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);
1230 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
1231 // bit set in (X & AndCST) will produce a result greater than RHSV.
1232 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1233 unsigned NTZ = AndCST->getValue().countTrailingZeros();
1234 if ((NTZ < AndCST->getBitWidth()) &&
1235 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
1236 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1237 Constant::getNullValue(RHS->getType()));
1241 // Try to optimize things like "A[i]&42 == 0" to index computations.
1242 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1243 if (GetElementPtrInst *GEP =
1244 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1245 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1246 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1247 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1248 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1249 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1255 case Instruction::Or: {
1256 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1259 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1260 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1261 // -> and (icmp eq P, null), (icmp eq Q, null).
1262 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1263 Constant::getNullValue(P->getType()));
1264 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1265 Constant::getNullValue(Q->getType()));
1267 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1268 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1270 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1276 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1277 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1280 uint32_t TypeBits = RHSV.getBitWidth();
1282 // Check that the shift amount is in range. If not, don't perform
1283 // undefined shifts. When the shift is visited it will be
1285 if (ShAmt->uge(TypeBits))
1288 if (ICI.isEquality()) {
1289 // If we are comparing against bits always shifted out, the
1290 // comparison cannot succeed.
1292 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1294 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1295 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1297 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1298 return ReplaceInstUsesWith(ICI, Cst);
1301 // If the shift is NUW, then it is just shifting out zeros, no need for an
1303 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1304 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1305 ConstantExpr::getLShr(RHS, ShAmt));
1307 if (LHSI->hasOneUse()) {
1308 // Otherwise strength reduce the shift into an and.
1309 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1311 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1312 TypeBits-ShAmtVal));
1315 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1316 return new ICmpInst(ICI.getPredicate(), And,
1317 ConstantExpr::getLShr(RHS, ShAmt));
1321 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1322 bool TrueIfSigned = false;
1323 if (LHSI->hasOneUse() &&
1324 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1325 // (X << 31) <s 0 --> (X&1) != 0
1326 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1327 APInt::getOneBitSet(TypeBits,
1328 TypeBits-ShAmt->getZExtValue()-1));
1330 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1331 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1332 And, Constant::getNullValue(And->getType()));
1337 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1338 case Instruction::AShr: {
1339 // Handle equality comparisons of shift-by-constant.
1340 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1341 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1342 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1346 // Handle exact shr's.
1347 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1348 if (RHSV.isMinValue())
1349 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1354 case Instruction::SDiv:
1355 case Instruction::UDiv:
1356 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1357 // Fold this div into the comparison, producing a range check.
1358 // Determine, based on the divide type, what the range is being
1359 // checked. If there is an overflow on the low or high side, remember
1360 // it, otherwise compute the range [low, hi) bounding the new value.
1361 // See: InsertRangeTest above for the kinds of replacements possible.
1362 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1363 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1368 case Instruction::Add:
1369 // Fold: icmp pred (add X, C1), C2
1370 if (!ICI.isEquality()) {
1371 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1373 const APInt &LHSV = LHSC->getValue();
1375 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1378 if (ICI.isSigned()) {
1379 if (CR.getLower().isSignBit()) {
1380 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1381 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1382 } else if (CR.getUpper().isSignBit()) {
1383 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1384 ConstantInt::get(ICI.getContext(),CR.getLower()));
1387 if (CR.getLower().isMinValue()) {
1388 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1389 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1390 } else if (CR.getUpper().isMinValue()) {
1391 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1392 ConstantInt::get(ICI.getContext(),CR.getLower()));
1399 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1400 if (ICI.isEquality()) {
1401 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1403 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1404 // the second operand is a constant, simplify a bit.
1405 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1406 switch (BO->getOpcode()) {
1407 case Instruction::SRem:
1408 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1409 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1410 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1411 if (V.sgt(1) && V.isPowerOf2()) {
1413 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1415 return new ICmpInst(ICI.getPredicate(), NewRem,
1416 Constant::getNullValue(BO->getType()));
1420 case Instruction::Add:
1421 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1422 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1423 if (BO->hasOneUse())
1424 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1425 ConstantExpr::getSub(RHS, BOp1C));
1426 } else if (RHSV == 0) {
1427 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1428 // efficiently invertible, or if the add has just this one use.
1429 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1431 if (Value *NegVal = dyn_castNegVal(BOp1))
1432 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1433 if (Value *NegVal = dyn_castNegVal(BOp0))
1434 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1435 if (BO->hasOneUse()) {
1436 Value *Neg = Builder->CreateNeg(BOp1);
1438 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1442 case Instruction::Xor:
1443 // For the xor case, we can xor two constants together, eliminating
1444 // the explicit xor.
1445 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1446 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1447 ConstantExpr::getXor(RHS, BOC));
1448 } else if (RHSV == 0) {
1449 // Replace ((xor A, B) != 0) with (A != B)
1450 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1454 case Instruction::Sub:
1455 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1456 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1457 if (BO->hasOneUse())
1458 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1459 ConstantExpr::getSub(BOp0C, RHS));
1460 } else if (RHSV == 0) {
1461 // Replace ((sub A, B) != 0) with (A != B)
1462 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1466 case Instruction::Or:
1467 // If bits are being or'd in that are not present in the constant we
1468 // are comparing against, then the comparison could never succeed!
1469 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1470 Constant *NotCI = ConstantExpr::getNot(RHS);
1471 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1472 return ReplaceInstUsesWith(ICI,
1473 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1478 case Instruction::And:
1479 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1480 // If bits are being compared against that are and'd out, then the
1481 // comparison can never succeed!
