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/IntrinsicInst.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/Target/TargetData.h"
19 #include "llvm/Support/ConstantRange.h"
20 #include "llvm/Support/GetElementPtrTypeIterator.h"
21 #include "llvm/Support/PatternMatch.h"
23 using namespace PatternMatch;
25 static ConstantInt *getOne(Constant *C) {
26 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
29 /// AddOne - Add one to a ConstantInt
30 static Constant *AddOne(Constant *C) {
31 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
33 /// SubOne - Subtract one from a ConstantInt
34 static Constant *SubOne(Constant *C) {
35 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
38 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
39 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
42 static bool HasAddOverflow(ConstantInt *Result,
43 ConstantInt *In1, ConstantInt *In2,
46 if (In2->getValue().isNegative())
47 return Result->getValue().sgt(In1->getValue());
49 return Result->getValue().slt(In1->getValue());
51 return Result->getValue().ult(In1->getValue());
54 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
55 /// overflowed for this type.
56 static bool AddWithOverflow(Constant *&Result, Constant *In1,
57 Constant *In2, bool IsSigned = false) {
58 Result = ConstantExpr::getAdd(In1, In2);
60 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
61 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
62 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
63 if (HasAddOverflow(ExtractElement(Result, Idx),
64 ExtractElement(In1, Idx),
65 ExtractElement(In2, Idx),
72 return HasAddOverflow(cast<ConstantInt>(Result),
73 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
77 static bool HasSubOverflow(ConstantInt *Result,
78 ConstantInt *In1, ConstantInt *In2,
81 if (In2->getValue().isNegative())
82 return Result->getValue().slt(In1->getValue());
84 return Result->getValue().sgt(In1->getValue());
86 return Result->getValue().ugt(In1->getValue());
89 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
90 /// overflowed for this type.
91 static bool SubWithOverflow(Constant *&Result, Constant *In1,
92 Constant *In2, bool IsSigned = false) {
93 Result = ConstantExpr::getSub(In1, In2);
95 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
96 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
97 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
98 if (HasSubOverflow(ExtractElement(Result, Idx),
99 ExtractElement(In1, Idx),
100 ExtractElement(In2, Idx),
107 return HasSubOverflow(cast<ConstantInt>(Result),
108 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
112 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
113 /// comparison only checks the sign bit. If it only checks the sign bit, set
114 /// TrueIfSigned if the result of the comparison is true when the input value is
116 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
117 bool &TrueIfSigned) {
119 case ICmpInst::ICMP_SLT: // True if LHS s< 0
121 return RHS->isZero();
122 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
124 return RHS->isAllOnesValue();
125 case ICmpInst::ICMP_SGT: // True if LHS s> -1
126 TrueIfSigned = false;
127 return RHS->isAllOnesValue();
128 case ICmpInst::ICMP_UGT:
129 // True if LHS u> RHS and RHS == high-bit-mask - 1
131 return RHS->getValue() ==
132 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
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 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer());
208 if (Init == 0 || Init->getNumOperands() > 1024) return 0;
210 // There are many forms of this optimization we can handle, for now, just do
211 // the simple index into a single-dimensional array.
213 // Require: GEP GV, 0, i {{, constant indices}}
214 if (GEP->getNumOperands() < 3 ||
215 !isa<ConstantInt>(GEP->getOperand(1)) ||
216 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
217 isa<Constant>(GEP->getOperand(2)))
220 // Check that indices after the variable are constants and in-range for the
221 // type they index. Collect the indices. This is typically for arrays of
223 SmallVector<unsigned, 4> LaterIndices;
225 const Type *EltTy = cast<ArrayType>(Init->getType())->getElementType();
226 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
227 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
228 if (Idx == 0) return 0; // Variable index.
230 uint64_t IdxVal = Idx->getZExtValue();
231 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
233 if (const StructType *STy = dyn_cast<StructType>(EltTy))
234 EltTy = STy->getElementType(IdxVal);
235 else if (const ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
236 if (IdxVal >= ATy->getNumElements()) return 0;
237 EltTy = ATy->getElementType();
239 return 0; // Unknown type.
242 LaterIndices.push_back(IdxVal);
245 enum { Overdefined = -3, Undefined = -2 };
247 // Variables for our state machines.
