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/DataLayout.h"
19 #include "llvm/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 // isHighOnes - Return true if the constant is of the form 1+0+.
143 // This is the same as lowones(~X).
144 static bool isHighOnes(const ConstantInt *CI) {
145 return (~CI->getValue() + 1).isPowerOf2();
148 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
149 /// set of known zero and one bits, compute the maximum and minimum values that
150 /// could have the specified known zero and known one bits, returning them in
152 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
153 const APInt& KnownOne,
154 APInt& Min, APInt& Max) {
155 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
156 KnownZero.getBitWidth() == Min.getBitWidth() &&
157 KnownZero.getBitWidth() == Max.getBitWidth() &&
158 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
159 APInt UnknownBits = ~(KnownZero|KnownOne);
161 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
162 // bit if it is unknown.
164 Max = KnownOne|UnknownBits;
166 if (UnknownBits.isNegative()) { // Sign bit is unknown
167 Min.setBit(Min.getBitWidth()-1);
168 Max.clearBit(Max.getBitWidth()-1);
172 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
173 // a set of known zero and one bits, compute the maximum and minimum values that
174 // could have the specified known zero and known one bits, returning them in
176 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
177 const APInt &KnownOne,
178 APInt &Min, APInt &Max) {
179 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
180 KnownZero.getBitWidth() == Min.getBitWidth() &&
181 KnownZero.getBitWidth() == Max.getBitWidth() &&
182 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
183 APInt UnknownBits = ~(KnownZero|KnownOne);
185 // The minimum value is when the unknown bits are all zeros.
187 // The maximum value is when the unknown bits are all ones.
188 Max = KnownOne|UnknownBits;
193 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
194 /// cmp pred (load (gep GV, ...)), cmpcst
195 /// where GV is a global variable with a constant initializer. Try to simplify
196 /// this into some simple computation that does not need the load. For example
197 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
199 /// If AndCst is non-null, then the loaded value is masked with that constant
200 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
201 Instruction *InstCombiner::
202 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
203 CmpInst &ICI, ConstantInt *AndCst) {
204 // We need TD information to know the pointer size unless this is inbounds.
205 if (!GEP->isInBounds() && TD == 0) return 0;
207 Constant *Init = GV->getInitializer();
208 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
211 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
212 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays.
214 // There are many forms of this optimization we can handle, for now, just do
215 // the simple index into a single-dimensional array.
217 // Require: GEP GV, 0, i {{, constant indices}}
218 if (GEP->getNumOperands() < 3 ||
219 !isa<ConstantInt>(GEP->getOperand(1)) ||
220 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
221 isa<Constant>(GEP->getOperand(2)))
224 // Check that indices after the variable are constants and in-range for the
225 // type they index. Collect the indices. This is typically for arrays of
227 SmallVector<unsigned, 4> LaterIndices;
229 Type *EltTy = Init->getType()->getArrayElementType();
230 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
231 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
232 if (Idx == 0) return 0; // Variable index.
234 uint64_t IdxVal = Idx->getZExtValue();
235 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
237 if (StructType *STy = dyn_cast<StructType>(EltTy))
238 EltTy = STy->getElementType(IdxVal);
239 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
240 if (IdxVal >= ATy->getNumElements()) return 0;
241 EltTy = ATy->getElementType();
243 return 0; // Unknown type.
246 LaterIndices.push_back(IdxVal);
249 enum { Overdefined = -3, Undefined = -2 };
251 // Variables for our state machines.
253 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
254 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
255 // and 87 is the second (and last) index. FirstTrueElement is -2 when
256 // undefined, otherwise set to the first true element. SecondTrueElement is
257 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
258 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
260 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
261 // form "i != 47 & i != 87". Same state transitions as for true elements.
262 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
264 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
265 /// define a state machine that triggers for ranges of values that the index
266 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
267 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
268 /// index in the range (inclusive). We use -2 for undefined here because we
269 /// use relative comparisons and don't want 0-1 to match -1.
270 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
272 // MagicBitvector - This is a magic bitvector where we set a bit if the
273 // comparison is true for element 'i'. If there are 64 elements or less in
274 // the array, this will fully represent all the comparison results.
275 uint64_t MagicBitvector = 0;
278 // Scan the array and see if one of our patterns matches.
279 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
280 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
281 Constant *Elt = Init->getAggregateElement(i);
282 if (Elt == 0) return 0;
284 // If this is indexing an array of structures, get the structure element.
285 if (!LaterIndices.empty())
286 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
288 // If the element is masked, handle it.
289 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
291 // Find out if the comparison would be true or false for the i'th element.
292 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
293 CompareRHS, TD, TLI);
294 // If the result is undef for this element, ignore it.
295 if (isa<UndefValue>(C)) {
296 // Extend range state machines to cover this element in case there is an
297 // undef in the middle of the range.
298 if (TrueRangeEnd == (int)i-1)
300 if (FalseRangeEnd == (int)i-1)
305 // If we can't compute the result for any of the elements, we have to give
306 // up evaluating the entire conditional.
307 if (!isa<ConstantInt>(C)) return 0;
309 // Otherwise, we know if the comparison is true or false for this element,
310 // update our state machines.
311 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
313 // State machine for single/double/range index comparison.
315 // Update the TrueElement state machine.
316 if (FirstTrueElement == Undefined)
317 FirstTrueElement = TrueRangeEnd = i; // First true element.
319 // Update double-compare state machine.
320 if (SecondTrueElement == Undefined)
321 SecondTrueElement = i;
323 SecondTrueElement = Overdefined;
325 // Update range state machine.
326 if (TrueRangeEnd == (int)i-1)
329 TrueRangeEnd = Overdefined;
332 // Update the FalseElement state machine.
333 if (FirstFalseElement == Undefined)
334 FirstFalseElement = FalseRangeEnd = i; // First false element.
336 // Update double-compare state machine.
337 if (SecondFalseElement == Undefined)
338 SecondFalseElement = i;
340 SecondFalseElement = Overdefined;
342 // Update range state machine.
343 if (FalseRangeEnd == (int)i-1)
346 FalseRangeEnd = Overdefined;
351 // If this element is in range, update our magic bitvector.
352 if (i < 64 && IsTrueForElt)
353 MagicBitvector |= 1ULL << i;
355 // If all of our states become overdefined, bail out early. Since the
356 // predicate is expensive, only check it every 8 elements. This is only
357 // really useful for really huge arrays.
358 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
359 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
360 FalseRangeEnd == Overdefined)
364 // Now that we've scanned the entire array, emit our new comparison(s). We
365 // order the state machines in complexity of the generated code.
366 Value *Idx = GEP->getOperand(2);
368 // If the index is larger than the pointer size of the target, truncate the
369 // index down like the GEP would do implicitly. We don't have to do this for
370 // an inbounds GEP because the index can't be out of range.
371 if (!GEP->isInBounds() &&
372 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
373 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
375 // If the comparison is only true for one or two elements, emit direct
377 if (SecondTrueElement != Overdefined) {
378 // None true -> false.
379 if (FirstTrueElement == Undefined)
380 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
382 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
384 // True for one element -> 'i == 47'.
385 if (SecondTrueElement == Undefined)
386 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
388 // True for two elements -> 'i == 47 | i == 72'.
389 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
390 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
391 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
392 return BinaryOperator::CreateOr(C1, C2);
395 // If the comparison is only false for one or two elements, emit direct
397 if (SecondFalseElement != Overdefined) {
398 // None false -> true.
399 if (FirstFalseElement == Undefined)
400 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
402 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
404 // False for one element -> 'i != 47'.
405 if (SecondFalseElement == Undefined)
406 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
408 // False for two elements -> 'i != 47 & i != 72'.
409 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
410 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
411 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
412 return BinaryOperator::CreateAnd(C1, C2);
415 // If the comparison can be replaced with a range comparison for the elements
416 // where it is true, emit the range check.
417 if (TrueRangeEnd != Overdefined) {
418 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
420 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
421 if (FirstTrueElement) {
422 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
423 Idx = Builder->CreateAdd(Idx, Offs);
426 Value *End = ConstantInt::get(Idx->getType(),
427 TrueRangeEnd-FirstTrueElement+1);
428 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
431 // False range check.
432 if (FalseRangeEnd != Overdefined) {
433 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
434 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
435 if (FirstFalseElement) {
436 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
437 Idx = Builder->CreateAdd(Idx, Offs);
440 Value *End = ConstantInt::get(Idx->getType(),
441 FalseRangeEnd-FirstFalseElement);
442 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
446 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
447 // of this load, replace it with computation that does:
448 // ((magic_cst >> i) & 1) != 0
449 if (ArrayElementCount <= 32 ||
450 (TD && ArrayElementCount <= 64 && TD->isLegalInteger(64))) {
452 if (ArrayElementCount <= 32)
453 Ty = Type::getInt32Ty(Init->getContext());
455 Ty = Type::getInt64Ty(Init->getContext());
456 Value *V = Builder->CreateIntCast(Idx, Ty, false);
457 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
458 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
459 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
466 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
467 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
468 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
469 /// be complex, and scales are involved. The above expression would also be
470 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
471 /// This later form is less amenable to optimization though, and we are allowed
472 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
474 /// If we can't emit an optimized form for this expression, this returns null.