1482 if ((RHSV & ~BOC->getValue()) != 0)
1483 return ReplaceInstUsesWith(ICI,
1484 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1487 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1488 if (RHS == BOC && RHSV.isPowerOf2())
1489 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1490 ICmpInst::ICMP_NE, LHSI,
1491 Constant::getNullValue(RHS->getType()));
1493 // Don't perform the following transforms if the AND has multiple uses
1494 if (!BO->hasOneUse())
1497 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1498 if (BOC->getValue().isSignBit()) {
1499 Value *X = BO->getOperand(0);
1500 Constant *Zero = Constant::getNullValue(X->getType());
1501 ICmpInst::Predicate pred = isICMP_NE ?
1502 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1503 return new ICmpInst(pred, X, Zero);
1506 // ((X & ~7) == 0) --> X < 8
1507 if (RHSV == 0 && isHighOnes(BOC)) {
1508 Value *X = BO->getOperand(0);
1509 Constant *NegX = ConstantExpr::getNeg(BOC);
1510 ICmpInst::Predicate pred = isICMP_NE ?
1511 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1512 return new ICmpInst(pred, X, NegX);
1517 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1518 // Handle icmp {eq|ne} <intrinsic>, intcst.
1519 switch (II->getIntrinsicID()) {
1520 case Intrinsic::bswap:
1522 ICI.setOperand(0, II->getArgOperand(0));
1523 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1525 case Intrinsic::ctlz:
1526 case Intrinsic::cttz:
1527 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1528 if (RHSV == RHS->getType()->getBitWidth()) {
1530 ICI.setOperand(0, II->getArgOperand(0));
1531 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1535 case Intrinsic::ctpop:
1536 // popcount(A) == 0 -> A == 0 and likewise for !=
1537 if (RHS->isZero()) {
1539 ICI.setOperand(0, II->getArgOperand(0));
1540 ICI.setOperand(1, RHS);
1552 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1553 /// We only handle extending casts so far.
1555 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1556 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1557 Value *LHSCIOp = LHSCI->getOperand(0);
1558 Type *SrcTy = LHSCIOp->getType();
1559 Type *DestTy = LHSCI->getType();
1562 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1563 // integer type is the same size as the pointer type.
1564 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1565 TD->getPointerSizeInBits() ==
1566 cast<IntegerType>(DestTy)->getBitWidth()) {
1568 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1569 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1570 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1571 RHSOp = RHSC->getOperand(0);
1572 // If the pointer types don't match, insert a bitcast.
1573 if (LHSCIOp->getType() != RHSOp->getType())
1574 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1578 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1581 // The code below only handles extension cast instructions, so far.
1583 if (LHSCI->getOpcode() != Instruction::ZExt &&
1584 LHSCI->getOpcode() != Instruction::SExt)
1587 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1588 bool isSignedCmp = ICI.isSigned();
1590 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1591 // Not an extension from the same type?
1592 RHSCIOp = CI->getOperand(0);
1593 if (RHSCIOp->getType() != LHSCIOp->getType())
1596 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1597 // and the other is a zext), then we can't handle this.
1598 if (CI->getOpcode() != LHSCI->getOpcode())
1601 // Deal with equality cases early.
1602 if (ICI.isEquality())
1603 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1605 // A signed comparison of sign extended values simplifies into a
1606 // signed comparison.
1607 if (isSignedCmp && isSignedExt)
1608 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1610 // The other three cases all fold into an unsigned comparison.
1611 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1614 // If we aren't dealing with a constant on the RHS, exit early
1615 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1619 // Compute the constant that would happen if we truncated to SrcTy then
1620 // reextended to DestTy.
1621 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1622 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1625 // If the re-extended constant didn't change...
1627 // Deal with equality cases early.
1628 if (ICI.isEquality())
1629 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1631 // A signed comparison of sign extended values simplifies into a
1632 // signed comparison.
1633 if (isSignedExt && isSignedCmp)
1634 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1636 // The other three cases all fold into an unsigned comparison.
1637 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1640 // The re-extended constant changed so the constant cannot be represented
1641 // in the shorter type. Consequently, we cannot emit a simple comparison.
1642 // All the cases that fold to true or false will have already been handled
1643 // by SimplifyICmpInst, so only deal with the tricky case.
1645 if (isSignedCmp || !isSignedExt)
1648 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1649 // should have been folded away previously and not enter in here.
1651 // We're performing an unsigned comp with a sign extended value.
1652 // This is true if the input is >= 0. [aka >s -1]
1653 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1654 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1656 // Finally, return the value computed.
1657 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1658 return ReplaceInstUsesWith(ICI, Result);
1660 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1661 return BinaryOperator::CreateNot(Result);
1664 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1665 /// I = icmp ugt (add (add A, B), CI2), CI1
1666 /// If this is of the form:
1668 /// if (sum+128 >u 255)
1669 /// Then replace it with llvm.sadd.with.overflow.i8.
1671 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1672 ConstantInt *CI2, ConstantInt *CI1,
1674 // The transformation we're trying to do here is to transform this into an
1675 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1676 // with a narrower add, and discard the add-with-constant that is part of the
1677 // range check (if we can't eliminate it, this isn't profitable).
1679 // In order to eliminate the add-with-constant, the compare can be its only
1681 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1682 if (!AddWithCst->hasOneUse()) return 0;
1684 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1685 if (!CI2->getValue().isPowerOf2()) return 0;
1686 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1687 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1689 // The width of the new add formed is 1 more than the bias.
1692 // Check to see that CI1 is an all-ones value with NewWidth bits.