249 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
250 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
251 // and 87 is the second (and last) index. FirstTrueElement is -2 when
252 // undefined, otherwise set to the first true element. SecondTrueElement is
253 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
254 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
256 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
257 // form "i != 47 & i != 87". Same state transitions as for true elements.
258 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
260 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
261 /// define a state machine that triggers for ranges of values that the index
262 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
263 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
264 /// index in the range (inclusive). We use -2 for undefined here because we
265 /// use relative comparisons and don't want 0-1 to match -1.
266 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
268 // MagicBitvector - This is a magic bitvector where we set a bit if the
269 // comparison is true for element 'i'. If there are 64 elements or less in
270 // the array, this will fully represent all the comparison results.
271 uint64_t MagicBitvector = 0;
274 // Scan the array and see if one of our patterns matches.
275 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
276 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) {
277 Constant *Elt = Init->getOperand(i);
279 // If this is indexing an array of structures, get the structure element.
280 if (!LaterIndices.empty())
281 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
283 // If the element is masked, handle it.
284 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
286 // Find out if the comparison would be true or false for the i'th element.
287 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
289 // If the result is undef for this element, ignore it.
290 if (isa<UndefValue>(C)) {
291 // Extend range state machines to cover this element in case there is an
292 // undef in the middle of the range.
293 if (TrueRangeEnd == (int)i-1)
295 if (FalseRangeEnd == (int)i-1)
300 // If we can't compute the result for any of the elements, we have to give
301 // up evaluating the entire conditional.
302 if (!isa<ConstantInt>(C)) return 0;
304 // Otherwise, we know if the comparison is true or false for this element,
305 // update our state machines.
306 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
308 // State machine for single/double/range index comparison.
310 // Update the TrueElement state machine.
311 if (FirstTrueElement == Undefined)
312 FirstTrueElement = TrueRangeEnd = i; // First true element.
314 // Update double-compare state machine.
315 if (SecondTrueElement == Undefined)
316 SecondTrueElement = i;
318 SecondTrueElement = Overdefined;
320 // Update range state machine.
321 if (TrueRangeEnd == (int)i-1)
324 TrueRangeEnd = Overdefined;
327 // Update the FalseElement state machine.
328 if (FirstFalseElement == Undefined)
329 FirstFalseElement = FalseRangeEnd = i; // First false element.
331 // Update double-compare state machine.
332 if (SecondFalseElement == Undefined)
333 SecondFalseElement = i;
335 SecondFalseElement = Overdefined;
337 // Update range state machine.
338 if (FalseRangeEnd == (int)i-1)
341 FalseRangeEnd = Overdefined;
346 // If this element is in range, update our magic bitvector.
347 if (i < 64 && IsTrueForElt)
348 MagicBitvector |= 1ULL << i;
350 // If all of our states become overdefined, bail out early. Since the
351 // predicate is expensive, only check it every 8 elements. This is only
352 // really useful for really huge arrays.
353 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
354 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
355 FalseRangeEnd == Overdefined)
359 // Now that we've scanned the entire array, emit our new comparison(s). We
360 // order the state machines in complexity of the generated code.
361 Value *Idx = GEP->getOperand(2);
363 // If the index is larger than the pointer size of the target, truncate the
364 // index down like the GEP would do implicitly. We don't have to do this for
365 // an inbounds GEP because the index can't be out of range.
366 if (!GEP->isInBounds() &&
367 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
368 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
370 // If the comparison is only true for one or two elements, emit direct
372 if (SecondTrueElement != Overdefined) {
373 // None true -> false.
374 if (FirstTrueElement == Undefined)
375 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
377 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
379 // True for one element -> 'i == 47'.
380 if (SecondTrueElement == Undefined)
381 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
383 // True for two elements -> 'i == 47 | i == 72'.
384 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
385 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
386 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
387 return BinaryOperator::CreateOr(C1, C2);
390 // If the comparison is only false for one or two elements, emit direct
392 if (SecondFalseElement != Overdefined) {
393 // None false -> true.
394 if (FirstFalseElement == Undefined)
395 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
397 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
399 // False for one element -> 'i != 47'.