476 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
477 DataLayout &TD = *IC.getDataLayout();
478 gep_type_iterator GTI = gep_type_begin(GEP);
480 // Check to see if this gep only has a single variable index. If so, and if
481 // any constant indices are a multiple of its scale, then we can compute this
482 // in terms of the scale of the variable index. For example, if the GEP
483 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
484 // because the expression will cross zero at the same point.
485 unsigned i, e = GEP->getNumOperands();
487 for (i = 1; i != e; ++i, ++GTI) {
488 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
489 // Compute the aggregate offset of constant indices.
490 if (CI->isZero()) continue;
492 // Handle a struct index, which adds its field offset to the pointer.
493 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
494 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
496 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
497 Offset += Size*CI->getSExtValue();
500 // Found our variable index.
505 // If there are no variable indices, we must have a constant offset, just
506 // evaluate it the general way.
507 if (i == e) return 0;
509 Value *VariableIdx = GEP->getOperand(i);
510 // Determine the scale factor of the variable element. For example, this is
511 // 4 if the variable index is into an array of i32.
512 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
514 // Verify that there are no other variable indices. If so, emit the hard way.
515 for (++i, ++GTI; i != e; ++i, ++GTI) {
516 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
519 // Compute the aggregate offset of constant indices.
520 if (CI->isZero()) continue;
522 // Handle a struct index, which adds its field offset to the pointer.
523 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
524 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
526 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
527 Offset += Size*CI->getSExtValue();
531 // Okay, we know we have a single variable index, which must be a
532 // pointer/array/vector index. If there is no offset, life is simple, return
534 unsigned IntPtrWidth = TD.getPointerSizeInBits();
536 // Cast to intptrty in case a truncation occurs. If an extension is needed,
537 // we don't need to bother extending: the extension won't affect where the
538 // computation crosses zero.
539 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
540 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
541 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
546 // Otherwise, there is an index. The computation we will do will be modulo
547 // the pointer size, so get it.
548 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
550 Offset &= PtrSizeMask;
551 VariableScale &= PtrSizeMask;
553 // To do this transformation, any constant index must be a multiple of the
554 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
555 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
556 // multiple of the variable scale.
557 int64_t NewOffs = Offset / (int64_t)VariableScale;
558 if (Offset != NewOffs*(int64_t)VariableScale)
561 // Okay, we can do this evaluation. Start by converting the index to intptr.
562 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
563 if (VariableIdx->getType() != IntPtrTy)
564 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
566 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
567 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
570 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
571 /// else. At this point we know that the GEP is on the LHS of the comparison.
572 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
573 ICmpInst::Predicate Cond,
575 // Don't transform signed compares of GEPs into index compares. Even if the
576 // GEP is inbounds, the final add of the base pointer can have signed overflow
577 // and would change the result of the icmp.
578 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
579 // the maximum signed value for the pointer type.
580 if (ICmpInst::isSigned(Cond))
583 // Look through bitcasts.
584 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
585 RHS = BCI->getOperand(0);
587 Value *PtrBase = GEPLHS->getOperand(0);
588 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
589 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
590 // This transformation (ignoring the base and scales) is valid because we
591 // know pointers can't overflow since the gep is inbounds. See if we can
592 // output an optimized form.
593 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
595 // If not, synthesize the offset the hard way.
597 Offset = EmitGEPOffset(GEPLHS);
598 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
599 Constant::getNullValue(Offset->getType()));
600 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
601 // If the base pointers are different, but the indices are the same, just
602 // compare the base pointer.
603 if (PtrBase != GEPRHS->getOperand(0)) {
604 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
605 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
606 GEPRHS->getOperand(0)->getType();
608 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
609 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
610 IndicesTheSame = false;
614 // If all indices are the same, just compare the base pointers.
616 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
617 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
619 // If we're comparing GEPs with two base pointers that only differ in type
620 // and both GEPs have only constant indices or just one use, then fold
621 // the compare with the adjusted indices.
622 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
623 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
624 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
625 PtrBase->stripPointerCasts() ==
626 GEPRHS->getOperand(0)->stripPointerCasts()) {
627 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
628 EmitGEPOffset(GEPLHS),
629 EmitGEPOffset(GEPRHS));
630 return ReplaceInstUsesWith(I, Cmp);
633 // Otherwise, the base pointers are different and the indices are
634 // different, bail out.
638 // If one of the GEPs has all zero indices, recurse.
639 bool AllZeros = true;
640 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
641 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
642 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
647 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
648 ICmpInst::getSwappedPredicate(Cond), I);
650 // If the other GEP has all zero indices, recurse.
652 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
653 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
654 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
659 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
661 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
662 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
663 // If the GEPs only differ by one index, compare it.
664 unsigned NumDifferences = 0; // Keep track of # differences.
665 unsigned DiffOperand = 0; // The operand that differs.
666 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
667 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
668 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
669 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
670 // Irreconcilable differences.
674 if (NumDifferences++) break;
679 if (NumDifferences == 0) // SAME GEP?
680 return ReplaceInstUsesWith(I, // No comparison is needed here.
681 ConstantInt::get(Type::getInt1Ty(I.getContext()),
682 ICmpInst::isTrueWhenEqual(Cond)));
684 else if (NumDifferences == 1 && GEPsInBounds) {
685 Value *LHSV = GEPLHS->getOperand(DiffOperand);
686 Value *RHSV = GEPRHS->getOperand(DiffOperand);
687 // Make sure we do a signed comparison here.
688 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
692 // Only lower this if the icmp is the only user of the GEP or if we expect
693 // the result to fold to a constant!
696 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
697 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
698 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
699 Value *L = EmitGEPOffset(GEPLHS);
700 Value *R = EmitGEPOffset(GEPRHS);
701 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
707 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
708 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
709 Value *X, ConstantInt *CI,
710 ICmpInst::Predicate Pred,
712 // If we have X+0, exit early (simplifying logic below) and let it get folded
713 // elsewhere. icmp X+0, X -> icmp X, X
715 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
716 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
719 // (X+4) == X -> false.
720 if (Pred == ICmpInst::ICMP_EQ)
721 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
723 // (X+4) != X -> true.
724 if (Pred == ICmpInst::ICMP_NE)
725 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
727 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
728 // so the values can never be equal. Similarly for all other "or equals"
731 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
732 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
733 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
734 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
736 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
737 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
740 // (X+1) >u X --> X <u (0-1) --> X != 255
741 // (X+2) >u X --> X <u (0-2) --> X <u 254
742 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
743 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
744 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
746 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
747 ConstantInt *SMax = ConstantInt::get(X->getContext(),
748 APInt::getSignedMaxValue(BitWidth));
750 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
751 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
752 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
753 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
754 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
755 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
756 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
757 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
759 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
760 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
761 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
762 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
763 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
764 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
766 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
767 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
768 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
771 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
772 /// and CmpRHS are both known to be integer constants.
773 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
774 ConstantInt *DivRHS) {
775 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
776 const APInt &CmpRHSV = CmpRHS->getValue();
778 // FIXME: If the operand types don't match the type of the divide
779 // then don't attempt this transform. The code below doesn't have the
780 // logic to deal with a signed divide and an unsigned compare (and
781 // vice versa). This is because (x /s C1) <s C2 produces different
782 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
783 // (x /u C1) <u C2. Simply casting the operands and result won't
784 // work. :( The if statement below tests that condition and bails
786 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
787 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
789 if (DivRHS->isZero())
790 return 0; // The ProdOV computation fails on divide by zero.
791 if (DivIsSigned && DivRHS->isAllOnesValue())
792 return 0; // The overflow computation also screws up here
793 if (DivRHS->isOne()) {
794 // This eliminates some funny cases with INT_MIN.
795 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
799 // Compute Prod = CI * DivRHS. We are essentially solving an equation
800 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
801 // C2 (CI). By solving for X we can turn this into a range check
802 // instead of computing a divide.
803 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
805 // Determine if the product overflows by seeing if the product is
806 // not equal to the divide. Make sure we do the same kind of divide
807 // as in the LHS instruction that we're folding.
808 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
809 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
811 // Get the ICmp opcode
812 ICmpInst::Predicate Pred = ICI.getPredicate();
814 /// If the division is known to be exact, then there is no remainder from the
815 /// divide, so the covered range size is unit, otherwise it is the divisor.
816 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
818 // Figure out the interval that is being checked. For example, a comparison
819 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
820 // Compute this interval based on the constants involved and the signedness of
821 // the compare/divide. This computes a half-open interval, keeping track of
822 // whether either value in the interval overflows. After analysis each
823 // overflow variable is set to 0 if it's corresponding bound variable is valid
824 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
825 int LoOverflow = 0, HiOverflow = 0;
826 Constant *LoBound = 0, *HiBound = 0;
828 if (!DivIsSigned) { // udiv
829 // e.g. X/5 op 3 --> [15, 20)
831 HiOverflow = LoOverflow = ProdOV;
833 // If this is not an exact divide, then many values in the range collapse
834 // to the same result value.