1693 if (CI1->getBitWidth() == NewWidth ||
1694 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1697 // This is only really a signed overflow check if the inputs have been
1698 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1699 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1700 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1701 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1702 IC.ComputeNumSignBits(B) < NeededSignBits)
1705 // In order to replace the original add with a narrower
1706 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1707 // and truncates that discard the high bits of the add. Verify that this is
1709 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1710 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1712 if (*UI == AddWithCst) continue;
1714 // Only accept truncates for now. We would really like a nice recursive
1715 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1716 // chain to see which bits of a value are actually demanded. If the
1717 // original add had another add which was then immediately truncated, we
1718 // could still do the transformation.
1719 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1721 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1724 // If the pattern matches, truncate the inputs to the narrower type and
1725 // use the sadd_with_overflow intrinsic to efficiently compute both the
1726 // result and the overflow bit.
1727 Module *M = I.getParent()->getParent()->getParent();
1729 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1730 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1733 InstCombiner::BuilderTy *Builder = IC.Builder;
1735 // Put the new code above the original add, in case there are any uses of the
1736 // add between the add and the compare.
1737 Builder->SetInsertPoint(OrigAdd);
1739 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1740 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1741 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1742 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1743 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1745 // The inner add was the result of the narrow add, zero extended to the
1746 // wider type. Replace it with the result computed by the intrinsic.
1747 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1749 // The original icmp gets replaced with the overflow value.
1750 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1753 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1755 // Don't bother doing this transformation for pointers, don't do it for
1757 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1759 // If the add is a constant expr, then we don't bother transforming it.
1760 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1761 if (OrigAdd == 0) return 0;
1763 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1765 // Put the new code above the original add, in case there are any uses of the
1766 // add between the add and the compare.
1767 InstCombiner::BuilderTy *Builder = IC.Builder;
1768 Builder->SetInsertPoint(OrigAdd);
1770 Module *M = I.getParent()->getParent()->getParent();
1771 Type *Ty = LHS->getType();
1772 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1773 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1774 Value *Add = Builder->CreateExtractValue(Call, 0);
1776 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1778 // The original icmp gets replaced with the overflow value.
1779 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1782 // DemandedBitsLHSMask - When performing a comparison against a constant,
1783 // it is possible that not all the bits in the LHS are demanded. This helper
1784 // method computes the mask that IS demanded.
1785 static APInt DemandedBitsLHSMask(ICmpInst &I,
1786 unsigned BitWidth, bool isSignCheck) {
1788 return APInt::getSignBit(BitWidth);
1790 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1791 if (!CI) return APInt::getAllOnesValue(BitWidth);
1792 const APInt &RHS = CI->getValue();
1794 switch (I.getPredicate()) {
1795 // For a UGT comparison, we don't care about any bits that
1796 // correspond to the trailing ones of the comparand. The value of these
1797 // bits doesn't impact the outcome of the comparison, because any value
1798 // greater than the RHS must differ in a bit higher than these due to carry.
1799 case ICmpInst::ICMP_UGT: {
1800 unsigned trailingOnes = RHS.countTrailingOnes();
1801 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1805 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1806 // Any value less than the RHS must differ in a higher bit because of carries.
1807 case ICmpInst::ICMP_ULT: {
1808 unsigned trailingZeros = RHS.countTrailingZeros();
1809 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1814 return APInt::getAllOnesValue(BitWidth);
1819 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1820 bool Changed = false;
1821 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1823 /// Orders the operands of the compare so that they are listed from most
1824 /// complex to least complex. This puts constants before unary operators,
1825 /// before binary operators.
1826 if (getComplexity(Op0) < getComplexity(Op1)) {
1828 std::swap(Op0, Op1);
1832 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1833 return ReplaceInstUsesWith(I, V);
1835 // comparing -val or val with non-zero is the same as just comparing val
1836 // ie, abs(val) != 0 -> val != 0
1837 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
1839 Value *Cond, *SelectTrue, *SelectFalse;
1840 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
1841 m_Value(SelectFalse)))) {
1842 if (Value *V = dyn_castNegVal(SelectTrue)) {
1843 if (V == SelectFalse)
1844 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1846 else if (Value *V = dyn_castNegVal(SelectFalse)) {
1847 if (V == SelectTrue)
1848 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
1853 Type *Ty = Op0->getType();
1855 // icmp's with boolean values can always be turned into bitwise operations
1856 if (Ty->isIntegerTy(1)) {
1857 switch (I.getPredicate()) {
1858 default: llvm_unreachable("Invalid icmp instruction!");
1859 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1860 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1861 return BinaryOperator::CreateNot(Xor);
1863 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1864 return BinaryOperator::CreateXor(Op0, Op1);
1866 case ICmpInst::ICMP_UGT:
1867 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1869 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1870 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1871 return BinaryOperator::CreateAnd(Not, Op1);
1873 case ICmpInst::ICMP_SGT:
1874 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1876 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1877 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1878 return BinaryOperator::CreateAnd(Not, Op0);
1880 case ICmpInst::ICMP_UGE:
1881 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1883 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1884 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1885 return BinaryOperator::CreateOr(Not, Op1);
1887 case ICmpInst::ICMP_SGE:
1888 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1890 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1891 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1892 return BinaryOperator::CreateOr(Not, Op0);
1897 unsigned BitWidth = 0;
1898 if (Ty->isIntOrIntVectorTy())
1899 BitWidth = Ty->getScalarSizeInBits();
1900 else if (TD) // Pointers require TD info to get their size.
1901 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1903 bool isSignBit = false;
1905 // See if we are doing a comparison with a constant.
1906 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1907 Value *A = 0, *B = 0;
1909 // Match the following pattern, which is a common idiom when writing
1910 // overflow-safe integer arithmetic function. The source performs an
1911 // addition in wider type, and explicitly checks for overflow using
1912 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1913 // sadd_with_overflow intrinsic.