400 if (SecondFalseElement == Undefined)
401 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
403 // False for two elements -> 'i != 47 & i != 72'.
404 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
405 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
406 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
407 return BinaryOperator::CreateAnd(C1, C2);
410 // If the comparison can be replaced with a range comparison for the elements
411 // where it is true, emit the range check.
412 if (TrueRangeEnd != Overdefined) {
413 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
415 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
416 if (FirstTrueElement) {
417 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
418 Idx = Builder->CreateAdd(Idx, Offs);
421 Value *End = ConstantInt::get(Idx->getType(),
422 TrueRangeEnd-FirstTrueElement+1);
423 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
426 // False range check.
427 if (FalseRangeEnd != Overdefined) {
428 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
429 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
430 if (FirstFalseElement) {
431 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
432 Idx = Builder->CreateAdd(Idx, Offs);
435 Value *End = ConstantInt::get(Idx->getType(),
436 FalseRangeEnd-FirstFalseElement);
437 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
441 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
442 // of this load, replace it with computation that does:
443 // ((magic_cst >> i) & 1) != 0
444 if (Init->getNumOperands() <= 32 ||
445 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) {
447 if (Init->getNumOperands() <= 32)
448 Ty = Type::getInt32Ty(Init->getContext());
450 Ty = Type::getInt64Ty(Init->getContext());
451 Value *V = Builder->CreateIntCast(Idx, Ty, false);
452 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
453 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
454 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
461 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
462 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
463 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
464 /// be complex, and scales are involved. The above expression would also be
465 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
466 /// This later form is less amenable to optimization though, and we are allowed
467 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
469 /// If we can't emit an optimized form for this expression, this returns null.
471 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
472 TargetData &TD = *IC.getTargetData();
473 gep_type_iterator GTI = gep_type_begin(GEP);
475 // Check to see if this gep only has a single variable index. If so, and if
476 // any constant indices are a multiple of its scale, then we can compute this
477 // in terms of the scale of the variable index. For example, if the GEP
478 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
479 // because the expression will cross zero at the same point.
480 unsigned i, e = GEP->getNumOperands();
482 for (i = 1; i != e; ++i, ++GTI) {
483 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
484 // Compute the aggregate offset of constant indices.
485 if (CI->isZero()) continue;
487 // Handle a struct index, which adds its field offset to the pointer.
488 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
489 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
491 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
492 Offset += Size*CI->getSExtValue();
495 // Found our variable index.
500 // If there are no variable indices, we must have a constant offset, just
501 // evaluate it the general way.
502 if (i == e) return 0;
504 Value *VariableIdx = GEP->getOperand(i);
505 // Determine the scale factor of the variable element. For example, this is
506 // 4 if the variable index is into an array of i32.
507 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
509 // Verify that there are no other variable indices. If so, emit the hard way.
510 for (++i, ++GTI; i != e; ++i, ++GTI) {
511 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
514 // Compute the aggregate offset of constant indices.
515 if (CI->isZero()) continue;
517 // Handle a struct index, which adds its field offset to the pointer.
518 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
519 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
521 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
522 Offset += Size*CI->getSExtValue();
526 // Okay, we know we have a single variable index, which must be a
527 // pointer/array/vector index. If there is no offset, life is simple, return
529 unsigned IntPtrWidth = TD.getPointerSizeInBits();
531 // Cast to intptrty in case a truncation occurs. If an extension is needed,
532 // we don't need to bother extending: the extension won't affect where the
533 // computation crosses zero.
534 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
535 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
536 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
541 // Otherwise, there is an index. The computation we will do will be modulo
542 // the pointer size, so get it.
543 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
545 Offset &= PtrSizeMask;
546 VariableScale &= PtrSizeMask;
548 // To do this transformation, any constant index must be a multiple of the
549 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
550 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
551 // multiple of the variable scale.
552 int64_t NewOffs = Offset / (int64_t)VariableScale;
553 if (Offset != NewOffs*(int64_t)VariableScale)
556 // Okay, we can do this evaluation. Start by converting the index to intptr.
557 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
558 if (VariableIdx->getType() != IntPtrTy)
559 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
561 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
562 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
565 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
566 /// else. At this point we know that the GEP is on the LHS of the comparison.