835 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
838 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
839 if (CmpRHSV == 0) { // (X / pos) op 0
840 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
841 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
843 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
844 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
845 HiOverflow = LoOverflow = ProdOV;
847 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
848 } else { // (X / pos) op neg
849 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
850 HiBound = AddOne(Prod);
851 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
853 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
854 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
857 } else if (DivRHS->isNegative()) { // Divisor is < 0.
859 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
860 if (CmpRHSV == 0) { // (X / neg) op 0
861 // e.g. X/-5 op 0 --> [-4, 5)
862 LoBound = AddOne(RangeSize);
863 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
864 if (HiBound == DivRHS) { // -INTMIN = INTMIN
865 HiOverflow = 1; // [INTMIN+1, overflow)
866 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
868 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
869 // e.g. X/-5 op 3 --> [-19, -14)
870 HiBound = AddOne(Prod);
871 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
873 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
874 } else { // (X / neg) op neg
875 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
876 LoOverflow = HiOverflow = ProdOV;
878 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
881 // Dividing by a negative swaps the condition. LT <-> GT
882 Pred = ICmpInst::getSwappedPredicate(Pred);
885 Value *X = DivI->getOperand(0);
887 default: llvm_unreachable("Unhandled icmp opcode!");
888 case ICmpInst::ICMP_EQ:
889 if (LoOverflow && HiOverflow)
890 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
892 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
893 ICmpInst::ICMP_UGE, X, LoBound);
895 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
896 ICmpInst::ICMP_ULT, X, HiBound);
897 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
899 case ICmpInst::ICMP_NE:
900 if (LoOverflow && HiOverflow)
901 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
903 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
904 ICmpInst::ICMP_ULT, X, LoBound);
906 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
907 ICmpInst::ICMP_UGE, X, HiBound);
908 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
909 DivIsSigned, false));
910 case ICmpInst::ICMP_ULT:
911 case ICmpInst::ICMP_SLT:
912 if (LoOverflow == +1) // Low bound is greater than input range.
913 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
914 if (LoOverflow == -1) // Low bound is less than input range.
915 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
916 return new ICmpInst(Pred, X, LoBound);
917 case ICmpInst::ICMP_UGT:
918 case ICmpInst::ICMP_SGT:
919 if (HiOverflow == +1) // High bound greater than input range.
920 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
921 if (HiOverflow == -1) // High bound less than input range.
922 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
923 if (Pred == ICmpInst::ICMP_UGT)
924 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
925 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
929 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
930 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
931 ConstantInt *ShAmt) {
932 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
934 // Check that the shift amount is in range. If not, don't perform
935 // undefined shifts. When the shift is visited it will be
937 uint32_t TypeBits = CmpRHSV.getBitWidth();
938 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
939 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
942 if (!ICI.isEquality()) {
943 // If we have an unsigned comparison and an ashr, we can't simplify this.
944 // Similarly for signed comparisons with lshr.
945 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
948 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
949 // by a power of 2. Since we already have logic to simplify these,
950 // transform to div and then simplify the resultant comparison.
951 if (Shr->getOpcode() == Instruction::AShr &&
952 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
955 // Revisit the shift (to delete it).
959 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
962 Shr->getOpcode() == Instruction::AShr ?
963 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
964 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
966 ICI.setOperand(0, Tmp);
968 // If the builder folded the binop, just return it.
969 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
973 // Otherwise, fold this div/compare.
974 assert(TheDiv->getOpcode() == Instruction::SDiv ||
975 TheDiv->getOpcode() == Instruction::UDiv);
977 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
978 assert(Res && "This div/cst should have folded!");
983 // If we are comparing against bits always shifted out, the
984 // comparison cannot succeed.
985 APInt Comp = CmpRHSV << ShAmtVal;
986 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
987 if (Shr->getOpcode() == Instruction::LShr)
988 Comp = Comp.lshr(ShAmtVal);
990 Comp = Comp.ashr(ShAmtVal);
992 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
993 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
994 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
996 return ReplaceInstUsesWith(ICI, Cst);
999 // Otherwise, check to see if the bits shifted out are known to be zero.
1000 // If so, we can compare against the unshifted value:
1001 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1002 if (Shr->hasOneUse() && Shr->isExact())
1003 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1005 if (Shr->hasOneUse()) {
1006 // Otherwise strength reduce the shift into an and.
1007 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1008 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
1010 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1011 Mask, Shr->getName()+".mask");
1012 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1018 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1020 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1023 const APInt &RHSV = RHS->getValue();
1025 switch (LHSI->getOpcode()) {
1026 case Instruction::Trunc:
1027 if (ICI.isEquality() && LHSI->hasOneUse()) {
1028 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1029 // of the high bits truncated out of x are known.
1030 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1031 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1032 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1033 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1035 // If all the high bits are known, we can do this xform.
1036 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1037 // Pull in the high bits from known-ones set.
1038 APInt NewRHS = RHS->getValue().zext(SrcBits);
1039 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1040 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1041 ConstantInt::get(ICI.getContext(), NewRHS));
1046 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1047 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1048 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1050 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1051 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1052 Value *CompareVal = LHSI->getOperand(0);
1054 // If the sign bit of the XorCST is not set, there is no change to
1055 // the operation, just stop using the Xor.
1056 if (!XorCST->isNegative()) {
1057 ICI.setOperand(0, CompareVal);
1062 // Was the old condition true if the operand is positive?
1063 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1065 // If so, the new one isn't.
1066 isTrueIfPositive ^= true;
1068 if (isTrueIfPositive)
1069 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1072 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1076 if (LHSI->hasOneUse()) {
1077 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1078 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1079 const APInt &SignBit = XorCST->getValue();
1080 ICmpInst::Predicate Pred = ICI.isSigned()
1081 ? ICI.getUnsignedPredicate()
1082 : ICI.getSignedPredicate();
1083 return new ICmpInst(Pred, LHSI->getOperand(0),
1084 ConstantInt::get(ICI.getContext(),
1088 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1089 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1090 const APInt &NotSignBit = XorCST->getValue();
1091 ICmpInst::Predicate Pred = ICI.isSigned()
1092 ? ICI.getUnsignedPredicate()
1093 : ICI.getSignedPredicate();
1094 Pred = ICI.getSwappedPredicate(Pred);
1095 return new ICmpInst(Pred, LHSI->getOperand(0),
1096 ConstantInt::get(ICI.getContext(),
1097 RHSV ^ NotSignBit));
1102 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1103 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1104 LHSI->getOperand(0)->hasOneUse()) {
1105 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1107 // If the LHS is an AND of a truncating cast, we can widen the
1108 // and/compare to be the input width without changing the value
1109 // produced, eliminating a cast.
1110 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1111 // We can do this transformation if either the AND constant does not
1112 // have its sign bit set or if it is an equality comparison.
1113 // Extending a relational comparison when we're checking the sign
1114 // bit would not work.
1115 if (ICI.isEquality() ||
1116 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1118 Builder->CreateAnd(Cast->getOperand(0),
1119 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1120 NewAnd->takeName(LHSI);
1121 return new ICmpInst(ICI.getPredicate(), NewAnd,
1122 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1126 // If the LHS is an AND of a zext, and we have an equality compare, we can
1127 // shrink the and/compare to the smaller type, eliminating the cast.
1128 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1129 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1130 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1131 // should fold the icmp to true/false in that case.
1132 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1134 Builder->CreateAnd(Cast->getOperand(0),
1135 ConstantExpr::getTrunc(AndCST, Ty));
1136 NewAnd->takeName(LHSI);
1137 return new ICmpInst(ICI.getPredicate(), NewAnd,
1138 ConstantExpr::getTrunc(RHS, Ty));
1142 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1143 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1144 // happens a LOT in code produced by the C front-end, for bitfield
1146 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1147 if (Shift && !Shift->isShift())
1151 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1152 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1153 Type *AndTy = AndCST->getType(); // Type of the and.
1155 // We can fold this as long as we can't shift unknown bits
1156 // into the mask. This can only happen with signed shift
1157 // rights, as they sign-extend.
1159 bool CanFold = Shift->isLogicalShift();
1161 // To test for the bad case of the signed shr, see if any
1162 // of the bits shifted in could be tested after the mask.
1163 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1164 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1166 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1167 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1168 AndCST->getValue()) == 0)
1174 if (Shift->getOpcode() == Instruction::Shl)
1175 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1177 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1179 // Check to see if we are shifting out any of the bits being
1181 if (ConstantExpr::get(Shift->getOpcode(),
1182 NewCst, ShAmt) != RHS) {
1183 // If we shifted bits out, the fold is not going to work out.
1184 // As a special case, check to see if this means that the
1185 // result is always true or false now.