1915 // TODO: This could probably be generalized to handle other overflow-safe
1916 // operations if we worked out the formulas to compute the appropriate
1920 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1922 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1923 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1924 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1925 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1929 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1930 if (I.isEquality() && CI->isZero() &&
1931 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1932 // (icmp cond A B) if cond is equality
1933 return new ICmpInst(I.getPredicate(), A, B);
1936 // If we have an icmp le or icmp ge instruction, turn it into the
1937 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1938 // them being folded in the code below. The SimplifyICmpInst code has
1939 // already handled the edge cases for us, so we just assert on them.
1940 switch (I.getPredicate()) {
1942 case ICmpInst::ICMP_ULE:
1943 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1944 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1945 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1946 case ICmpInst::ICMP_SLE:
1947 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1948 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1949 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1950 case ICmpInst::ICMP_UGE:
1951 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1952 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1953 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1954 case ICmpInst::ICMP_SGE:
1955 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1956 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1957 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1960 // If this comparison is a normal comparison, it demands all
1961 // bits, if it is a sign bit comparison, it only demands the sign bit.
1963 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1966 // See if we can fold the comparison based on range information we can get
1967 // by checking whether bits are known to be zero or one in the input.
1968 if (BitWidth != 0) {
1969 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1970 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1972 if (SimplifyDemandedBits(I.getOperandUse(0),
1973 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1974 Op0KnownZero, Op0KnownOne, 0))
1976 if (SimplifyDemandedBits(I.getOperandUse(1),
1977 APInt::getAllOnesValue(BitWidth),
1978 Op1KnownZero, Op1KnownOne, 0))
1981 // Given the known and unknown bits, compute a range that the LHS could be
1982 // in. Compute the Min, Max and RHS values based on the known bits. For the
1983 // EQ and NE we use unsigned values.
1984 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1985 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1987 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1989 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1992 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1994 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1998 // If Min and Max are known to be the same, then SimplifyDemandedBits
1999 // figured out that the LHS is a constant. Just constant fold this now so
2000 // that code below can assume that Min != Max.
2001 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2002 return new ICmpInst(I.getPredicate(),
2003 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2004 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2005 return new ICmpInst(I.getPredicate(), Op0,
2006 ConstantInt::get(Op1->getType(), Op1Min));
2008 // Based on the range information we know about the LHS, see if we can
2009 // simplify this comparison. For example, (x&4) < 8 is always true.
2010 switch (I.getPredicate()) {
2011 default: llvm_unreachable("Unknown icmp opcode!");
2012 case ICmpInst::ICMP_EQ: {
2013 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2014 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2016 // If all bits are known zero except for one, then we know at most one
2017 // bit is set. If the comparison is against zero, then this is a check
2018 // to see if *that* bit is set.
2019 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2020 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2021 // If the LHS is an AND with the same constant, look through it.
2023 ConstantInt *LHSC = 0;
2024 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2025 LHSC->getValue() != Op0KnownZeroInverted)
2028 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2029 // then turn "((1 << x)&8) == 0" into "x != 3".
2031 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2032 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2033 return new ICmpInst(ICmpInst::ICMP_NE, X,
2034 ConstantInt::get(X->getType(), CmpVal));
2037 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2038 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2040 if (Op0KnownZeroInverted == 1 &&
2041 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2042 return new ICmpInst(ICmpInst::ICMP_NE, X,
2043 ConstantInt::get(X->getType(),
2044 CI->countTrailingZeros()));
2049 case ICmpInst::ICMP_NE: {
2050 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2051 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2053 // If all bits are known zero except for one, then we know at most one
2054 // bit is set. If the comparison is against zero, then this is a check
2055 // to see if *that* bit is set.
2056 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2057 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2058 // If the LHS is an AND with the same constant, look through it.
2060 ConstantInt *LHSC = 0;
2061 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2062 LHSC->getValue() != Op0KnownZeroInverted)
2065 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2066 // then turn "((1 << x)&8) != 0" into "x == 3".
2068 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2069 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2070 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2071 ConstantInt::get(X->getType(), CmpVal));
2074 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2075 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2077 if (Op0KnownZeroInverted == 1 &&
2078 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2079 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2080 ConstantInt::get(X->getType(),
2081 CI->countTrailingZeros()));
2086 case ICmpInst::ICMP_ULT:
2087 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2088 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2089 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2090 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2091 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2092 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2093 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2094 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2095 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2096 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2098 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2099 if (CI->isMinValue(true))
2100 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2101 Constant::getAllOnesValue(Op0->getType()));
2104 case ICmpInst::ICMP_UGT:
2105 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2106 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2107 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2108 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2110 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2111 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2112 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2113 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2114 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2115 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2117 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2118 if (CI->isMaxValue(true))
2119 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2120 Constant::getNullValue(Op0->getType()));
2123 case ICmpInst::ICMP_SLT:
2124 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2125 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2126 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2127 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2128 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2129 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2130 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2131 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2132 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2133 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2136 case ICmpInst::ICMP_SGT:
2137 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2138 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2139 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2140 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2142 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2143 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2144 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2145 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2146 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2147 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2150 case ICmpInst::ICMP_SGE:
2151 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2152 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2153 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2154 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2155 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2157 case ICmpInst::ICMP_SLE:
2158 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2159 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2160 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2161 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2162 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2164 case ICmpInst::ICMP_UGE:
2165 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2166 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2167 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2168 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2169 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2171 case ICmpInst::ICMP_ULE:
2172 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2173 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2174 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2175 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2176 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2180 // Turn a signed comparison into an unsigned one if both operands
2181 // are known to have the same sign.
