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1160 // If the LHS is an AND of a zext, and we have an equality compare, we can
1161 // shrink the and/compare to the smaller type, eliminating the cast.
1162 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1163 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1164 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1165 // should fold the icmp to true/false in that case.
1166 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1168 Builder->CreateAnd(Cast->getOperand(0),
1169 ConstantExpr::getTrunc(AndCST, Ty));
1170 NewAnd->takeName(LHSI);
1171 return new ICmpInst(ICI.getPredicate(), NewAnd,
1172 ConstantExpr::getTrunc(RHS, Ty));
1176 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1177 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1178 // happens a LOT in code produced by the C front-end, for bitfield
1180 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1181 if (Shift && !Shift->isShift())
1185 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1186 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1187 Type *AndTy = AndCST->getType(); // Type of the and.
1189 // We can fold this as long as we can't shift unknown bits
1190 // into the mask. This can only happen with signed shift
1191 // rights, as they sign-extend.
1193 bool CanFold = Shift->isLogicalShift();
1195 // To test for the bad case of the signed shr, see if any
1196 // of the bits shifted in could be tested after the mask.
1197 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1198 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1200 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1201 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1202 AndCST->getValue()) == 0)
1208 if (Shift->getOpcode() == Instruction::Shl)
1209 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1211 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1213 // Check to see if we are shifting out any of the bits being
1215 if (ConstantExpr::get(Shift->getOpcode(),
1216 NewCst, ShAmt) != RHS) {
1217 // If we shifted bits out, the fold is not going to work out.
1218 // As a special case, check to see if this means that the
1219 // result is always true or false now.
1220 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1221 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1222 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1223 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1225 ICI.setOperand(1, NewCst);
1226 Constant *NewAndCST;
1227 if (Shift->getOpcode() == Instruction::Shl)
1228 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1230 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1231 LHSI->setOperand(1, NewAndCST);
1232 LHSI->setOperand(0, Shift->getOperand(0));
1233 Worklist.Add(Shift); // Shift is dead.
1239 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1240 // preferable because it allows the C<<Y expression to be hoisted out
1241 // of a loop if Y is invariant and X is not.
1242 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1243 ICI.isEquality() && !Shift->isArithmeticShift() &&
1244 !isa<Constant>(Shift->getOperand(0))) {
1247 if (Shift->getOpcode() == Instruction::LShr) {
1248 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1250 // Insert a logical shift.
1251 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1254 // Compute X & (C << Y).
1256 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1258 ICI.setOperand(0, NewAnd);
1262 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
1263 // bit set in (X & AndCST) will produce a result greater than RHSV.
1264 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1265 unsigned NTZ = AndCST->getValue().countTrailingZeros();
1266 if ((NTZ < AndCST->getBitWidth()) &&
1267 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
1268 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1269 Constant::getNullValue(RHS->getType()));
1273 // Try to optimize things like "A[i]&42 == 0" to index computations.
1274 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1275 if (GetElementPtrInst *GEP =
1276 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1277 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1278 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1279 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1280 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1281 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1286 // X & -C == -C -> X > u ~C
1287 // X & -C != -C -> X <= u ~C
1288 // iff C is a power of 2
1289 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1290 return new ICmpInst(
1291 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1292 : ICmpInst::ICMP_ULE,
1293 LHSI->getOperand(0), SubOne(RHS));
1296 case Instruction::Or: {
1297 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1300 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1301 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1302 // -> and (icmp eq P, null), (icmp eq Q, null).
1303 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1304 Constant::getNullValue(P->getType()));
1305 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1306 Constant::getNullValue(Q->getType()));
1308 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1309 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1311 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1317 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1318 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1321 // If this is a signed comparison to 0 and the mul is sign preserving,
1322 // use the mul LHS operand instead.
1323 ICmpInst::Predicate pred = ICI.getPredicate();
1324 if (isSignTest(pred, RHS) && !Val->isZero() &&
1325 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1326 return new ICmpInst(Val->isNegative() ?
1327 ICmpInst::getSwappedPredicate(pred) : pred,
1328 LHSI->getOperand(0),
1329 Constant::getNullValue(RHS->getType()));
1334 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1335 uint32_t TypeBits = RHSV.getBitWidth();
1336 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1339 // (1 << X) pred P2 -> X pred Log2(P2)
1340 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1341 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1342 ICmpInst::Predicate Pred = ICI.getPredicate();
1343 if (ICI.isUnsigned()) {
1344 if (!RHSVIsPowerOf2) {
1345 // (1 << X) < 30 -> X <= 4
1346 // (1 << X) <= 30 -> X <= 4
1347 // (1 << X) >= 30 -> X > 4
1348 // (1 << X) > 30 -> X > 4
1349 if (Pred == ICmpInst::ICMP_ULT)
1350 Pred = ICmpInst::ICMP_ULE;
1351 else if (Pred == ICmpInst::ICMP_UGE)
1352 Pred = ICmpInst::ICMP_UGT;
1354 unsigned RHSLog2 = RHSV.logBase2();
1356 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1357 // (1 << X) > 2147483648 -> X > 31 -> false
1358 // (1 << X) <= 2147483648 -> X <= 31 -> true
1359 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1360 if (RHSLog2 == TypeBits-1) {
1361 if (Pred == ICmpInst::ICMP_UGE)
1362 Pred = ICmpInst::ICMP_EQ;
1363 else if (Pred == ICmpInst::ICMP_UGT)
1364 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1365 else if (Pred == ICmpInst::ICMP_ULE)
1366 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1367 else if (Pred == ICmpInst::ICMP_ULT)
1368 Pred = ICmpInst::ICMP_NE;
1371 return new ICmpInst(Pred, X,
1372 ConstantInt::get(RHS->getType(), RHSLog2));
1373 } else if (ICI.isSigned()) {
1374 if (RHSV.isAllOnesValue()) {
1375 // (1 << X) <= -1 -> X == 31
1376 if (Pred == ICmpInst::ICMP_SLE)
1377 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1378 ConstantInt::get(RHS->getType(), TypeBits-1));
1380 // (1 << X) > -1 -> X != 31
1381 if (Pred == ICmpInst::ICMP_SGT)
1382 return new ICmpInst(ICmpInst::ICMP_NE, X,
1383 ConstantInt::get(RHS->getType(), TypeBits-1));
1385 // (1 << X) < 0 -> X == 31
1386 // (1 << X) <= 0 -> X == 31
1387 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1388 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1389 ConstantInt::get(RHS->getType(), TypeBits-1));
1391 // (1 << X) >= 0 -> X != 31
1392 // (1 << X) > 0 -> X != 31
1393 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1394 return new ICmpInst(ICmpInst::ICMP_NE, X,
1395 ConstantInt::get(RHS->getType(), TypeBits-1));
1397 } else if (ICI.isEquality()) {
1399 return new ICmpInst(
1400 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1402 return ReplaceInstUsesWith(
1403 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse()
1404 : Builder->getTrue());
1410 // Check that the shift amount is in range. If not, don't perform
1411 // undefined shifts. When the shift is visited it will be
1413 if (ShAmt->uge(TypeBits))
1416 if (ICI.isEquality()) {
1417 // If we are comparing against bits always shifted out, the
1418 // comparison cannot succeed.