1186 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1187 return ReplaceInstUsesWith(ICI,
1188 ConstantInt::getFalse(ICI.getContext()));
1189 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1190 return ReplaceInstUsesWith(ICI,
1191 ConstantInt::getTrue(ICI.getContext()));
1193 ICI.setOperand(1, NewCst);
1194 Constant *NewAndCST;
1195 if (Shift->getOpcode() == Instruction::Shl)
1196 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1198 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1199 LHSI->setOperand(1, NewAndCST);
1200 LHSI->setOperand(0, Shift->getOperand(0));
1201 Worklist.Add(Shift); // Shift is dead.
1207 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1208 // preferable because it allows the C<<Y expression to be hoisted out
1209 // of a loop if Y is invariant and X is not.
1210 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1211 ICI.isEquality() && !Shift->isArithmeticShift() &&
1212 !isa<Constant>(Shift->getOperand(0))) {
1215 if (Shift->getOpcode() == Instruction::LShr) {
1216 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1218 // Insert a logical shift.
1219 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1222 // Compute X & (C << Y).
1224 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1226 ICI.setOperand(0, NewAnd);
1231 // Try to optimize things like "A[i]&42 == 0" to index computations.
1232 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1233 if (GetElementPtrInst *GEP =
1234 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1235 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1236 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1237 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1238 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1239 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1245 case Instruction::Or: {
1246 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1249 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1250 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1251 // -> and (icmp eq P, null), (icmp eq Q, null).
1252 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1253 Constant::getNullValue(P->getType()));
1254 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1255 Constant::getNullValue(Q->getType()));
1257 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1258 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1260 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1266 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1267 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1270 uint32_t TypeBits = RHSV.getBitWidth();
1272 // Check that the shift amount is in range. If not, don't perform
1273 // undefined shifts. When the shift is visited it will be
1275 if (ShAmt->uge(TypeBits))
1278 if (ICI.isEquality()) {
1279 // If we are comparing against bits always shifted out, the
1280 // comparison cannot succeed.
1282 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1284 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1285 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1287 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1288 return ReplaceInstUsesWith(ICI, Cst);
1291 // If the shift is NUW, then it is just shifting out zeros, no need for an
1293 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1294 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1295 ConstantExpr::getLShr(RHS, ShAmt));
1297 if (LHSI->hasOneUse()) {
1298 // Otherwise strength reduce the shift into an and.
1299 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1301 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1302 TypeBits-ShAmtVal));
1305 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1306 return new ICmpInst(ICI.getPredicate(), And,
1307 ConstantExpr::getLShr(RHS, ShAmt));
1311 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1312 bool TrueIfSigned = false;
1313 if (LHSI->hasOneUse() &&
1314 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1315 // (X << 31) <s 0 --> (X&1) != 0
1316 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1317 APInt::getOneBitSet(TypeBits,
1318 TypeBits-ShAmt->getZExtValue()-1));
1320 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1321 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1322 And, Constant::getNullValue(And->getType()));
1327 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1328 case Instruction::AShr: {
1329 // Handle equality comparisons of shift-by-constant.
1330 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1331 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1332 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1336 // Handle exact shr's.
1337 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1338 if (RHSV.isMinValue())
1339 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1344 case Instruction::SDiv:
1345 case Instruction::UDiv:
1346 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1347 // Fold this div into the comparison, producing a range check.
1348 // Determine, based on the divide type, what the range is being
1349 // checked. If there is an overflow on the low or high side, remember
1350 // it, otherwise compute the range [low, hi) bounding the new value.
1351 // See: InsertRangeTest above for the kinds of replacements possible.
1352 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1353 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1358 case Instruction::Add:
1359 // Fold: icmp pred (add X, C1), C2
1360 if (!ICI.isEquality()) {
1361 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1363 const APInt &LHSV = LHSC->getValue();
1365 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1368 if (ICI.isSigned()) {
1369 if (CR.getLower().isSignBit()) {
1370 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1371 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1372 } else if (CR.getUpper().isSignBit()) {
1373 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1374 ConstantInt::get(ICI.getContext(),CR.getLower()));
1377 if (CR.getLower().isMinValue()) {
1378 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1379 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1380 } else if (CR.getUpper().isMinValue()) {
1381 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1382 ConstantInt::get(ICI.getContext(),CR.getLower()));
1389 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1390 if (ICI.isEquality()) {
1391 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1393 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1394 // the second operand is a constant, simplify a bit.
1395 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1396 switch (BO->getOpcode()) {
1397 case Instruction::SRem:
1398 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1399 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1400 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1401 if (V.sgt(1) && V.isPowerOf2()) {
1403 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1405 return new ICmpInst(ICI.getPredicate(), NewRem,
1406 Constant::getNullValue(BO->getType()));
1410 case Instruction::Add:
1411 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1412 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1413 if (BO->hasOneUse())
1414 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1415 ConstantExpr::getSub(RHS, BOp1C));
1416 } else if (RHSV == 0) {
1417 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1418 // efficiently invertible, or if the add has just this one use.
1419 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1421 if (Value *NegVal = dyn_castNegVal(BOp1))
1422 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1423 if (Value *NegVal = dyn_castNegVal(BOp0))
1424 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1425 if (BO->hasOneUse()) {
1426 Value *Neg = Builder->CreateNeg(BOp1);
1428 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1432 case Instruction::Xor:
1433 // For the xor case, we can xor two constants together, eliminating
1434 // the explicit xor.
1435 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1436 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1437 ConstantExpr::getXor(RHS, BOC));
1438 } else if (RHSV == 0) {
1439 // Replace ((xor A, B) != 0) with (A != B)
1440 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1444 case Instruction::Sub:
1445 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1446 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1447 if (BO->hasOneUse())
1448 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1449 ConstantExpr::getSub(BOp0C, RHS));
1450 } else if (RHSV == 0) {
1451 // Replace ((sub A, B) != 0) with (A != B)
1452 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1456 case Instruction::Or:
1457 // If bits are being or'd in that are not present in the constant we
1458 // are comparing against, then the comparison could never succeed!
1459 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1460 Constant *NotCI = ConstantExpr::getNot(RHS);
1461 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1462 return ReplaceInstUsesWith(ICI,
1463 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1468 case Instruction::And:
1469 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1470 // If bits are being compared against that are and'd out, then the
1471 // comparison can never succeed!
1472 if ((RHSV & ~BOC->getValue()) != 0)
1473 return ReplaceInstUsesWith(ICI,
1474 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1477 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1478 if (RHS == BOC && RHSV.isPowerOf2())
1479 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1480 ICmpInst::ICMP_NE, LHSI,
1481 Constant::getNullValue(RHS->getType()));
1483 // Don't perform the following transforms if the AND has multiple uses
1484 if (!BO->hasOneUse())
1487 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1488 if (BOC->getValue().isSignBit()) {
1489 Value *X = BO->getOperand(0);
1490 Constant *Zero = Constant::getNullValue(X->getType());
1491 ICmpInst::Predicate pred = isICMP_NE ?
1492 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1493 return new ICmpInst(pred, X, Zero);
1496 // ((X & ~7) == 0) --> X < 8
1497 if (RHSV == 0 && isHighOnes(BOC)) {
1498 Value *X = BO->getOperand(0);
1499 Constant *NegX = ConstantExpr::getNeg(BOC);
1500 ICmpInst::Predicate pred = isICMP_NE ?
1501 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1502 return new ICmpInst(pred, X, NegX);
1507 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1508 // Handle icmp {eq|ne} <intrinsic>, intcst.
1509 switch (II->getIntrinsicID()) {
1510 case Intrinsic::bswap:
1512 ICI.setOperand(0, II->getArgOperand(0));
1513 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1515 case Intrinsic::ctlz:
1516 case Intrinsic::cttz:
1517 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1518 if (RHSV == RHS->getType()->getBitWidth()) {
1520 ICI.setOperand(0, II->getArgOperand(0));
1521 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1525 case Intrinsic::ctpop:
1526 // popcount(A) == 0 -> A == 0 and likewise for !=
1527 if (RHS->isZero()) {
1529 ICI.setOperand(0, II->getArgOperand(0));
1530 ICI.setOperand(1, RHS);
1542 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1543 /// We only handle extending casts so far.
1545 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1546 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1547 Value *LHSCIOp = LHSCI->getOperand(0);
1548 Type *SrcTy = LHSCIOp->getType();
1549 Type *DestTy = LHSCI->getType();
1552 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1553 // integer type is the same size as the pointer type.
1554 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1555 TD->getPointerSizeInBits() ==
1556 cast<IntegerType>(DestTy)->getBitWidth()) {
1558 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1559 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1560 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1561 RHSOp = RHSC->getOperand(0);
1562 // If the pointer types don't match, insert a bitcast.
1563 if (LHSCIOp->getType() != RHSOp->getType())
1564 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1568 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1571 // The code below only handles extension cast instructions, so far.
1573 if (LHSCI->getOpcode() != Instruction::ZExt &&
1574 LHSCI->getOpcode() != Instruction::SExt)
1577 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1578 bool isSignedCmp = ICI.isSigned();
1580 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1581 // Not an extension from the same type?
1582 RHSCIOp = CI->getOperand(0);
1583 if (RHSCIOp->getType() != LHSCIOp->getType())
1586 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1587 // and the other is a zext), then we can't handle this.
1588 if (CI->getOpcode() != LHSCI->getOpcode())
1591 // Deal with equality cases early.