2183 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2184 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2185 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2188 // Test if the ICmpInst instruction is used exclusively by a select as
2189 // part of a minimum or maximum operation. If so, refrain from doing
2190 // any other folding. This helps out other analyses which understand
2191 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2192 // and CodeGen. And in this case, at least one of the comparison
2193 // operands has at least one user besides the compare (the select),
2194 // which would often largely negate the benefit of folding anyway.
2196 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2197 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2198 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2201 // See if we are doing a comparison between a constant and an instruction that
2202 // can be folded into the comparison.
2203 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2204 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2205 // instruction, see if that instruction also has constants so that the
2206 // instruction can be folded into the icmp
2207 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2208 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2212 // Handle icmp with constant (but not simple integer constant) RHS
2213 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2214 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2215 switch (LHSI->getOpcode()) {
2216 case Instruction::GetElementPtr:
2217 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2218 if (RHSC->isNullValue() &&
2219 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2220 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2221 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2223 case Instruction::PHI:
2224 // Only fold icmp into the PHI if the phi and icmp are in the same
2225 // block. If in the same block, we're encouraging jump threading. If
2226 // not, we are just pessimizing the code by making an i1 phi.
2227 if (LHSI->getParent() == I.getParent())
2228 if (Instruction *NV = FoldOpIntoPhi(I))
2231 case Instruction::Select: {
2232 // If either operand of the select is a constant, we can fold the
2233 // comparison into the select arms, which will cause one to be
2234 // constant folded and the select turned into a bitwise or.
2235 Value *Op1 = 0, *Op2 = 0;
2236 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2237 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2238 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2239 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2241 // We only want to perform this transformation if it will not lead to
2242 // additional code. This is true if either both sides of the select
2243 // fold to a constant (in which case the icmp is replaced with a select
2244 // which will usually simplify) or this is the only user of the
2245 // select (in which case we are trading a select+icmp for a simpler
2247 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2249 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2252 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2254 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2258 case Instruction::IntToPtr:
2259 // icmp pred inttoptr(X), null -> icmp pred X, 0
2260 if (RHSC->isNullValue() && TD &&
2261 TD->getIntPtrType(RHSC->getContext()) ==
2262 LHSI->getOperand(0)->getType())
2263 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2264 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2267 case Instruction::Load:
2268 // Try to optimize things like "A[i] > 4" to index computations.
2269 if (GetElementPtrInst *GEP =
2270 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2271 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2272 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2273 !cast<LoadInst>(LHSI)->isVolatile())
2274 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2281 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2282 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2283 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2285 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2286 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2287 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2290 // Test to see if the operands of the icmp are casted versions of other
2291 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2293 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2294 if (Op0->getType()->isPointerTy() &&
2295 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2296 // We keep moving the cast from the left operand over to the right
2297 // operand, where it can often be eliminated completely.
2298 Op0 = CI->getOperand(0);
2300 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2301 // so eliminate it as well.
2302 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2303 Op1 = CI2->getOperand(0);
2305 // If Op1 is a constant, we can fold the cast into the constant.
2306 if (Op0->getType() != Op1->getType()) {
2307 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2308 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2310 // Otherwise, cast the RHS right before the icmp
2311 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2314 return new ICmpInst(I.getPredicate(), Op0, Op1);
2318 if (isa<CastInst>(Op0)) {
2319 // Handle the special case of: icmp (cast bool to X), <cst>
2320 // This comes up when you have code like
2323 // For generality, we handle any zero-extension of any operand comparison
2324 // with a constant or another cast from the same type.
2325 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2326 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2330 // Special logic for binary operators.
2331 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2332 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2334 CmpInst::Predicate Pred = I.getPredicate();
2335 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2336 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2337 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2338 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2339 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2340 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2341 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2342 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2343 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2345 // Analyze the case when either Op0 or Op1 is an add instruction.
2346 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2347 Value *A = 0, *B = 0, *C = 0, *D = 0;
2348 if (BO0 && BO0->getOpcode() == Instruction::Add)
2349 A = BO0->getOperand(0), B = BO0->getOperand(1);
2350 if (BO1 && BO1->getOpcode() == Instruction::Add)
2351 C = BO1->getOperand(0), D = BO1->getOperand(1);
2353 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2354 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2355 return new ICmpInst(Pred, A == Op1 ? B : A,
2356 Constant::getNullValue(Op1->getType()));
2358 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2359 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2360 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2363 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2364 if (A && C && (A == C || A == D || B == C || B == D) &&
2365 NoOp0WrapProblem && NoOp1WrapProblem &&
2366 // Try not to increase register pressure.
2367 BO0->hasOneUse() && BO1->hasOneUse()) {
2368 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2371 // C + B == C + D -> B == D
2374 } else if (A == D) {
2375 // D + B == C + D -> B == C
2378 } else if (B == C) {
2379 // A + C == C + D -> A == D
2384 // A + D == C + D -> A == C
2388 return new ICmpInst(Pred, Y, Z);
2391 // Analyze the case when either Op0 or Op1 is a sub instruction.
2392 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2393 A = 0; B = 0; C = 0; D = 0;
2394 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2395 A = BO0->getOperand(0), B = BO0->getOperand(1);
2396 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2397 C = BO1->getOperand(0), D = BO1->getOperand(1);
2399 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2400 if (A == Op1 && NoOp0WrapProblem)
2401 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2403 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2404 if (C == Op0 && NoOp1WrapProblem)
2405 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2407 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2408 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2409 // Try not to increase register pressure.
2410 BO0->hasOneUse() && BO1->hasOneUse())
2411 return new ICmpInst(Pred, A, C);
2413 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2414 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2415 // Try not to increase register pressure.