1420 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1422 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1423 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1424 Constant *Cst = Builder->getInt1(IsICMP_NE);
1425 return ReplaceInstUsesWith(ICI, Cst);
1428 // If the shift is NUW, then it is just shifting out zeros, no need for an
1430 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1431 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1432 ConstantExpr::getLShr(RHS, ShAmt));
1434 // If the shift is NSW and we compare to 0, then it is just shifting out
1435 // sign bits, no need for an AND either.
1436 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1437 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1438 ConstantExpr::getLShr(RHS, ShAmt));
1440 if (LHSI->hasOneUse()) {
1441 // Otherwise strength reduce the shift into an and.
1442 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1443 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1444 TypeBits - ShAmtVal));
1447 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1448 return new ICmpInst(ICI.getPredicate(), And,
1449 ConstantExpr::getLShr(RHS, ShAmt));
1453 // If this is a signed comparison to 0 and the shift is sign preserving,
1454 // use the shift LHS operand instead.
1455 ICmpInst::Predicate pred = ICI.getPredicate();
1456 if (isSignTest(pred, RHS) &&
1457 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1458 return new ICmpInst(pred,
1459 LHSI->getOperand(0),
1460 Constant::getNullValue(RHS->getType()));
1462 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1463 bool TrueIfSigned = false;
1464 if (LHSI->hasOneUse() &&
1465 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1466 // (X << 31) <s 0 --> (X&1) != 0
1467 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1468 APInt::getOneBitSet(TypeBits,
1469 TypeBits-ShAmt->getZExtValue()-1));
1471 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1472 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1473 And, Constant::getNullValue(And->getType()));
1476 // Transform (icmp pred iM (shl iM %v, N), CI)
1477 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1478 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1479 // This enables to get rid of the shift in favor of a trunc which can be
1480 // free on the target. It has the additional benefit of comparing to a
1481 // smaller constant, which will be target friendly.
1482 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1483 if (LHSI->hasOneUse() &&
1484 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1485 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1486 Constant *NCI = ConstantExpr::getTrunc(
1487 ConstantExpr::getAShr(RHS,
1488 ConstantInt::get(RHS->getType(), Amt)),
1490 return new ICmpInst(ICI.getPredicate(),
1491 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1498 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1499 case Instruction::AShr: {
1500 // Handle equality comparisons of shift-by-constant.
1501 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1502 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1503 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1507 // Handle exact shr's.
1508 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1509 if (RHSV.isMinValue())
1510 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1515 case Instruction::SDiv:
1516 case Instruction::UDiv:
1517 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1518 // Fold this div into the comparison, producing a range check.
1519 // Determine, based on the divide type, what the range is being
1520 // checked. If there is an overflow on the low or high side, remember
1521 // it, otherwise compute the range [low, hi) bounding the new value.
1522 // See: InsertRangeTest above for the kinds of replacements possible.
1523 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1524 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1529 case Instruction::Sub: {
1530 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1532 const APInt &LHSV = LHSC->getValue();
1534 // C1-X <u C2 -> (X|(C2-1)) == C1
1535 // iff C1 & (C2-1) == C2-1
1536 // C2 is a power of 2
1537 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1538 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1539 return new ICmpInst(ICmpInst::ICMP_EQ,
1540 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1543 // C1-X >u C2 -> (X|C2) != C1
1544 // iff C1 & C2 == C2
1545 // C2+1 is a power of 2
1546 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1547 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1548 return new ICmpInst(ICmpInst::ICMP_NE,
1549 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1553 case Instruction::Add:
1554 // Fold: icmp pred (add X, C1), C2
1555 if (!ICI.isEquality()) {
1556 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1558 const APInt &LHSV = LHSC->getValue();
1560 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1563 if (ICI.isSigned()) {
1564 if (CR.getLower().isSignBit()) {
1565 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1566 Builder->getInt(CR.getUpper()));
1567 } else if (CR.getUpper().isSignBit()) {
1568 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1569 Builder->getInt(CR.getLower()));
1572 if (CR.getLower().isMinValue()) {
1573 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1574 Builder->getInt(CR.getUpper()));
1575 } else if (CR.getUpper().isMinValue()) {
1576 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1577 Builder->getInt(CR.getLower()));
1581 // X-C1 <u C2 -> (X & -C2) == C1
1582 // iff C1 & (C2-1) == 0
1583 // C2 is a power of 2
1584 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1585 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1586 return new ICmpInst(ICmpInst::ICMP_EQ,
1587 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1588 ConstantExpr::getNeg(LHSC));
1590 // X-C1 >u C2 -> (X & ~C2) != C1
1592 // C2+1 is a power of 2
1593 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1594 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1595 return new ICmpInst(ICmpInst::ICMP_NE,
1596 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1597 ConstantExpr::getNeg(LHSC));
1602 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1603 if (ICI.isEquality()) {
1604 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1606 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1607 // the second operand is a constant, simplify a bit.
1608 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1609 switch (BO->getOpcode()) {
1610 case Instruction::SRem:
1611 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1612 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1613 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1614 if (V.sgt(1) && V.isPowerOf2()) {
1616 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1618 return new ICmpInst(ICI.getPredicate(), NewRem,
1619 Constant::getNullValue(BO->getType()));
1623 case Instruction::Add:
1624 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1625 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1626 if (BO->hasOneUse())
1627 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1628 ConstantExpr::getSub(RHS, BOp1C));
1629 } else if (RHSV == 0) {
1630 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1631 // efficiently invertible, or if the add has just this one use.
1632 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1634 if (Value *NegVal = dyn_castNegVal(BOp1))
1635 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1636 if (Value *NegVal = dyn_castNegVal(BOp0))
1637 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1638 if (BO->hasOneUse()) {
1639 Value *Neg = Builder->CreateNeg(BOp1);
1641 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1645 case Instruction::Xor:
1646 // For the xor case, we can xor two constants together, eliminating
1647 // the explicit xor.
1648 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1649 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1650 ConstantExpr::getXor(RHS, BOC));
1651 } else if (RHSV == 0) {
1652 // Replace ((xor A, B) != 0) with (A != B)
1653 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1657 case Instruction::Sub:
1658 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1659 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1660 if (BO->hasOneUse())
1661 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1662 ConstantExpr::getSub(BOp0C, RHS));
1663 } else if (RHSV == 0) {
1664 // Replace ((sub A, B) != 0) with (A != B)
1665 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1669 case Instruction::Or:
1670 // If bits are being or'd in that are not present in the constant we
1671 // are comparing against, then the comparison could never succeed!
1672 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1673 Constant *NotCI = ConstantExpr::getNot(RHS);
1674 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1675 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1679 case Instruction::And:
1680 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1681 // If bits are being compared against that are and'd out, then the
1682 // comparison can never succeed!