1592 if (ICI.isEquality())
1593 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1595 // A signed comparison of sign extended values simplifies into a
1596 // signed comparison.
1597 if (isSignedCmp && isSignedExt)
1598 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1600 // The other three cases all fold into an unsigned comparison.
1601 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1604 // If we aren't dealing with a constant on the RHS, exit early
1605 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1609 // Compute the constant that would happen if we truncated to SrcTy then
1610 // reextended to DestTy.
1611 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1612 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1615 // If the re-extended constant didn't change...
1617 // Deal with equality cases early.
1618 if (ICI.isEquality())
1619 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1621 // A signed comparison of sign extended values simplifies into a
1622 // signed comparison.
1623 if (isSignedExt && isSignedCmp)
1624 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1626 // The other three cases all fold into an unsigned comparison.
1627 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1630 // The re-extended constant changed so the constant cannot be represented
1631 // in the shorter type. Consequently, we cannot emit a simple comparison.
1632 // All the cases that fold to true or false will have already been handled
1633 // by SimplifyICmpInst, so only deal with the tricky case.
1635 if (isSignedCmp || !isSignedExt)
1638 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1639 // should have been folded away previously and not enter in here.
1641 // We're performing an unsigned comp with a sign extended value.
1642 // This is true if the input is >= 0. [aka >s -1]
1643 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1644 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1646 // Finally, return the value computed.
1647 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1648 return ReplaceInstUsesWith(ICI, Result);
1650 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1651 return BinaryOperator::CreateNot(Result);
1654 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1655 /// I = icmp ugt (add (add A, B), CI2), CI1
1656 /// If this is of the form:
1658 /// if (sum+128 >u 255)
1659 /// Then replace it with llvm.sadd.with.overflow.i8.
1661 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1662 ConstantInt *CI2, ConstantInt *CI1,
1664 // The transformation we're trying to do here is to transform this into an
1665 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1666 // with a narrower add, and discard the add-with-constant that is part of the
1667 // range check (if we can't eliminate it, this isn't profitable).
1669 // In order to eliminate the add-with-constant, the compare can be its only
1671 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1672 if (!AddWithCst->hasOneUse()) return 0;
1674 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1675 if (!CI2->getValue().isPowerOf2()) return 0;
1676 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1677 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1679 // The width of the new add formed is 1 more than the bias.
1682 // Check to see that CI1 is an all-ones value with NewWidth bits.
1683 if (CI1->getBitWidth() == NewWidth ||
1684 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1687 // This is only really a signed overflow check if the inputs have been
1688 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1689 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1690 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1691 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1692 IC.ComputeNumSignBits(B) < NeededSignBits)
1695 // In order to replace the original add with a narrower
1696 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1697 // and truncates that discard the high bits of the add. Verify that this is
1699 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1700 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1702 if (*UI == AddWithCst) continue;
1704 // Only accept truncates for now. We would really like a nice recursive
1705 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1706 // chain to see which bits of a value are actually demanded. If the
1707 // original add had another add which was then immediately truncated, we
1708 // could still do the transformation.
1709 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1711 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1714 // If the pattern matches, truncate the inputs to the narrower type and
1715 // use the sadd_with_overflow intrinsic to efficiently compute both the
1716 // result and the overflow bit.
1717 Module *M = I.getParent()->getParent()->getParent();
1719 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1720 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1723 InstCombiner::BuilderTy *Builder = IC.Builder;
1725 // Put the new code above the original add, in case there are any uses of the
1726 // add between the add and the compare.
1727 Builder->SetInsertPoint(OrigAdd);
1729 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1730 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1731 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1732 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1733 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1735 // The inner add was the result of the narrow add, zero extended to the
1736 // wider type. Replace it with the result computed by the intrinsic.
1737 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1739 // The original icmp gets replaced with the overflow value.
1740 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1743 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1745 // Don't bother doing this transformation for pointers, don't do it for
1747 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1749 // If the add is a constant expr, then we don't bother transforming it.
1750 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1751 if (OrigAdd == 0) return 0;
1753 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1755 // Put the new code above the original add, in case there are any uses of the
1756 // add between the add and the compare.
1757 InstCombiner::BuilderTy *Builder = IC.Builder;
1758 Builder->SetInsertPoint(OrigAdd);
1760 Module *M = I.getParent()->getParent()->getParent();
1761 Type *Ty = LHS->getType();
1762 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1763 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1764 Value *Add = Builder->CreateExtractValue(Call, 0);
1766 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1768 // The original icmp gets replaced with the overflow value.
1769 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1772 // DemandedBitsLHSMask - When performing a comparison against a constant,
1773 // it is possible that not all the bits in the LHS are demanded. This helper
1774 // method computes the mask that IS demanded.
1775 static APInt DemandedBitsLHSMask(ICmpInst &I,
1776 unsigned BitWidth, bool isSignCheck) {
1778 return APInt::getSignBit(BitWidth);
1780 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1781 if (!CI) return APInt::getAllOnesValue(BitWidth);
1782 const APInt &RHS = CI->getValue();
1784 switch (I.getPredicate()) {
1785 // For a UGT comparison, we don't care about any bits that
1786 // correspond to the trailing ones of the comparand. The value of these
1787 // bits doesn't impact the outcome of the comparison, because any value
1788 // greater than the RHS must differ in a bit higher than these due to carry.
1789 case ICmpInst::ICMP_UGT: {
1790 unsigned trailingOnes = RHS.countTrailingOnes();
1791 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1795 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1796 // Any value less than the RHS must differ in a higher bit because of carries.
1797 case ICmpInst::ICMP_ULT: {
1798 unsigned trailingZeros = RHS.countTrailingZeros();
1799 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1804 return APInt::getAllOnesValue(BitWidth);
1809 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1810 bool Changed = false;
1811 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1813 /// Orders the operands of the compare so that they are listed from most
1814 /// complex to least complex. This puts constants before unary operators,
1815 /// before binary operators.
1816 if (getComplexity(Op0) < getComplexity(Op1)) {
1818 std::swap(Op0, Op1);
1822 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1823 return ReplaceInstUsesWith(I, V);
1825 // comparing -val or val with non-zero is the same as just comparing val
1826 // ie, abs(val) != 0 -> val != 0
1827 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
1829 Value *Cond, *SelectTrue, *SelectFalse;
1830 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
1831 m_Value(SelectFalse)))) {
1832 if (Value *V = dyn_castNegVal(SelectTrue)) {
1833 if (V == SelectFalse)
1834 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1836 else if (Value *V = dyn_castNegVal(SelectFalse)) {
1837 if (V == SelectTrue)
1838 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1843 Type *Ty = Op0->getType();
1845 // icmp's with boolean values can always be turned into bitwise operations
1846 if (Ty->isIntegerTy(1)) {
1847 switch (I.getPredicate()) {
1848 default: llvm_unreachable("Invalid icmp instruction!");
1849 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1850 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1851 return BinaryOperator::CreateNot(Xor);
1853 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1854 return BinaryOperator::CreateXor(Op0, Op1);
1856 case ICmpInst::ICMP_UGT:
1857 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1859 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1860 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1861 return BinaryOperator::CreateAnd(Not, Op1);
1863 case ICmpInst::ICMP_SGT:
1864 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1866 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1867 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1868 return BinaryOperator::CreateAnd(Not, Op0);
1870 case ICmpInst::ICMP_UGE:
1871 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1873 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1874 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1875 return BinaryOperator::CreateOr(Not, Op1);
1877 case ICmpInst::ICMP_SGE:
1878 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1880 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1881 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1882 return BinaryOperator::CreateOr(Not, Op0);
1887 unsigned BitWidth = 0;
1888 if (Ty->isIntOrIntVectorTy())
1889 BitWidth = Ty->getScalarSizeInBits();
1890 else if (TD) // Pointers require TD info to get their size.
1891 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1893 bool isSignBit = false;
1895 // See if we are doing a comparison with a constant.
1896 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1897 Value *A = 0, *B = 0;
1899 // Match the following pattern, which is a common idiom when writing
1900 // overflow-safe integer arithmetic function. The source performs an
1901 // addition in wider type, and explicitly checks for overflow using
1902 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1903 // sadd_with_overflow intrinsic.
1905 // TODO: This could probably be generalized to handle other overflow-safe
1906 // operations if we worked out the formulas to compute the appropriate
1910 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1912 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1913 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1914 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1915 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1919 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1920 if (I.isEquality() && CI->isZero() &&
1921 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1922 // (icmp cond A B) if cond is equality
1923 return new ICmpInst(I.getPredicate(), A, B);
1926 // If we have an icmp le or icmp ge instruction, turn it into the
1927 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1928 // them being folded in the code below. The SimplifyICmpInst code has
1929 // already handled the edge cases for us, so we just assert on them.
1930 switch (I.getPredicate()) {
1932 case ICmpInst::ICMP_ULE:
1933 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1934 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1935 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1936 case ICmpInst::ICMP_SLE:
1937 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1938 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1939 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1940 case ICmpInst::ICMP_UGE:
1941 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1942 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1943 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1944 case ICmpInst::ICMP_SGE:
1945 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1946 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1947 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1950 // If this comparison is a normal comparison, it demands all
1951 // bits, if it is a sign bit comparison, it only demands the sign bit.