2416 BO0->hasOneUse() && BO1->hasOneUse())
2417 return new ICmpInst(Pred, D, B);
2419 BinaryOperator *SRem = NULL;
2420 // icmp (srem X, Y), Y
2421 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2422 Op1 == BO0->getOperand(1))
2424 // icmp Y, (srem X, Y)
2425 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2426 Op0 == BO1->getOperand(1))
2429 // We don't check hasOneUse to avoid increasing register pressure because
2430 // the value we use is the same value this instruction was already using.
2431 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2433 case ICmpInst::ICMP_EQ:
2434 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2435 case ICmpInst::ICMP_NE:
2436 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2437 case ICmpInst::ICMP_SGT:
2438 case ICmpInst::ICMP_SGE:
2439 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2440 Constant::getAllOnesValue(SRem->getType()));
2441 case ICmpInst::ICMP_SLT:
2442 case ICmpInst::ICMP_SLE:
2443 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2444 Constant::getNullValue(SRem->getType()));
2448 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2449 BO0->hasOneUse() && BO1->hasOneUse() &&
2450 BO0->getOperand(1) == BO1->getOperand(1)) {
2451 switch (BO0->getOpcode()) {
2453 case Instruction::Add:
2454 case Instruction::Sub:
2455 case Instruction::Xor:
2456 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2457 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2458 BO1->getOperand(0));
2459 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2460 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2461 if (CI->getValue().isSignBit()) {
2462 ICmpInst::Predicate Pred = I.isSigned()
2463 ? I.getUnsignedPredicate()
2464 : I.getSignedPredicate();
2465 return new ICmpInst(Pred, BO0->getOperand(0),
2466 BO1->getOperand(0));
2469 if (CI->isMaxValue(true)) {
2470 ICmpInst::Predicate Pred = I.isSigned()
2471 ? I.getUnsignedPredicate()
2472 : I.getSignedPredicate();
2473 Pred = I.getSwappedPredicate(Pred);
2474 return new ICmpInst(Pred, BO0->getOperand(0),
2475 BO1->getOperand(0));
2479 case Instruction::Mul:
2480 if (!I.isEquality())
2483 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2484 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2485 // Mask = -1 >> count-trailing-zeros(Cst).
2486 if (!CI->isZero() && !CI->isOne()) {
2487 const APInt &AP = CI->getValue();
2488 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2489 APInt::getLowBitsSet(AP.getBitWidth(),
2491 AP.countTrailingZeros()));
2492 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2493 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2494 return new ICmpInst(I.getPredicate(), And1, And2);
2498 case Instruction::UDiv:
2499 case Instruction::LShr:
2503 case Instruction::SDiv:
2504 case Instruction::AShr:
2505 if (!BO0->isExact() || !BO1->isExact())
2507 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2508 BO1->getOperand(0));
2509 case Instruction::Shl: {
2510 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2511 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2514 if (!NSW && I.isSigned())
2516 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2517 BO1->getOperand(0));
2524 // ~x < ~y --> y < x
2525 // ~x < cst --> ~cst < x
2526 if (match(Op0, m_Not(m_Value(A)))) {
2527 if (match(Op1, m_Not(m_Value(B))))
2528 return new ICmpInst(I.getPredicate(), B, A);
2529 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2530 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2533 // (a+b) <u a --> llvm.uadd.with.overflow.
2534 // (a+b) <u b --> llvm.uadd.with.overflow.
2535 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2536 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2537 (Op1 == A || Op1 == B))
2538 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2541 // a >u (a+b) --> llvm.uadd.with.overflow.
2542 // b >u (a+b) --> llvm.uadd.with.overflow.
2543 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2544 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2545 (Op0 == A || Op0 == B))
2546 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2550 if (I.isEquality()) {
2551 Value *A, *B, *C, *D;
2553 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2554 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2555 Value *OtherVal = A == Op1 ? B : A;
2556 return new ICmpInst(I.getPredicate(), OtherVal,
2557 Constant::getNullValue(A->getType()));
2560 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2561 // A^c1 == C^c2 --> A == C^(c1^c2)
2562 ConstantInt *C1, *C2;
2563 if (match(B, m_ConstantInt(C1)) &&
2564 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2565 Constant *NC = ConstantInt::get(I.getContext(),
2566 C1->getValue() ^ C2->getValue());
2567 Value *Xor = Builder->CreateXor(C, NC);
2568 return new ICmpInst(I.getPredicate(), A, Xor);
2571 // A^B == A^D -> B == D
2572 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2573 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2574 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2575 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2579 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2580 (A == Op0 || B == Op0)) {
2581 // A == (A^B) -> B == 0
2582 Value *OtherVal = A == Op0 ? B : A;
2583 return new ICmpInst(I.getPredicate(), OtherVal,
2584 Constant::getNullValue(A->getType()));
2587 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2588 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2589 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2590 Value *X = 0, *Y = 0, *Z = 0;
2593 X = B; Y = D; Z = A;
2594 } else if (A == D) {
2595 X = B; Y = C; Z = A;
2596 } else if (B == C) {
2597 X = A; Y = D; Z = B;
2598 } else if (B == D) {
2599 X = A; Y = C; Z = B;
2602 if (X) { // Build (X^Y) & Z
2603 Op1 = Builder->CreateXor(X, Y);
2604 Op1 = Builder->CreateAnd(Op1, Z);
2605 I.setOperand(0, Op1);
2606 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2611 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2612 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2614 if ((Op0->hasOneUse() &&
2615 match(Op0, m_ZExt(m_Value(A))) &&
2616 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2617 (Op1->hasOneUse() &&
2618 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2619 match(Op1, m_ZExt(m_Value(A))))) {
2620 APInt Pow2 = Cst1->getValue() + 1;
2621 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2622 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2623 return new ICmpInst(I.getPredicate(), A,
2624 Builder->CreateTrunc(B, A->getType()));
2627 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2628 // "icmp (and X, mask), cst"
2630 if (Op0->hasOneUse() &&
2631 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2632 m_ConstantInt(ShAmt))))) &&
2633 match(Op1, m_ConstantInt(Cst1)) &&
2634 // Only do this when A has multiple uses. This is most important to do
2635 // when it exposes other optimizations.