1683 if ((RHSV & ~BOC->getValue()) != 0)
1684 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1686 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1687 if (RHS == BOC && RHSV.isPowerOf2())
1688 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1689 ICmpInst::ICMP_NE, LHSI,
1690 Constant::getNullValue(RHS->getType()));
1692 // Don't perform the following transforms if the AND has multiple uses
1693 if (!BO->hasOneUse())
1696 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1697 if (BOC->getValue().isSignBit()) {
1698 Value *X = BO->getOperand(0);
1699 Constant *Zero = Constant::getNullValue(X->getType());
1700 ICmpInst::Predicate pred = isICMP_NE ?
1701 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1702 return new ICmpInst(pred, X, Zero);
1705 // ((X & ~7) == 0) --> X < 8
1706 if (RHSV == 0 && isHighOnes(BOC)) {
1707 Value *X = BO->getOperand(0);
1708 Constant *NegX = ConstantExpr::getNeg(BOC);
1709 ICmpInst::Predicate pred = isICMP_NE ?
1710 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1711 return new ICmpInst(pred, X, NegX);
1715 case Instruction::Mul:
1716 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1717 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1718 // The trivial case (mul X, 0) is handled by InstSimplify
1719 // General case : (mul X, C) != 0 iff X != 0
1720 // (mul X, C) == 0 iff X == 0
1722 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1723 Constant::getNullValue(RHS->getType()));
1729 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1730 // Handle icmp {eq|ne} <intrinsic>, intcst.
1731 switch (II->getIntrinsicID()) {
1732 case Intrinsic::bswap:
1734 ICI.setOperand(0, II->getArgOperand(0));
1735 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1737 case Intrinsic::ctlz:
1738 case Intrinsic::cttz:
1739 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1740 if (RHSV == RHS->getType()->getBitWidth()) {
1742 ICI.setOperand(0, II->getArgOperand(0));
1743 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1747 case Intrinsic::ctpop:
1748 // popcount(A) == 0 -> A == 0 and likewise for !=
1749 if (RHS->isZero()) {
1751 ICI.setOperand(0, II->getArgOperand(0));
1752 ICI.setOperand(1, RHS);
1764 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1765 /// We only handle extending casts so far.
1767 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1768 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1769 Value *LHSCIOp = LHSCI->getOperand(0);
1770 Type *SrcTy = LHSCIOp->getType();
1771 Type *DestTy = LHSCI->getType();
1774 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1775 // integer type is the same size as the pointer type.
1776 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1777 TD->getPointerSizeInBits() ==
1778 cast<IntegerType>(DestTy)->getBitWidth()) {
1780 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1781 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1782 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1783 RHSOp = RHSC->getOperand(0);
1784 // If the pointer types don't match, insert a bitcast.
1785 if (LHSCIOp->getType() != RHSOp->getType())
1786 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1790 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1793 // The code below only handles extension cast instructions, so far.
1795 if (LHSCI->getOpcode() != Instruction::ZExt &&
1796 LHSCI->getOpcode() != Instruction::SExt)
1799 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1800 bool isSignedCmp = ICI.isSigned();
1802 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1803 // Not an extension from the same type?
1804 RHSCIOp = CI->getOperand(0);
1805 if (RHSCIOp->getType() != LHSCIOp->getType())
1808 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1809 // and the other is a zext), then we can't handle this.
1810 if (CI->getOpcode() != LHSCI->getOpcode())
1813 // Deal with equality cases early.
1814 if (ICI.isEquality())
1815 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1817 // A signed comparison of sign extended values simplifies into a
1818 // signed comparison.
1819 if (isSignedCmp && isSignedExt)
1820 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1822 // The other three cases all fold into an unsigned comparison.
1823 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1826 // If we aren't dealing with a constant on the RHS, exit early
1827 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1831 // Compute the constant that would happen if we truncated to SrcTy then
1832 // reextended to DestTy.
1833 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1834 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1837 // If the re-extended constant didn't change...
1839 // Deal with equality cases early.
1840 if (ICI.isEquality())
1841 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1843 // A signed comparison of sign extended values simplifies into a
1844 // signed comparison.
1845 if (isSignedExt && isSignedCmp)
1846 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1848 // The other three cases all fold into an unsigned comparison.
1849 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1852 // The re-extended constant changed so the constant cannot be represented
1853 // in the shorter type. Consequently, we cannot emit a simple comparison.
1854 // All the cases that fold to true or false will have already been handled
1855 // by SimplifyICmpInst, so only deal with the tricky case.
1857 if (isSignedCmp || !isSignedExt)
1860 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1861 // should have been folded away previously and not enter in here.
1863 // We're performing an unsigned comp with a sign extended value.
1864 // This is true if the input is >= 0. [aka >s -1]
1865 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1866 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1868 // Finally, return the value computed.
1869 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1870 return ReplaceInstUsesWith(ICI, Result);
1872 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1873 return BinaryOperator::CreateNot(Result);
1876 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1877 /// I = icmp ugt (add (add A, B), CI2), CI1
1878 /// If this is of the form:
1880 /// if (sum+128 >u 255)
1881 /// Then replace it with llvm.sadd.with.overflow.i8.
1883 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1884 ConstantInt *CI2, ConstantInt *CI1,
1886 // The transformation we're trying to do here is to transform this into an
1887 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1888 // with a narrower add, and discard the add-with-constant that is part of the
1889 // range check (if we can't eliminate it, this isn't profitable).
1891 // In order to eliminate the add-with-constant, the compare can be its only
1893 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1894 if (!AddWithCst->hasOneUse()) return 0;
1896 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1897 if (!CI2->getValue().isPowerOf2()) return 0;
1898 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1899 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1901 // The width of the new add formed is 1 more than the bias.
1904 // Check to see that CI1 is an all-ones value with NewWidth bits.
1905 if (CI1->getBitWidth() == NewWidth ||
1906 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1909 // This is only really a signed overflow check if the inputs have been
1910 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1911 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1912 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1913 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1914 IC.ComputeNumSignBits(B) < NeededSignBits)
1917 // In order to replace the original add with a narrower
1918 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1919 // and truncates that discard the high bits of the add. Verify that this is
1921 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1922 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1924 if (*UI == AddWithCst) continue;
1926 // Only accept truncates for now. We would really like a nice recursive
1927 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1928 // chain to see which bits of a value are actually demanded. If the
1929 // original add had another add which was then immediately truncated, we
1930 // could still do the transformation.
1931 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1933 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1936 // If the pattern matches, truncate the inputs to the narrower type and
1937 // use the sadd_with_overflow intrinsic to efficiently compute both the
1938 // result and the overflow bit.
1939 Module *M = I.getParent()->getParent()->getParent();
1941 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1942 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1945 InstCombiner::BuilderTy *Builder = IC.Builder;
1947 // Put the new code above the original add, in case there are any uses of the
1948 // add between the add and the compare.
1949 Builder->SetInsertPoint(OrigAdd);
1951 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1952 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1953 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1954 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1955 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1957 // The inner add was the result of the narrow add, zero extended to the
1958 // wider type. Replace it with the result computed by the intrinsic.
1959 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1961 // The original icmp gets replaced with the overflow value.
1962 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1965 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1967 // Don't bother doing this transformation for pointers, don't do it for
1969 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1971 // If the add is a constant expr, then we don't bother transforming it.