1953 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1956 // See if we can fold the comparison based on range information we can get
1957 // by checking whether bits are known to be zero or one in the input.
1958 if (BitWidth != 0) {
1959 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1960 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1962 if (SimplifyDemandedBits(I.getOperandUse(0),
1963 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1964 Op0KnownZero, Op0KnownOne, 0))
1966 if (SimplifyDemandedBits(I.getOperandUse(1),
1967 APInt::getAllOnesValue(BitWidth),
1968 Op1KnownZero, Op1KnownOne, 0))
1971 // Given the known and unknown bits, compute a range that the LHS could be
1972 // in. Compute the Min, Max and RHS values based on the known bits. For the
1973 // EQ and NE we use unsigned values.
1974 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1975 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1977 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1979 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1982 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1984 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1988 // If Min and Max are known to be the same, then SimplifyDemandedBits
1989 // figured out that the LHS is a constant. Just constant fold this now so
1990 // that code below can assume that Min != Max.
1991 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1992 return new ICmpInst(I.getPredicate(),
1993 ConstantInt::get(Op0->getType(), Op0Min), Op1);
1994 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1995 return new ICmpInst(I.getPredicate(), Op0,
1996 ConstantInt::get(Op1->getType(), Op1Min));
1998 // Based on the range information we know about the LHS, see if we can
1999 // simplify this comparison. For example, (x&4) < 8 is always true.
2000 switch (I.getPredicate()) {
2001 default: llvm_unreachable("Unknown icmp opcode!");
2002 case ICmpInst::ICMP_EQ: {
2003 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2004 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2006 // If all bits are known zero except for one, then we know at most one
2007 // bit is set. If the comparison is against zero, then this is a check
2008 // to see if *that* bit is set.
2009 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2010 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2011 // If the LHS is an AND with the same constant, look through it.
2013 ConstantInt *LHSC = 0;
2014 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2015 LHSC->getValue() != Op0KnownZeroInverted)
2018 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2019 // then turn "((1 << x)&8) == 0" into "x != 3".
2021 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2022 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2023 return new ICmpInst(ICmpInst::ICMP_NE, X,
2024 ConstantInt::get(X->getType(), CmpVal));
2027 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2028 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2030 if (Op0KnownZeroInverted == 1 &&
2031 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2032 return new ICmpInst(ICmpInst::ICMP_NE, X,
2033 ConstantInt::get(X->getType(),
2034 CI->countTrailingZeros()));
2039 case ICmpInst::ICMP_NE: {
2040 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2041 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2043 // If all bits are known zero except for one, then we know at most one
2044 // bit is set. If the comparison is against zero, then this is a check
2045 // to see if *that* bit is set.
2046 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2047 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2048 // If the LHS is an AND with the same constant, look through it.
2050 ConstantInt *LHSC = 0;
2051 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2052 LHSC->getValue() != Op0KnownZeroInverted)
2055 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2056 // then turn "((1 << x)&8) != 0" into "x == 3".
2058 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2059 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2060 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2061 ConstantInt::get(X->getType(), CmpVal));
2064 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2065 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2067 if (Op0KnownZeroInverted == 1 &&
2068 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2069 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2070 ConstantInt::get(X->getType(),
2071 CI->countTrailingZeros()));
2076 case ICmpInst::ICMP_ULT:
2077 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2078 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2079 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2080 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2081 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2082 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2083 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2084 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2085 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2086 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2088 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2089 if (CI->isMinValue(true))
2090 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2091 Constant::getAllOnesValue(Op0->getType()));
2094 case ICmpInst::ICMP_UGT:
2095 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2096 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2097 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2098 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2100 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2101 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2102 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2103 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2104 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2105 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2107 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2108 if (CI->isMaxValue(true))
2109 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2110 Constant::getNullValue(Op0->getType()));
2113 case ICmpInst::ICMP_SLT:
2114 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2115 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2116 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2117 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2118 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2119 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2120 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2121 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2122 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2123 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2126 case ICmpInst::ICMP_SGT:
2127 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2128 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2129 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2130 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2132 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2133 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2134 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2135 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2136 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2137 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2140 case ICmpInst::ICMP_SGE:
2141 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2142 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2143 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2144 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2145 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2147 case ICmpInst::ICMP_SLE:
2148 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2149 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2150 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2151 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2152 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2154 case ICmpInst::ICMP_UGE:
2155 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2156 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2157 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2158 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2159 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2161 case ICmpInst::ICMP_ULE:
2162 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2163 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2164 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2165 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2166 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2170 // Turn a signed comparison into an unsigned one if both operands
2171 // are known to have the same sign.
2173 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2174 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2175 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2178 // Test if the ICmpInst instruction is used exclusively by a select as
2179 // part of a minimum or maximum operation. If so, refrain from doing
2180 // any other folding. This helps out other analyses which understand
2181 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2182 // and CodeGen. And in this case, at least one of the comparison
2183 // operands has at least one user besides the compare (the select),
2184 // which would often largely negate the benefit of folding anyway.
2186 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2187 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2188 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2191 // See if we are doing a comparison between a constant and an instruction that
2192 // can be folded into the comparison.
2193 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2194 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2195 // instruction, see if that instruction also has constants so that the
2196 // instruction can be folded into the icmp
2197 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2198 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2202 // Handle icmp with constant (but not simple integer constant) RHS
2203 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2204 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2205 switch (LHSI->getOpcode()) {
2206 case Instruction::GetElementPtr:
2207 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2208 if (RHSC->isNullValue() &&
2209 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2210 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2211 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2213 case Instruction::PHI:
2214 // Only fold icmp into the PHI if the phi and icmp are in the same
2215 // block. If in the same block, we're encouraging jump threading. If
2216 // not, we are just pessimizing the code by making an i1 phi.
2217 if (LHSI->getParent() == I.getParent())
2218 if (Instruction *NV = FoldOpIntoPhi(I))
2221 case Instruction::Select: {
2222 // If either operand of the select is a constant, we can fold the
2223 // comparison into the select arms, which will cause one to be
2224 // constant folded and the select turned into a bitwise or.
2225 Value *Op1 = 0, *Op2 = 0;
2226 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2227 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2228 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2229 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2231 // We only want to perform this transformation if it will not lead to
2232 // additional code. This is true if either both sides of the select
2233 // fold to a constant (in which case the icmp is replaced with a select
2234 // which will usually simplify) or this is the only user of the
2235 // select (in which case we are trading a select+icmp for a simpler
2237 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2239 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2242 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2244 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2248 case Instruction::IntToPtr:
2249 // icmp pred inttoptr(X), null -> icmp pred X, 0
2250 if (RHSC->isNullValue() && TD &&
2251 TD->getIntPtrType(RHSC->getContext()) ==
2252 LHSI->getOperand(0)->getType())
2253 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2254 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2257 case Instruction::Load:
2258 // Try to optimize things like "A[i] > 4" to index computations.
2259 if (GetElementPtrInst *GEP =
2260 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2261 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2262 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2263 !cast<LoadInst>(LHSI)->isVolatile())
2264 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2271 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2272 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2273 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2275 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2276 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2277 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2280 // Test to see if the operands of the icmp are casted versions of other
2281 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2283 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2284 if (Op0->getType()->isPointerTy() &&
2285 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2286 // We keep moving the cast from the left operand over to the right
2287 // operand, where it can often be eliminated completely.
2288 Op0 = CI->getOperand(0);
2290 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2291 // so eliminate it as well.
2292 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2293 Op1 = CI2->getOperand(0);
2295 // If Op1 is a constant, we can fold the cast into the constant.
2296 if (Op0->getType() != Op1->getType()) {
2297 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2298 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2300 // Otherwise, cast the RHS right before the icmp
2301 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2304 return new ICmpInst(I.getPredicate(), Op0, Op1);
2308 if (isa<CastInst>(Op0)) {
2309 // Handle the special case of: icmp (cast bool to X), <cst>
2310 // This comes up when you have code like
2313 // For generality, we handle any zero-extension of any operand comparison
2314 // with a constant or another cast from the same type.
2315 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2316 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2320 // Special logic for binary operators.
2321 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2322 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2324 CmpInst::Predicate Pred = I.getPredicate();
2325 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2326 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2327 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2328 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2329 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2330 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2331 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2332 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2333 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2335 // Analyze the case when either Op0 or Op1 is an add instruction.
2336 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2337 Value *A = 0, *B = 0, *C = 0, *D = 0;
2338 if (BO0 && BO0->getOpcode() == Instruction::Add)
2339 A = BO0->getOperand(0), B = BO0->getOperand(1);
2340 if (BO1 && BO1->getOpcode() == Instruction::Add)
2341 C = BO1->getOperand(0), D = BO1->getOperand(1);
2343 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2344 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2345 return new ICmpInst(Pred, A == Op1 ? B : A,
2346 Constant::getNullValue(Op1->getType()));
2348 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2349 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2350 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2353 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2354 if (A && C && (A == C || A == D || B == C || B == D) &&
2355 NoOp0WrapProblem && NoOp1WrapProblem &&
2356 // Try not to increase register pressure.