2637 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2639 if (ShAmt < ASize) {
2641 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2644 APInt CmpV = Cst1->getValue().zext(ASize);
2647 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2648 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2654 Value *X; ConstantInt *Cst;
2656 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2657 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2660 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2661 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2663 return Changed ? &I : 0;
2671 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2673 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2676 if (!isa<ConstantFP>(RHSC)) return 0;
2677 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2679 // Get the width of the mantissa. We don't want to hack on conversions that
2680 // might lose information from the integer, e.g. "i64 -> float"
2681 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2682 if (MantissaWidth == -1) return 0; // Unknown.
2684 // Check to see that the input is converted from an integer type that is small
2685 // enough that preserves all bits. TODO: check here for "known" sign bits.
2686 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2687 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2689 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2690 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2694 // If the conversion would lose info, don't hack on this.
2695 if ((int)InputSize > MantissaWidth)
2698 // Otherwise, we can potentially simplify the comparison. We know that it
2699 // will always come through as an integer value and we know the constant is
2700 // not a NAN (it would have been previously simplified).
2701 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2703 ICmpInst::Predicate Pred;
2704 switch (I.getPredicate()) {
2705 default: llvm_unreachable("Unexpected predicate!");
2706 case FCmpInst::FCMP_UEQ:
2707 case FCmpInst::FCMP_OEQ:
2708 Pred = ICmpInst::ICMP_EQ;
2710 case FCmpInst::FCMP_UGT:
2711 case FCmpInst::FCMP_OGT:
2712 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2714 case FCmpInst::FCMP_UGE:
2715 case FCmpInst::FCMP_OGE:
2716 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2718 case FCmpInst::FCMP_ULT:
2719 case FCmpInst::FCMP_OLT:
2720 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2722 case FCmpInst::FCMP_ULE:
2723 case FCmpInst::FCMP_OLE:
2724 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2726 case FCmpInst::FCMP_UNE:
2727 case FCmpInst::FCMP_ONE:
2728 Pred = ICmpInst::ICMP_NE;
2730 case FCmpInst::FCMP_ORD:
2731 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2732 case FCmpInst::FCMP_UNO:
2733 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2736 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2738 // Now we know that the APFloat is a normal number, zero or inf.
2740 // See if the FP constant is too large for the integer. For example,
2741 // comparing an i8 to 300.0.
2742 unsigned IntWidth = IntTy->getScalarSizeInBits();
2745 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2746 // and large values.
2747 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2748 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2749 APFloat::rmNearestTiesToEven);
2750 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2751 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2752 Pred == ICmpInst::ICMP_SLE)
2753 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2754 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2757 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2758 // +INF and large values.
2759 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2760 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2761 APFloat::rmNearestTiesToEven);
2762 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2763 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2764 Pred == ICmpInst::ICMP_ULE)
2765 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2766 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2771 // See if the RHS value is < SignedMin.
2772 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2773 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2774 APFloat::rmNearestTiesToEven);
2775 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2776 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2777 Pred == ICmpInst::ICMP_SGE)
2778 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2779 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2782 // See if the RHS value is < UnsignedMin.
2783 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2784 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
2785 APFloat::rmNearestTiesToEven);
2786 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
2787 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
2788 Pred == ICmpInst::ICMP_UGE)
2789 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2790 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2794 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2795 // [0, UMAX], but it may still be fractional. See if it is fractional by
2796 // casting the FP value to the integer value and back, checking for equality.
2797 // Don't do this for zero, because -0.0 is not fractional.
2798 Constant *RHSInt = LHSUnsigned
2799 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2800 : ConstantExpr::getFPToSI(RHSC, IntTy);
2801 if (!RHS.isZero()) {
2802 bool Equal = LHSUnsigned
2803 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2804 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2806 // If we had a comparison against a fractional value, we have to adjust
2807 // the compare predicate and sometimes the value. RHSC is rounded towards
2808 // zero at this point.
2810 default: llvm_unreachable("Unexpected integer comparison!");
2811 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2812 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2813 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2814 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2815 case ICmpInst::ICMP_ULE:
2816 // (float)int <= 4.4 --> int <= 4
2817 // (float)int <= -4.4 --> false
2818 if (RHS.isNegative())
2819 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2821 case ICmpInst::ICMP_SLE:
2822 // (float)int <= 4.4 --> int <= 4
2823 // (float)int <= -4.4 --> int < -4
2824 if (RHS.isNegative())
2825 Pred = ICmpInst::ICMP_SLT;
2827 case ICmpInst::ICMP_ULT:
2828 // (float)int < -4.4 --> false
2829 // (float)int < 4.4 --> int <= 4
2830 if (RHS.isNegative())
2831 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2832 Pred = ICmpInst::ICMP_ULE;
2834 case ICmpInst::ICMP_SLT:
2835 // (float)int < -4.4 --> int < -4
2836 // (float)int < 4.4 --> int <= 4
2837 if (!RHS.isNegative())
2838 Pred = ICmpInst::ICMP_SLE;
2840 case ICmpInst::ICMP_UGT:
2841 // (float)int > 4.4 --> int > 4
2842 // (float)int > -4.4 --> true
2843 if (RHS.isNegative())
2844 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2846 case ICmpInst::ICMP_SGT:
2847 // (float)int > 4.4 --> int > 4
2848 // (float)int > -4.4 --> int >= -4
2849 if (RHS.isNegative())
2850 Pred = ICmpInst::ICMP_SGE;
2852 case ICmpInst::ICMP_UGE:
2853 // (float)int >= -4.4 --> true
2854 // (float)int >= 4.4 --> int > 4
2855 if (RHS.isNegative())
2856 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2857 Pred = ICmpInst::ICMP_UGT;
2859 case ICmpInst::ICMP_SGE:
2860 // (float)int >= -4.4 --> int >= -4
2861 // (float)int >= 4.4 --> int > 4
2862 if (!RHS.isNegative())
2863 Pred = ICmpInst::ICMP_SGT;
2869 // Lower this FP comparison into an appropriate integer version of the
2871 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2874 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2875 bool Changed = false;
2877 /// Orders the operands of the compare so that they are listed from most
2878 /// complex to least complex. This puts constants before unary operators,
2879 /// before binary operators.