1972 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1973 if (OrigAdd == 0) return 0;
1975 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1977 // Put the new code above the original add, in case there are any uses of the
1978 // add between the add and the compare.
1979 InstCombiner::BuilderTy *Builder = IC.Builder;
1980 Builder->SetInsertPoint(OrigAdd);
1982 Module *M = I.getParent()->getParent()->getParent();
1983 Type *Ty = LHS->getType();
1984 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1985 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1986 Value *Add = Builder->CreateExtractValue(Call, 0);
1988 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1990 // The original icmp gets replaced with the overflow value.
1991 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1994 // DemandedBitsLHSMask - When performing a comparison against a constant,
1995 // it is possible that not all the bits in the LHS are demanded. This helper
1996 // method computes the mask that IS demanded.
1997 static APInt DemandedBitsLHSMask(ICmpInst &I,
1998 unsigned BitWidth, bool isSignCheck) {
2000 return APInt::getSignBit(BitWidth);
2002 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2003 if (!CI) return APInt::getAllOnesValue(BitWidth);
2004 const APInt &RHS = CI->getValue();
2006 switch (I.getPredicate()) {
2007 // For a UGT comparison, we don't care about any bits that
2008 // correspond to the trailing ones of the comparand. The value of these
2009 // bits doesn't impact the outcome of the comparison, because any value
2010 // greater than the RHS must differ in a bit higher than these due to carry.
2011 case ICmpInst::ICMP_UGT: {
2012 unsigned trailingOnes = RHS.countTrailingOnes();
2013 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2017 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2018 // Any value less than the RHS must differ in a higher bit because of carries.
2019 case ICmpInst::ICMP_ULT: {
2020 unsigned trailingZeros = RHS.countTrailingZeros();
2021 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2026 return APInt::getAllOnesValue(BitWidth);
2031 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2032 bool Changed = false;
2033 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2035 /// Orders the operands of the compare so that they are listed from most
2036 /// complex to least complex. This puts constants before unary operators,
2037 /// before binary operators.
2038 if (getComplexity(Op0) < getComplexity(Op1)) {
2040 std::swap(Op0, Op1);
2044 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
2045 return ReplaceInstUsesWith(I, V);
2047 // comparing -val or val with non-zero is the same as just comparing val
2048 // ie, abs(val) != 0 -> val != 0
2049 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2051 Value *Cond, *SelectTrue, *SelectFalse;
2052 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2053 m_Value(SelectFalse)))) {
2054 if (Value *V = dyn_castNegVal(SelectTrue)) {
2055 if (V == SelectFalse)
2056 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2058 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2059 if (V == SelectTrue)
2060 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2065 Type *Ty = Op0->getType();
2067 // icmp's with boolean values can always be turned into bitwise operations
2068 if (Ty->isIntegerTy(1)) {
2069 switch (I.getPredicate()) {
2070 default: llvm_unreachable("Invalid icmp instruction!");
2071 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2072 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2073 return BinaryOperator::CreateNot(Xor);
2075 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2076 return BinaryOperator::CreateXor(Op0, Op1);
2078 case ICmpInst::ICMP_UGT:
2079 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2081 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2082 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2083 return BinaryOperator::CreateAnd(Not, Op1);
2085 case ICmpInst::ICMP_SGT:
2086 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2088 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2089 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2090 return BinaryOperator::CreateAnd(Not, Op0);
2092 case ICmpInst::ICMP_UGE:
2093 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2095 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2096 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2097 return BinaryOperator::CreateOr(Not, Op1);
2099 case ICmpInst::ICMP_SGE:
2100 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2102 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2103 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2104 return BinaryOperator::CreateOr(Not, Op0);
2109 unsigned BitWidth = 0;
2110 if (Ty->isIntOrIntVectorTy())
2111 BitWidth = Ty->getScalarSizeInBits();
2112 else if (TD) // Pointers require TD info to get their size.
2113 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
2115 bool isSignBit = false;
2117 // See if we are doing a comparison with a constant.
2118 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2119 Value *A = 0, *B = 0;
2121 // Match the following pattern, which is a common idiom when writing
2122 // overflow-safe integer arithmetic function. The source performs an
2123 // addition in wider type, and explicitly checks for overflow using
2124 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2125 // sadd_with_overflow intrinsic.
2127 // TODO: This could probably be generalized to handle other overflow-safe
2128 // operations if we worked out the formulas to compute the appropriate
2132 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2134 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2135 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2136 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2137 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2141 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2142 if (I.isEquality() && CI->isZero() &&
2143 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2144 // (icmp cond A B) if cond is equality
2145 return new ICmpInst(I.getPredicate(), A, B);
2148 // If we have an icmp le or icmp ge instruction, turn it into the
2149 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2150 // them being folded in the code below. The SimplifyICmpInst code has
2151 // already handled the edge cases for us, so we just assert on them.
2152 switch (I.getPredicate()) {
2154 case ICmpInst::ICMP_ULE:
2155 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2156 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2157 Builder->getInt(CI->getValue()+1));
2158 case ICmpInst::ICMP_SLE:
2159 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2160 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2161 Builder->getInt(CI->getValue()+1));
2162 case ICmpInst::ICMP_UGE:
2163 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2164 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2165 Builder->getInt(CI->getValue()-1));
2166 case ICmpInst::ICMP_SGE:
2167 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2168 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2169 Builder->getInt(CI->getValue()-1));
2172 // If this comparison is a normal comparison, it demands all
2173 // bits, if it is a sign bit comparison, it only demands the sign bit.
2175 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2178 // See if we can fold the comparison based on range information we can get
2179 // by checking whether bits are known to be zero or one in the input.
2180 if (BitWidth != 0) {
2181 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2182 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2184 if (SimplifyDemandedBits(I.getOperandUse(0),
2185 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2186 Op0KnownZero, Op0KnownOne, 0))
2188 if (SimplifyDemandedBits(I.getOperandUse(1),
2189 APInt::getAllOnesValue(BitWidth),
2190 Op1KnownZero, Op1KnownOne, 0))
2193 // Given the known and unknown bits, compute a range that the LHS could be
2194 // in. Compute the Min, Max and RHS values based on the known bits. For the
2195 // EQ and NE we use unsigned values.
2196 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2197 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2199 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2201 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2204 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2206 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2210 // If Min and Max are known to be the same, then SimplifyDemandedBits
2211 // figured out that the LHS is a constant. Just constant fold this now so
2212 // that code below can assume that Min != Max.
2213 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2214 return new ICmpInst(I.getPredicate(),
2215 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2216 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2217 return new ICmpInst(I.getPredicate(), Op0,
2218 ConstantInt::get(Op1->getType(), Op1Min));
2220 // Based on the range information we know about the LHS, see if we can
2221 // simplify this comparison. For example, (x&4) < 8 is always true.
2222 switch (I.getPredicate()) {
2223 default: llvm_unreachable("Unknown icmp opcode!");
2224 case ICmpInst::ICMP_EQ: {
2225 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2226 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2228 // If all bits are known zero except for one, then we know at most one
2229 // bit is set. If the comparison is against zero, then this is a check
2230 // to see if *that* bit is set.