2357 BO0->hasOneUse() && BO1->hasOneUse()) {
2358 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2361 // C + B == C + D -> B == D
2364 } else if (A == D) {
2365 // D + B == C + D -> B == C
2368 } else if (B == C) {
2369 // A + C == C + D -> A == D
2374 // A + D == C + D -> A == C
2378 return new ICmpInst(Pred, Y, Z);
2381 // Analyze the case when either Op0 or Op1 is a sub instruction.
2382 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2383 A = 0; B = 0; C = 0; D = 0;
2384 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2385 A = BO0->getOperand(0), B = BO0->getOperand(1);
2386 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2387 C = BO1->getOperand(0), D = BO1->getOperand(1);
2389 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2390 if (A == Op1 && NoOp0WrapProblem)
2391 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2393 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2394 if (C == Op0 && NoOp1WrapProblem)
2395 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2397 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2398 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2399 // Try not to increase register pressure.
2400 BO0->hasOneUse() && BO1->hasOneUse())
2401 return new ICmpInst(Pred, A, C);
2403 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2404 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2405 // Try not to increase register pressure.
2406 BO0->hasOneUse() && BO1->hasOneUse())
2407 return new ICmpInst(Pred, D, B);
2409 BinaryOperator *SRem = NULL;
2410 // icmp (srem X, Y), Y
2411 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2412 Op1 == BO0->getOperand(1))
2414 // icmp Y, (srem X, Y)
2415 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2416 Op0 == BO1->getOperand(1))
2419 // We don't check hasOneUse to avoid increasing register pressure because
2420 // the value we use is the same value this instruction was already using.
2421 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2423 case ICmpInst::ICMP_EQ:
2424 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2425 case ICmpInst::ICMP_NE:
2426 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2427 case ICmpInst::ICMP_SGT:
2428 case ICmpInst::ICMP_SGE:
2429 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2430 Constant::getAllOnesValue(SRem->getType()));
2431 case ICmpInst::ICMP_SLT:
2432 case ICmpInst::ICMP_SLE:
2433 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2434 Constant::getNullValue(SRem->getType()));
2438 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2439 BO0->hasOneUse() && BO1->hasOneUse() &&
2440 BO0->getOperand(1) == BO1->getOperand(1)) {
2441 switch (BO0->getOpcode()) {
2443 case Instruction::Add:
2444 case Instruction::Sub:
2445 case Instruction::Xor:
2446 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2447 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2448 BO1->getOperand(0));
2449 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2450 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2451 if (CI->getValue().isSignBit()) {
2452 ICmpInst::Predicate Pred = I.isSigned()
2453 ? I.getUnsignedPredicate()
2454 : I.getSignedPredicate();
2455 return new ICmpInst(Pred, BO0->getOperand(0),
2456 BO1->getOperand(0));
2459 if (CI->isMaxValue(true)) {
2460 ICmpInst::Predicate Pred = I.isSigned()
2461 ? I.getUnsignedPredicate()
2462 : I.getSignedPredicate();
2463 Pred = I.getSwappedPredicate(Pred);
2464 return new ICmpInst(Pred, BO0->getOperand(0),
2465 BO1->getOperand(0));
2469 case Instruction::Mul:
2470 if (!I.isEquality())
2473 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2474 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2475 // Mask = -1 >> count-trailing-zeros(Cst).
2476 if (!CI->isZero() && !CI->isOne()) {
2477 const APInt &AP = CI->getValue();
2478 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2479 APInt::getLowBitsSet(AP.getBitWidth(),
2481 AP.countTrailingZeros()));
2482 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2483 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2484 return new ICmpInst(I.getPredicate(), And1, And2);
2488 case Instruction::UDiv:
2489 case Instruction::LShr:
2493 case Instruction::SDiv:
2494 case Instruction::AShr:
2495 if (!BO0->isExact() || !BO1->isExact())
2497 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2498 BO1->getOperand(0));
2499 case Instruction::Shl: {
2500 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2501 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2504 if (!NSW && I.isSigned())
2506 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2507 BO1->getOperand(0));
2514 // ~x < ~y --> y < x
2515 // ~x < cst --> ~cst < x
2516 if (match(Op0, m_Not(m_Value(A)))) {
2517 if (match(Op1, m_Not(m_Value(B))))
2518 return new ICmpInst(I.getPredicate(), B, A);
2519 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2520 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2523 // (a+b) <u a --> llvm.uadd.with.overflow.
2524 // (a+b) <u b --> llvm.uadd.with.overflow.
2525 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2526 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2527 (Op1 == A || Op1 == B))
2528 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2531 // a >u (a+b) --> llvm.uadd.with.overflow.
2532 // b >u (a+b) --> llvm.uadd.with.overflow.
2533 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2534 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2535 (Op0 == A || Op0 == B))
2536 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2540 if (I.isEquality()) {
2541 Value *A, *B, *C, *D;
2543 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2544 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2545 Value *OtherVal = A == Op1 ? B : A;
2546 return new ICmpInst(I.getPredicate(), OtherVal,
2547 Constant::getNullValue(A->getType()));
2550 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2551 // A^c1 == C^c2 --> A == C^(c1^c2)
2552 ConstantInt *C1, *C2;
2553 if (match(B, m_ConstantInt(C1)) &&
2554 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2555 Constant *NC = ConstantInt::get(I.getContext(),
2556 C1->getValue() ^ C2->getValue());
2557 Value *Xor = Builder->CreateXor(C, NC);
2558 return new ICmpInst(I.getPredicate(), A, Xor);
2561 // A^B == A^D -> B == D
2562 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2563 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2564 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2565 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2569 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2570 (A == Op0 || B == Op0)) {
2571 // A == (A^B) -> B == 0
2572 Value *OtherVal = A == Op0 ? B : A;
2573 return new ICmpInst(I.getPredicate(), OtherVal,
2574 Constant::getNullValue(A->getType()));
2577 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2578 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2579 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2580 Value *X = 0, *Y = 0, *Z = 0;
2583 X = B; Y = D; Z = A;
2584 } else if (A == D) {
2585 X = B; Y = C; Z = A;
2586 } else if (B == C) {
2587 X = A; Y = D; Z = B;
2588 } else if (B == D) {
2589 X = A; Y = C; Z = B;
2592 if (X) { // Build (X^Y) & Z
2593 Op1 = Builder->CreateXor(X, Y);
2594 Op1 = Builder->CreateAnd(Op1, Z);
2595 I.setOperand(0, Op1);
2596 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2601 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2602 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2604 if ((Op0->hasOneUse() &&
2605 match(Op0, m_ZExt(m_Value(A))) &&
2606 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2607 (Op1->hasOneUse() &&
2608 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2609 match(Op1, m_ZExt(m_Value(A))))) {
2610 APInt Pow2 = Cst1->getValue() + 1;
2611 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2612 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2613 return new ICmpInst(I.getPredicate(), A,
2614 Builder->CreateTrunc(B, A->getType()));
2617 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2618 // "icmp (and X, mask), cst"
2620 if (Op0->hasOneUse() &&
2621 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2622 m_ConstantInt(ShAmt))))) &&
2623 match(Op1, m_ConstantInt(Cst1)) &&
2624 // Only do this when A has multiple uses. This is most important to do
2625 // when it exposes other optimizations.
2627 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2629 if (ShAmt < ASize) {
2631 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2634 APInt CmpV = Cst1->getValue().zext(ASize);
2637 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2638 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2644 Value *X; ConstantInt *Cst;
2646 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2647 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2650 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2651 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2653 return Changed ? &I : 0;
2661 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2663 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2666 if (!isa<ConstantFP>(RHSC)) return 0;
2667 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2669 // Get the width of the mantissa. We don't want to hack on conversions that
2670 // might lose information from the integer, e.g. "i64 -> float"
2671 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2672 if (MantissaWidth == -1) return 0; // Unknown.
2674 // Check to see that the input is converted from an integer type that is small
2675 // enough that preserves all bits. TODO: check here for "known" sign bits.
2676 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2677 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2679 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2680 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2684 // If the conversion would lose info, don't hack on this.
2685 if ((int)InputSize > MantissaWidth)
2688 // Otherwise, we can potentially simplify the comparison. We know that it
2689 // will always come through as an integer value and we know the constant is
2690 // not a NAN (it would have been previously simplified).
2691 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2693 ICmpInst::Predicate Pred;
2694 switch (I.getPredicate()) {
2695 default: llvm_unreachable("Unexpected predicate!");
2696 case FCmpInst::FCMP_UEQ:
2697 case FCmpInst::FCMP_OEQ:
2698 Pred = ICmpInst::ICMP_EQ;
2700 case FCmpInst::FCMP_UGT:
2701 case FCmpInst::FCMP_OGT:
2702 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2704 case FCmpInst::FCMP_UGE:
2705 case FCmpInst::FCMP_OGE:
2706 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2708 case FCmpInst::FCMP_ULT:
2709 case FCmpInst::FCMP_OLT:
2710 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2712 case FCmpInst::FCMP_ULE:
2713 case FCmpInst::FCMP_OLE:
2714 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2716 case FCmpInst::FCMP_UNE:
2717 case FCmpInst::FCMP_ONE:
2718 Pred = ICmpInst::ICMP_NE;
2720 case FCmpInst::FCMP_ORD:
2721 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2722 case FCmpInst::FCMP_UNO:
2723 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2726 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2728 // Now we know that the APFloat is a normal number, zero or inf.