2880 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2885 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2887 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2888 return ReplaceInstUsesWith(I, V);
2890 // Simplify 'fcmp pred X, X'
2892 switch (I.getPredicate()) {
2893 default: llvm_unreachable("Unknown predicate!");
2894 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2895 case FCmpInst::FCMP_ULT: // True if unordered or less than
2896 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2897 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2898 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2899 I.setPredicate(FCmpInst::FCMP_UNO);
2900 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2903 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2904 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2905 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2906 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2907 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2908 I.setPredicate(FCmpInst::FCMP_ORD);
2909 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2914 // Handle fcmp with constant RHS
2915 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2916 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2917 switch (LHSI->getOpcode()) {
2918 case Instruction::FPExt: {
2919 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2920 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2921 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2925 const fltSemantics *Sem;
2926 // FIXME: This shouldn't be here.
2927 if (LHSExt->getSrcTy()->isHalfTy())
2928 Sem = &APFloat::IEEEhalf;
2929 else if (LHSExt->getSrcTy()->isFloatTy())
2930 Sem = &APFloat::IEEEsingle;
2931 else if (LHSExt->getSrcTy()->isDoubleTy())
2932 Sem = &APFloat::IEEEdouble;
2933 else if (LHSExt->getSrcTy()->isFP128Ty())
2934 Sem = &APFloat::IEEEquad;
2935 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2936 Sem = &APFloat::x87DoubleExtended;
2937 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
2938 Sem = &APFloat::PPCDoubleDouble;
2943 APFloat F = RHSF->getValueAPF();
2944 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2946 // Avoid lossy conversions and denormals. Zero is a special case
2947 // that's OK to convert.
2951 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
2952 APFloat::cmpLessThan) || Fabs.isZero()))
2954 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2955 ConstantFP::get(RHSC->getContext(), F));
2958 case Instruction::PHI:
2959 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2960 // block. If in the same block, we're encouraging jump threading. If
2961 // not, we are just pessimizing the code by making an i1 phi.
2962 if (LHSI->getParent() == I.getParent())
2963 if (Instruction *NV = FoldOpIntoPhi(I))
2966 case Instruction::SIToFP:
2967 case Instruction::UIToFP:
2968 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2971 case Instruction::Select: {
2972 // If either operand of the select is a constant, we can fold the
2973 // comparison into the select arms, which will cause one to be
2974 // constant folded and the select turned into a bitwise or.
2975 Value *Op1 = 0, *Op2 = 0;
2976 if (LHSI->hasOneUse()) {
2977 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2978 // Fold the known value into the constant operand.
2979 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2980 // Insert a new FCmp of the other select operand.
2981 Op2 = Builder->CreateFCmp(I.getPredicate(),
2982 LHSI->getOperand(2), RHSC, I.getName());
2983 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2984 // Fold the known value into the constant operand.
2985 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2986 // Insert a new FCmp of the other select operand.
2987 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2993 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2996 case Instruction::FSub: {
2997 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2999 if (match(LHSI, m_FNeg(m_Value(Op))))
3000 return new FCmpInst(I.getSwappedPredicate(), Op,
3001 ConstantExpr::getFNeg(RHSC));
3004 case Instruction::Load:
3005 if (GetElementPtrInst *GEP =
3006 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3007 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3008 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3009 !cast<LoadInst>(LHSI)->isVolatile())
3010 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3014 case Instruction::Call: {
3015 CallInst *CI = cast<CallInst>(LHSI);
3017 // Various optimization for fabs compared with zero.
3018 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3019 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3021 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3022 Func == LibFunc::fabsl) {
3023 switch (I.getPredicate()) {
3025 // fabs(x) < 0 --> false
3026 case FCmpInst::FCMP_OLT:
3027 return ReplaceInstUsesWith(I, Builder->getFalse());
3028 // fabs(x) > 0 --> x != 0
3029 case FCmpInst::FCMP_OGT:
3030 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3032 // fabs(x) <= 0 --> x == 0
3033 case FCmpInst::FCMP_OLE:
3034 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3036 // fabs(x) >= 0 --> !isnan(x)
3037 case FCmpInst::FCMP_OGE:
3038 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3040 // fabs(x) == 0 --> x == 0
3041 // fabs(x) != 0 --> x != 0
3042 case FCmpInst::FCMP_OEQ:
3043 case FCmpInst::FCMP_UEQ:
3044 case FCmpInst::FCMP_ONE:
3045 case FCmpInst::FCMP_UNE:
3046 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3055 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3057 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3058 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3060 // fcmp (fpext x), (fpext y) -> fcmp x, y
3061 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3062 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3063 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3064 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3065 RHSExt->getOperand(0));
3067 return Changed ? &I : 0;