2231 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2232 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2233 // If the LHS is an AND with the same constant, look through it.
2235 ConstantInt *LHSC = 0;
2236 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2237 LHSC->getValue() != Op0KnownZeroInverted)
2240 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2241 // then turn "((1 << x)&8) == 0" into "x != 3".
2243 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2244 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2245 return new ICmpInst(ICmpInst::ICMP_NE, X,
2246 ConstantInt::get(X->getType(), CmpVal));
2249 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2250 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2252 if (Op0KnownZeroInverted == 1 &&
2253 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2254 return new ICmpInst(ICmpInst::ICMP_NE, X,
2255 ConstantInt::get(X->getType(),
2256 CI->countTrailingZeros()));
2261 case ICmpInst::ICMP_NE: {
2262 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2263 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2265 // If all bits are known zero except for one, then we know at most one
2266 // bit is set. If the comparison is against zero, then this is a check
2267 // to see if *that* bit is set.
2268 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2269 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2270 // If the LHS is an AND with the same constant, look through it.
2272 ConstantInt *LHSC = 0;
2273 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2274 LHSC->getValue() != Op0KnownZeroInverted)
2277 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2278 // then turn "((1 << x)&8) != 0" into "x == 3".
2280 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2281 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2282 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2283 ConstantInt::get(X->getType(), CmpVal));
2286 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2287 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2289 if (Op0KnownZeroInverted == 1 &&
2290 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2291 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2292 ConstantInt::get(X->getType(),
2293 CI->countTrailingZeros()));
2298 case ICmpInst::ICMP_ULT:
2299 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2300 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2301 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2302 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2303 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2304 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2305 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2306 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2307 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2308 Builder->getInt(CI->getValue()-1));
2310 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2311 if (CI->isMinValue(true))
2312 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2313 Constant::getAllOnesValue(Op0->getType()));
2316 case ICmpInst::ICMP_UGT:
2317 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2318 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2319 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2320 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2322 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2323 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2324 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2325 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2326 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2327 Builder->getInt(CI->getValue()+1));
2329 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2330 if (CI->isMaxValue(true))
2331 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2332 Constant::getNullValue(Op0->getType()));
2335 case ICmpInst::ICMP_SLT:
2336 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2337 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2338 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2339 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2340 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2341 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2342 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2343 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2344 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2345 Builder->getInt(CI->getValue()-1));
2348 case ICmpInst::ICMP_SGT:
2349 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2350 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2351 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2352 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2354 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2355 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2356 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2357 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2358 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2359 Builder->getInt(CI->getValue()+1));
2362 case ICmpInst::ICMP_SGE:
2363 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2364 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2365 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2366 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2367 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2369 case ICmpInst::ICMP_SLE:
2370 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2371 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2372 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2373 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2374 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2376 case ICmpInst::ICMP_UGE:
2377 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2378 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2379 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2380 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2381 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2383 case ICmpInst::ICMP_ULE:
2384 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2385 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2386 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2387 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2388 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2392 // Turn a signed comparison into an unsigned one if both operands
2393 // are known to have the same sign.
2395 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2396 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2397 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2400 // Test if the ICmpInst instruction is used exclusively by a select as
2401 // part of a minimum or maximum operation. If so, refrain from doing
2402 // any other folding. This helps out other analyses which understand
2403 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2404 // and CodeGen. And in this case, at least one of the comparison
2405 // operands has at least one user besides the compare (the select),
2406 // which would often largely negate the benefit of folding anyway.
2408 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2409 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2410 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2413 // See if we are doing a comparison between a constant and an instruction that
2414 // can be folded into the comparison.
2415 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2416 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2417 // instruction, see if that instruction also has constants so that the
2418 // instruction can be folded into the icmp
2419 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2420 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2424 // Handle icmp with constant (but not simple integer constant) RHS
2425 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2426 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2427 switch (LHSI->getOpcode()) {
2428 case Instruction::GetElementPtr:
2429 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2430 if (RHSC->isNullValue() &&
2431 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2432 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2433 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2435 case Instruction::PHI:
2436 // Only fold icmp into the PHI if the phi and icmp are in the same
2437 // block. If in the same block, we're encouraging jump threading. If
2438 // not, we are just pessimizing the code by making an i1 phi.
2439 if (LHSI->getParent() == I.getParent())
2440 if (Instruction *NV = FoldOpIntoPhi(I))
2443 case Instruction::Select: {
2444 // If either operand of the select is a constant, we can fold the
2445 // comparison into the select arms, which will cause one to be
2446 // constant folded and the select turned into a bitwise or.
2447 Value *Op1 = 0, *Op2 = 0;
2448 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2449 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2450 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2451 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2453 // We only want to perform this transformation if it will not lead to
2454 // additional code. This is true if either both sides of the select
2455 // fold to a constant (in which case the icmp is replaced with a select
2456 // which will usually simplify) or this is the only user of the
2457 // select (in which case we are trading a select+icmp for a simpler
2459 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2461 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2464 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2466 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2470 case Instruction::IntToPtr:
2471 // icmp pred inttoptr(X), null -> icmp pred X, 0
2472 if (RHSC->isNullValue() && TD &&
2473 TD->getIntPtrType(RHSC->getContext()) ==
2474 LHSI->getOperand(0)->getType())
2475 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2476 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2479 case Instruction::Load:
2480 // Try to optimize things like "A[i] > 4" to index computations.
2481 if (GetElementPtrInst *GEP =
2482 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2483 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2484 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2485 !cast<LoadInst>(LHSI)->isVolatile())
2486 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2493 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2494 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2495 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2497 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2498 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2499 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2502 // Test to see if the operands of the icmp are casted versions of other
2503 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2505 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2506 if (Op0->getType()->isPointerTy() &&
2507 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2508 // We keep moving the cast from the left operand over to the right
2509 // operand, where it can often be eliminated completely.
2510 Op0 = CI->getOperand(0);
2512 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2513 // so eliminate it as well.
2514 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2515 Op1 = CI2->getOperand(0);
2517 // If Op1 is a constant, we can fold the cast into the constant.
2518 if (Op0->getType() != Op1->getType()) {
2519 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2520 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2522 // Otherwise, cast the RHS right before the icmp
2523 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2526 return new ICmpInst(I.getPredicate(), Op0, Op1);
2530 if (isa<CastInst>(Op0)) {
2531 // Handle the special case of: icmp (cast bool to X), <cst>
2532 // This comes up when you have code like
2535 // For generality, we handle any zero-extension of any operand comparison
2536 // with a constant or another cast from the same type.
2537 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2538 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2542 // Special logic for binary operators.
2543 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2544 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2546 CmpInst::Predicate Pred = I.getPredicate();
2547 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2548 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2549 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2550 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2551 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2552 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2553 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2554 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2555 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2557 // Analyze the case when either Op0 or Op1 is an add instruction.