2730 // See if the FP constant is too large for the integer. For example,
2731 // comparing an i8 to 300.0.
2732 unsigned IntWidth = IntTy->getScalarSizeInBits();
2735 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2736 // and large values.
2737 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2738 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2739 APFloat::rmNearestTiesToEven);
2740 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2741 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2742 Pred == ICmpInst::ICMP_SLE)
2743 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2744 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2747 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2748 // +INF and large values.
2749 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2750 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2751 APFloat::rmNearestTiesToEven);
2752 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2753 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2754 Pred == ICmpInst::ICMP_ULE)
2755 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2756 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2761 // See if the RHS value is < SignedMin.
2762 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2763 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2764 APFloat::rmNearestTiesToEven);
2765 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2766 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2767 Pred == ICmpInst::ICMP_SGE)
2768 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2769 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2772 // See if the RHS value is < UnsignedMin.
2773 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2774 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
2775 APFloat::rmNearestTiesToEven);
2776 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
2777 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
2778 Pred == ICmpInst::ICMP_UGE)
2779 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2780 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2784 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2785 // [0, UMAX], but it may still be fractional. See if it is fractional by
2786 // casting the FP value to the integer value and back, checking for equality.
2787 // Don't do this for zero, because -0.0 is not fractional.
2788 Constant *RHSInt = LHSUnsigned
2789 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2790 : ConstantExpr::getFPToSI(RHSC, IntTy);
2791 if (!RHS.isZero()) {
2792 bool Equal = LHSUnsigned
2793 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2794 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2796 // If we had a comparison against a fractional value, we have to adjust
2797 // the compare predicate and sometimes the value. RHSC is rounded towards
2798 // zero at this point.
2800 default: llvm_unreachable("Unexpected integer comparison!");
2801 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2802 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2803 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2804 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2805 case ICmpInst::ICMP_ULE:
2806 // (float)int <= 4.4 --> int <= 4
2807 // (float)int <= -4.4 --> false
2808 if (RHS.isNegative())
2809 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2811 case ICmpInst::ICMP_SLE:
2812 // (float)int <= 4.4 --> int <= 4
2813 // (float)int <= -4.4 --> int < -4
2814 if (RHS.isNegative())
2815 Pred = ICmpInst::ICMP_SLT;
2817 case ICmpInst::ICMP_ULT:
2818 // (float)int < -4.4 --> false
2819 // (float)int < 4.4 --> int <= 4
2820 if (RHS.isNegative())
2821 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2822 Pred = ICmpInst::ICMP_ULE;
2824 case ICmpInst::ICMP_SLT:
2825 // (float)int < -4.4 --> int < -4
2826 // (float)int < 4.4 --> int <= 4
2827 if (!RHS.isNegative())
2828 Pred = ICmpInst::ICMP_SLE;
2830 case ICmpInst::ICMP_UGT:
2831 // (float)int > 4.4 --> int > 4
2832 // (float)int > -4.4 --> true
2833 if (RHS.isNegative())
2834 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2836 case ICmpInst::ICMP_SGT:
2837 // (float)int > 4.4 --> int > 4
2838 // (float)int > -4.4 --> int >= -4
2839 if (RHS.isNegative())
2840 Pred = ICmpInst::ICMP_SGE;
2842 case ICmpInst::ICMP_UGE:
2843 // (float)int >= -4.4 --> true
2844 // (float)int >= 4.4 --> int > 4
2845 if (RHS.isNegative())
2846 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2847 Pred = ICmpInst::ICMP_UGT;
2849 case ICmpInst::ICMP_SGE:
2850 // (float)int >= -4.4 --> int >= -4
2851 // (float)int >= 4.4 --> int > 4
2852 if (!RHS.isNegative())
2853 Pred = ICmpInst::ICMP_SGT;
2859 // Lower this FP comparison into an appropriate integer version of the
2861 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2864 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2865 bool Changed = false;
2867 /// Orders the operands of the compare so that they are listed from most
2868 /// complex to least complex. This puts constants before unary operators,
2869 /// before binary operators.
2870 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2875 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2877 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2878 return ReplaceInstUsesWith(I, V);
2880 // Simplify 'fcmp pred X, X'
2882 switch (I.getPredicate()) {
2883 default: llvm_unreachable("Unknown predicate!");
2884 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2885 case FCmpInst::FCMP_ULT: // True if unordered or less than
2886 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2887 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2888 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2889 I.setPredicate(FCmpInst::FCMP_UNO);
2890 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2893 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2894 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2895 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2896 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2897 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2898 I.setPredicate(FCmpInst::FCMP_ORD);
2899 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2904 // Handle fcmp with constant RHS
2905 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2906 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2907 switch (LHSI->getOpcode()) {
2908 case Instruction::FPExt: {
2909 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2910 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2911 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2915 const fltSemantics *Sem;
2916 // FIXME: This shouldn't be here.
2917 if (LHSExt->getSrcTy()->isHalfTy())
2918 Sem = &APFloat::IEEEhalf;
2919 else if (LHSExt->getSrcTy()->isFloatTy())
2920 Sem = &APFloat::IEEEsingle;
2921 else if (LHSExt->getSrcTy()->isDoubleTy())
2922 Sem = &APFloat::IEEEdouble;
2923 else if (LHSExt->getSrcTy()->isFP128Ty())
2924 Sem = &APFloat::IEEEquad;
2925 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2926 Sem = &APFloat::x87DoubleExtended;
2927 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
2928 Sem = &APFloat::PPCDoubleDouble;
2933 APFloat F = RHSF->getValueAPF();
2934 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2936 // Avoid lossy conversions and denormals. Zero is a special case
2937 // that's OK to convert.
2941 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
2942 APFloat::cmpLessThan) || Fabs.isZero()))
2944 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2945 ConstantFP::get(RHSC->getContext(), F));
2948 case Instruction::PHI:
2949 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2950 // block. If in the same block, we're encouraging jump threading. If
2951 // not, we are just pessimizing the code by making an i1 phi.
2952 if (LHSI->getParent() == I.getParent())
2953 if (Instruction *NV = FoldOpIntoPhi(I))
2956 case Instruction::SIToFP:
2957 case Instruction::UIToFP:
2958 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2961 case Instruction::Select: {
2962 // If either operand of the select is a constant, we can fold the
2963 // comparison into the select arms, which will cause one to be
2964 // constant folded and the select turned into a bitwise or.
2965 Value *Op1 = 0, *Op2 = 0;
2966 if (LHSI->hasOneUse()) {
2967 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2968 // Fold the known value into the constant operand.
2969 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2970 // Insert a new FCmp of the other select operand.
2971 Op2 = Builder->CreateFCmp(I.getPredicate(),
2972 LHSI->getOperand(2), RHSC, I.getName());
2973 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2974 // Fold the known value into the constant operand.
2975 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2976 // Insert a new FCmp of the other select operand.
2977 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2983 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2986 case Instruction::FSub: {
2987 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2989 if (match(LHSI, m_FNeg(m_Value(Op))))
2990 return new FCmpInst(I.getSwappedPredicate(), Op,
2991 ConstantExpr::getFNeg(RHSC));
2994 case Instruction::Load:
2995 if (GetElementPtrInst *GEP =
2996 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2997 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2998 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2999 !cast<LoadInst>(LHSI)->isVolatile())
3000 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3004 case Instruction::Call: {
3005 CallInst *CI = cast<CallInst>(LHSI);
3007 // Various optimization for fabs compared with zero.
3008 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3009 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3011 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3012 Func == LibFunc::fabsl) {
3013 switch (I.getPredicate()) {
3015 // fabs(x) < 0 --> false
3016 case FCmpInst::FCMP_OLT:
3017 return ReplaceInstUsesWith(I, Builder->getFalse());
3018 // fabs(x) > 0 --> x != 0
3019 case FCmpInst::FCMP_OGT:
3020 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3022 // fabs(x) <= 0 --> x == 0
3023 case FCmpInst::FCMP_OLE:
3024 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3026 // fabs(x) >= 0 --> !isnan(x)
3027 case FCmpInst::FCMP_OGE:
3028 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3030 // fabs(x) == 0 --> x == 0
3031 // fabs(x) != 0 --> x != 0
3032 case FCmpInst::FCMP_OEQ:
3033 case FCmpInst::FCMP_UEQ:
3034 case FCmpInst::FCMP_ONE:
3035 case FCmpInst::FCMP_UNE:
3036 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3045 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3047 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3048 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3050 // fcmp (fpext x), (fpext y) -> fcmp x, y
3051 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3052 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3053 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3054 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3055 RHSExt->getOperand(0));
3057 return Changed ? &I : 0;