2558 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2559 Value *A = 0, *B = 0, *C = 0, *D = 0;
2560 if (BO0 && BO0->getOpcode() == Instruction::Add)
2561 A = BO0->getOperand(0), B = BO0->getOperand(1);
2562 if (BO1 && BO1->getOpcode() == Instruction::Add)
2563 C = BO1->getOperand(0), D = BO1->getOperand(1);
2565 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2566 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2567 return new ICmpInst(Pred, A == Op1 ? B : A,
2568 Constant::getNullValue(Op1->getType()));
2570 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2571 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2572 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2575 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2576 if (A && C && (A == C || A == D || B == C || B == D) &&
2577 NoOp0WrapProblem && NoOp1WrapProblem &&
2578 // Try not to increase register pressure.
2579 BO0->hasOneUse() && BO1->hasOneUse()) {
2580 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2583 // C + B == C + D -> B == D
2586 } else if (A == D) {
2587 // D + B == C + D -> B == C
2590 } else if (B == C) {
2591 // A + C == C + D -> A == D
2596 // A + D == C + D -> A == C
2600 return new ICmpInst(Pred, Y, Z);
2603 // icmp slt (X + -1), Y -> icmp sle X, Y
2604 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2605 match(B, m_AllOnes()))
2606 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2608 // icmp sge (X + -1), Y -> icmp sgt X, Y
2609 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2610 match(B, m_AllOnes()))
2611 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2613 // icmp sle (X + 1), Y -> icmp slt X, Y
2614 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
2616 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2618 // icmp sgt (X + 1), Y -> icmp sge X, Y
2619 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
2621 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2623 // if C1 has greater magnitude than C2:
2624 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2625 // s.t. C3 = C1 - C2
2627 // if C2 has greater magnitude than C1:
2628 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2629 // s.t. C3 = C2 - C1
2630 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2631 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2632 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2633 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2634 const APInt &AP1 = C1->getValue();
2635 const APInt &AP2 = C2->getValue();
2636 if (AP1.isNegative() == AP2.isNegative()) {
2637 APInt AP1Abs = C1->getValue().abs();
2638 APInt AP2Abs = C2->getValue().abs();
2639 if (AP1Abs.uge(AP2Abs)) {
2640 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2641 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2642 return new ICmpInst(Pred, NewAdd, C);
2644 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2645 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2646 return new ICmpInst(Pred, A, NewAdd);
2652 // Analyze the case when either Op0 or Op1 is a sub instruction.
2653 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2654 A = 0; B = 0; C = 0; D = 0;
2655 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2656 A = BO0->getOperand(0), B = BO0->getOperand(1);
2657 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2658 C = BO1->getOperand(0), D = BO1->getOperand(1);
2660 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2661 if (A == Op1 && NoOp0WrapProblem)
2662 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2664 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2665 if (C == Op0 && NoOp1WrapProblem)
2666 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2668 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2669 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2670 // Try not to increase register pressure.
2671 BO0->hasOneUse() && BO1->hasOneUse())
2672 return new ICmpInst(Pred, A, C);
2674 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2675 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2676 // Try not to increase register pressure.
2677 BO0->hasOneUse() && BO1->hasOneUse())
2678 return new ICmpInst(Pred, D, B);
2680 BinaryOperator *SRem = NULL;
2681 // icmp (srem X, Y), Y
2682 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2683 Op1 == BO0->getOperand(1))
2685 // icmp Y, (srem X, Y)
2686 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2687 Op0 == BO1->getOperand(1))
2690 // We don't check hasOneUse to avoid increasing register pressure because
2691 // the value we use is the same value this instruction was already using.
2692 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2694 case ICmpInst::ICMP_EQ:
2695 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2696 case ICmpInst::ICMP_NE:
2697 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2698 case ICmpInst::ICMP_SGT:
2699 case ICmpInst::ICMP_SGE:
2700 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2701 Constant::getAllOnesValue(SRem->getType()));
2702 case ICmpInst::ICMP_SLT:
2703 case ICmpInst::ICMP_SLE:
2704 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2705 Constant::getNullValue(SRem->getType()));
2709 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2710 BO0->hasOneUse() && BO1->hasOneUse() &&
2711 BO0->getOperand(1) == BO1->getOperand(1)) {
2712 switch (BO0->getOpcode()) {
2714 case Instruction::Add:
2715 case Instruction::Sub:
2716 case Instruction::Xor:
2717 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2718 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2719 BO1->getOperand(0));
2720 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2721 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2722 if (CI->getValue().isSignBit()) {
2723 ICmpInst::Predicate Pred = I.isSigned()
2724 ? I.getUnsignedPredicate()
2725 : I.getSignedPredicate();
2726 return new ICmpInst(Pred, BO0->getOperand(0),
2727 BO1->getOperand(0));
2730 if (CI->isMaxValue(true)) {
2731 ICmpInst::Predicate Pred = I.isSigned()
2732 ? I.getUnsignedPredicate()
2733 : I.getSignedPredicate();
2734 Pred = I.getSwappedPredicate(Pred);
2735 return new ICmpInst(Pred, BO0->getOperand(0),
2736 BO1->getOperand(0));
2740 case Instruction::Mul:
2741 if (!I.isEquality())
2744 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2745 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2746 // Mask = -1 >> count-trailing-zeros(Cst).
2747 if (!CI->isZero() && !CI->isOne()) {
2748 const APInt &AP = CI->getValue();
2749 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2750 APInt::getLowBitsSet(AP.getBitWidth(),
2752 AP.countTrailingZeros()));
2753 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2754 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2755 return new ICmpInst(I.getPredicate(), And1, And2);
2759 case Instruction::UDiv:
2760 case Instruction::LShr:
2764 case Instruction::SDiv:
2765 case Instruction::AShr:
2766 if (!BO0->isExact() || !BO1->isExact())
2768 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2769 BO1->getOperand(0));
2770 case Instruction::Shl: {
2771 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2772 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2775 if (!NSW && I.isSigned())
2777 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2778 BO1->getOperand(0));
2785 // Transform (A & ~B) == 0 --> (A & B) != 0
2786 // and (A & ~B) != 0 --> (A & B) == 0
2787 // if A is a power of 2.
2788 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2789 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
2790 return new ICmpInst(I.getInversePredicate(),
2791 Builder->CreateAnd(A, B),
2794 // ~x < ~y --> y < x
2795 // ~x < cst --> ~cst < x
2796 if (match(Op0, m_Not(m_Value(A)))) {
2797 if (match(Op1, m_Not(m_Value(B))))
2798 return new ICmpInst(I.getPredicate(), B, A);
2799 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2800 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2803 // (a+b) <u a --> llvm.uadd.with.overflow.
2804 // (a+b) <u b --> llvm.uadd.with.overflow.
2805 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2806 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2807 (Op1 == A || Op1 == B))
2808 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2811 // a >u (a+b) --> llvm.uadd.with.overflow.
2812 // b >u (a+b) --> llvm.uadd.with.overflow.
2813 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2814 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2815 (Op0 == A || Op0 == B))
2816 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2820 if (I.isEquality()) {
2821 Value *A, *B, *C, *D;
2823 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2824 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2825 Value *OtherVal = A == Op1 ? B : A;
2826 return new ICmpInst(I.getPredicate(), OtherVal,
2827 Constant::getNullValue(A->getType()));
2830 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2831 // A^c1 == C^c2 --> A == C^(c1^c2)
2832 ConstantInt *C1, *C2;
2833 if (match(B, m_ConstantInt(C1)) &&
2834 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2835 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
2836 Value *Xor = Builder->CreateXor(C, NC);
2837 return new ICmpInst(I.getPredicate(), A, Xor);
2840 // A^B == A^D -> B == D
2841 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2842 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2843 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2844 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2848 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2849 (A == Op0 || B == Op0)) {
2850 // A == (A^B) -> B == 0
2851 Value *OtherVal = A == Op0 ? B : A;
2852 return new ICmpInst(I.getPredicate(), OtherVal,
2853 Constant::getNullValue(A->getType()));
2856 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2857 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2858 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2859 Value *X = 0, *Y = 0, *Z = 0;
2862 X = B; Y = D; Z = A;
2863 } else if (A == D) {
2864 X = B; Y = C; Z = A;
2865 } else if (B == C) {
2866 X = A; Y = D; Z = B;
2867 } else if (B == D) {
2868 X = A; Y = C; Z = B;
2871 if (X) { // Build (X^Y) & Z
2872 Op1 = Builder->CreateXor(X, Y);
2873 Op1 = Builder->CreateAnd(Op1, Z);
2874 I.setOperand(0, Op1);
2875 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2880 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2881 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2883 if ((Op0->hasOneUse() &&
2884 match(Op0, m_ZExt(m_Value(A))) &&
2885 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2886 (Op1->hasOneUse() &&
2887 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2888 match(Op1, m_ZExt(m_Value(A))))) {
2889 APInt Pow2 = Cst1->getValue() + 1;
2890 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2891 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2892 return new ICmpInst(I.getPredicate(), A,
2893 Builder->CreateTrunc(B, A->getType()));
2896 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2897 // "icmp (and X, mask), cst"
2899 if (Op0->hasOneUse() &&
2900 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2901 m_ConstantInt(ShAmt))))) &&
2902 match(Op1, m_ConstantInt(Cst1)) &&
2903 // Only do this when A has multiple uses. This is most important to do
2904 // when it exposes other optimizations.
2906 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2908 if (ShAmt < ASize) {
2910 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2913 APInt CmpV = Cst1->getValue().zext(ASize);
2916 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2917 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2923 Value *X; ConstantInt *Cst;
2925 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2926 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2929 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2930 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2932 return Changed ? &I : 0;
2940 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2942 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2945 if (!isa<ConstantFP>(RHSC)) return 0;
2946 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2948 // Get the width of the mantissa. We don't want to hack on conversions that
2949 // might lose information from the integer, e.g. "i64 -> float"
2950 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2951 if (MantissaWidth == -1) return 0; // Unknown.
2953 // Check to see that the input is converted from an integer type that is small
2954 // enough that preserves all bits. TODO: check here for "known" sign bits.
2955 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2956 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2958 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2959 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2963 // If the conversion would lose info, don't hack on this.
2964 if ((int)InputSize > MantissaWidth)
2967 // Otherwise, we can potentially simplify the comparison. We know that it
2968 // will always come through as an integer value and we know the constant is
2969 // not a NAN (it would have been previously simplified).
2970 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2972 ICmpInst::Predicate Pred;
2973 switch (I.getPredicate()) {
2974 default: llvm_unreachable("Unexpected predicate!");
2975 case FCmpInst::FCMP_UEQ:
2976 case FCmpInst::FCMP_OEQ:
2977 Pred = ICmpInst::ICMP_EQ;
2979 case FCmpInst::FCMP_UGT:
2980 case FCmpInst::FCMP_OGT:
2981 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2983 case FCmpInst::FCMP_UGE:
2984 case FCmpInst::FCMP_OGE:
2985 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2987 case FCmpInst::FCMP_ULT:
2988 case FCmpInst::FCMP_OLT:
2989 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2991 case FCmpInst::FCMP_ULE:
2992 case FCmpInst::FCMP_OLE:
2993 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2995 case FCmpInst::FCMP_UNE:
2996 case FCmpInst::FCMP_ONE:
2997 Pred = ICmpInst::ICMP_NE;
2999 case FCmpInst::FCMP_ORD:
3000 return ReplaceInstUsesWith(I, Builder->getTrue());
3001 case FCmpInst::FCMP_UNO:
3002 return ReplaceInstUsesWith(I, Builder->getFalse());
3005 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3007 // Now we know that the APFloat is a normal number, zero or inf.
3009 // See if the FP constant is too large for the integer. For example,
3010 // comparing an i8 to 300.0.
3011 unsigned IntWidth = IntTy->getScalarSizeInBits();
3014 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3015 // and large values.
3016 APFloat SMax(RHS.getSemantics());
3017 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3018 APFloat::rmNearestTiesToEven);
3019 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3020 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3021 Pred == ICmpInst::ICMP_SLE)
3022 return ReplaceInstUsesWith(I, Builder->getTrue());
3023 return ReplaceInstUsesWith(I, Builder->getFalse());
3026 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3027 // +INF and large values.
3028 APFloat UMax(RHS.getSemantics());
3029 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3030 APFloat::rmNearestTiesToEven);
3031 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3032 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3033 Pred == ICmpInst::ICMP_ULE)
3034 return ReplaceInstUsesWith(I, Builder->getTrue());
3035 return ReplaceInstUsesWith(I, Builder->getFalse());
3040 // See if the RHS value is < SignedMin.
3041 APFloat SMin(RHS.getSemantics());
3042 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3043 APFloat::rmNearestTiesToEven);
3044 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3045 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3046 Pred == ICmpInst::ICMP_SGE)
3047 return ReplaceInstUsesWith(I, Builder->getTrue());
3048 return ReplaceInstUsesWith(I, Builder->getFalse());
3051 // See if the RHS value is < UnsignedMin.
3052 APFloat SMin(RHS.getSemantics());
3053 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3054 APFloat::rmNearestTiesToEven);
3055 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3056 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3057 Pred == ICmpInst::ICMP_UGE)
3058 return ReplaceInstUsesWith(I, Builder->getTrue());
3059 return ReplaceInstUsesWith(I, Builder->getFalse());
3063 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3064 // [0, UMAX], but it may still be fractional. See if it is fractional by
3065 // casting the FP value to the integer value and back, checking for equality.
3066 // Don't do this for zero, because -0.0 is not fractional.
3067 Constant *RHSInt = LHSUnsigned
3068 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3069 : ConstantExpr::getFPToSI(RHSC, IntTy);
3070 if (!RHS.isZero()) {
3071 bool Equal = LHSUnsigned
3072 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3073 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3075 // If we had a comparison against a fractional value, we have to adjust
3076 // the compare predicate and sometimes the value. RHSC is rounded towards
3077 // zero at this point.
3079 default: llvm_unreachable("Unexpected integer comparison!");
3080 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3081 return ReplaceInstUsesWith(I, Builder->getTrue());
3082 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3083 return ReplaceInstUsesWith(I, Builder->getFalse());
3084 case ICmpInst::ICMP_ULE:
3085 // (float)int <= 4.4 --> int <= 4
3086 // (float)int <= -4.4 --> false
3087 if (RHS.isNegative())
3088 return ReplaceInstUsesWith(I, Builder->getFalse());
3090 case ICmpInst::ICMP_SLE:
3091 // (float)int <= 4.4 --> int <= 4
3092 // (float)int <= -4.4 --> int < -4
3093 if (RHS.isNegative())
3094 Pred = ICmpInst::ICMP_SLT;
3096 case ICmpInst::ICMP_ULT:
3097 // (float)int < -4.4 --> false
3098 // (float)int < 4.4 --> int <= 4
3099 if (RHS.isNegative())
3100 return ReplaceInstUsesWith(I, Builder->getFalse());
3101 Pred = ICmpInst::ICMP_ULE;
3103 case ICmpInst::ICMP_SLT:
3104 // (float)int < -4.4 --> int < -4
3105 // (float)int < 4.4 --> int <= 4
3106 if (!RHS.isNegative())
3107 Pred = ICmpInst::ICMP_SLE;
3109 case ICmpInst::ICMP_UGT:
3110 // (float)int > 4.4 --> int > 4
3111 // (float)int > -4.4 --> true
3112 if (RHS.isNegative())
3113 return ReplaceInstUsesWith(I, Builder->getTrue());
3115 case ICmpInst::ICMP_SGT:
3116 // (float)int > 4.4 --> int > 4
3117 // (float)int > -4.4 --> int >= -4
3118 if (RHS.isNegative())
3119 Pred = ICmpInst::ICMP_SGE;
3121 case ICmpInst::ICMP_UGE:
3122 // (float)int >= -4.4 --> true
3123 // (float)int >= 4.4 --> int > 4
3124 if (RHS.isNegative())
3125 return ReplaceInstUsesWith(I, Builder->getTrue());
3126 Pred = ICmpInst::ICMP_UGT;
3128 case ICmpInst::ICMP_SGE:
3129 // (float)int >= -4.4 --> int >= -4
3130 // (float)int >= 4.4 --> int > 4
3131 if (!RHS.isNegative())
3132 Pred = ICmpInst::ICMP_SGT;
3138 // Lower this FP comparison into an appropriate integer version of the
3140 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3143 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3144 bool Changed = false;
3146 /// Orders the operands of the compare so that they are listed from most
3147 /// complex to least complex. This puts constants before unary operators,
3148 /// before binary operators.
3149 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3154 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3156 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
3157 return ReplaceInstUsesWith(I, V);
3159 // Simplify 'fcmp pred X, X'
3161 switch (I.getPredicate()) {
3162 default: llvm_unreachable("Unknown predicate!");
3163 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3164 case FCmpInst::FCMP_ULT: // True if unordered or less than
3165 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3166 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3167 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3168 I.setPredicate(FCmpInst::FCMP_UNO);
3169 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3172 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3173 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3174 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3175 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3176 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3177 I.setPredicate(FCmpInst::FCMP_ORD);
3178 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3183 // Handle fcmp with constant RHS
3184 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3185 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3186 switch (LHSI->getOpcode()) {
3187 case Instruction::FPExt: {
3188 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3189 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3190 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3194 const fltSemantics *Sem;
3195 // FIXME: This shouldn't be here.
3196 if (LHSExt->getSrcTy()->isHalfTy())
3197 Sem = &APFloat::IEEEhalf;
3198 else if (LHSExt->getSrcTy()->isFloatTy())
3199 Sem = &APFloat::IEEEsingle;
3200 else if (LHSExt->getSrcTy()->isDoubleTy())
3201 Sem = &APFloat::IEEEdouble;
3202 else if (LHSExt->getSrcTy()->isFP128Ty())
3203 Sem = &APFloat::IEEEquad;
3204 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3205 Sem = &APFloat::x87DoubleExtended;
3206 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3207 Sem = &APFloat::PPCDoubleDouble;
3212 APFloat F = RHSF->getValueAPF();
3213 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3215 // Avoid lossy conversions and denormals. Zero is a special case
3216 // that's OK to convert.
3220 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3221 APFloat::cmpLessThan) || Fabs.isZero()))
3223 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3224 ConstantFP::get(RHSC->getContext(), F));
3227 case Instruction::PHI:
3228 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3229 // block. If in the same block, we're encouraging jump threading. If
3230 // not, we are just pessimizing the code by making an i1 phi.
3231 if (LHSI->getParent() == I.getParent())
3232 if (Instruction *NV = FoldOpIntoPhi(I))
3235 case Instruction::SIToFP:
3236 case Instruction::UIToFP:
3237 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3240 case Instruction::Select: {
3241 // If either operand of the select is a constant, we can fold the
3242 // comparison into the select arms, which will cause one to be
3243 // constant folded and the select turned into a bitwise or.
3244 Value *Op1 = 0, *Op2 = 0;
3245 if (LHSI->hasOneUse()) {
3246 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3247 // Fold the known value into the constant operand.
3248 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3249 // Insert a new FCmp of the other select operand.
3250 Op2 = Builder->CreateFCmp(I.getPredicate(),
3251 LHSI->getOperand(2), RHSC, I.getName());
3252 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3253 // Fold the known value into the constant operand.
3254 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3255 // Insert a new FCmp of the other select operand.
3256 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
3262 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3265 case Instruction::FSub: {
3266 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3268 if (match(LHSI, m_FNeg(m_Value(Op))))
3269 return new FCmpInst(I.getSwappedPredicate(), Op,
3270 ConstantExpr::getFNeg(RHSC));
3273 case Instruction::Load:
3274 if (GetElementPtrInst *GEP =
3275 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3276 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3277 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3278 !cast<LoadInst>(LHSI)->isVolatile())
3279 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3283 case Instruction::Call: {
3284 CallInst *CI = cast<CallInst>(LHSI);
3286 // Various optimization for fabs compared with zero.
3287 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3288 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3290 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3291 Func == LibFunc::fabsl) {
3292 switch (I.getPredicate()) {
3294 // fabs(x) < 0 --> false
3295 case FCmpInst::FCMP_OLT:
3296 return ReplaceInstUsesWith(I, Builder->getFalse());
3297 // fabs(x) > 0 --> x != 0
3298 case FCmpInst::FCMP_OGT:
3299 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3301 // fabs(x) <= 0 --> x == 0
3302 case FCmpInst::FCMP_OLE:
3303 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3305 // fabs(x) >= 0 --> !isnan(x)
3306 case FCmpInst::FCMP_OGE:
3307 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3309 // fabs(x) == 0 --> x == 0
3310 // fabs(x) != 0 --> x != 0
3311 case FCmpInst::FCMP_OEQ:
3312 case FCmpInst::FCMP_UEQ:
3313 case FCmpInst::FCMP_ONE:
3314 case FCmpInst::FCMP_UNE:
3315 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3324 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3326 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3327 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3329 // fcmp (fpext x), (fpext y) -> fcmp x, y
3330 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3331 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3332 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3333 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3334 RHSExt->getOperand(0));
3336 return Changed ? &I : 0;