567 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
568 ICmpInst::Predicate Cond,
570 // Look through bitcasts.
571 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
572 RHS = BCI->getOperand(0);
574 Value *PtrBase = GEPLHS->getOperand(0);
575 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
576 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
577 // This transformation (ignoring the base and scales) is valid because we
578 // know pointers can't overflow since the gep is inbounds. See if we can
579 // output an optimized form.
580 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
582 // If not, synthesize the offset the hard way.
584 Offset = EmitGEPOffset(GEPLHS);
585 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
586 Constant::getNullValue(Offset->getType()));
587 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
588 // If the base pointers are different, but the indices are the same, just
589 // compare the base pointer.
590 if (PtrBase != GEPRHS->getOperand(0)) {
591 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
592 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
593 GEPRHS->getOperand(0)->getType();
595 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
596 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
597 IndicesTheSame = false;
601 // If all indices are the same, just compare the base pointers.
603 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
604 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
606 // Otherwise, the base pointers are different and the indices are
607 // different, bail out.
611 // If one of the GEPs has all zero indices, recurse.
612 bool AllZeros = true;
613 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
614 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
615 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
620 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
621 ICmpInst::getSwappedPredicate(Cond), I);
623 // If the other GEP has all zero indices, recurse.
625 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
626 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
627 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
632 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
634 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
635 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
636 // If the GEPs only differ by one index, compare it.
637 unsigned NumDifferences = 0; // Keep track of # differences.
638 unsigned DiffOperand = 0; // The operand that differs.
639 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
640 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
641 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
642 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
643 // Irreconcilable differences.
647 if (NumDifferences++) break;
652 if (NumDifferences == 0) // SAME GEP?
653 return ReplaceInstUsesWith(I, // No comparison is needed here.
654 ConstantInt::get(Type::getInt1Ty(I.getContext()),
655 ICmpInst::isTrueWhenEqual(Cond)));
657 else if (NumDifferences == 1 && GEPsInBounds) {
658 Value *LHSV = GEPLHS->getOperand(DiffOperand);
659 Value *RHSV = GEPRHS->getOperand(DiffOperand);
660 // Make sure we do a signed comparison here.
661 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
665 // Only lower this if the icmp is the only user of the GEP or if we expect
666 // the result to fold to a constant!
669 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
670 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
671 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
672 Value *L = EmitGEPOffset(GEPLHS);
673 Value *R = EmitGEPOffset(GEPRHS);
674 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
680 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
681 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
682 Value *X, ConstantInt *CI,
683 ICmpInst::Predicate Pred,
685 // If we have X+0, exit early (simplifying logic below) and let it get folded
686 // elsewhere. icmp X+0, X -> icmp X, X
688 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
689 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
692 // (X+4) == X -> false.
693 if (Pred == ICmpInst::ICMP_EQ)
694 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
696 // (X+4) != X -> true.
697 if (Pred == ICmpInst::ICMP_NE)
698 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
700 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
701 // so the values can never be equal. Similarly for all other "or equals"
704 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
705 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
706 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
707 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
709 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
710 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
713 // (X+1) >u X --> X <u (0-1) --> X != 255
714 // (X+2) >u X --> X <u (0-2) --> X <u 254
715 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
716 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
717 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
719 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
720 ConstantInt *SMax = ConstantInt::get(X->getContext(),
721 APInt::getSignedMaxValue(BitWidth));
723 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
724 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
725 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
726 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
727 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
728 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
729 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
730 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
732 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
733 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
734 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
735 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
736 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
737 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
739 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
740 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
741 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
744 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
745 /// and CmpRHS are both known to be integer constants.
746 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
747 ConstantInt *DivRHS) {
748 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
749 const APInt &CmpRHSV = CmpRHS->getValue();
751 // FIXME: If the operand types don't match the type of the divide
752 // then don't attempt this transform. The code below doesn't have the
753 // logic to deal with a signed divide and an unsigned compare (and
754 // vice versa). This is because (x /s C1) <s C2 produces different
755 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
756 // (x /u C1) <u C2. Simply casting the operands and result won't
757 // work. :( The if statement below tests that condition and bails
759 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
760 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
762 if (DivRHS->isZero())
763 return 0; // The ProdOV computation fails on divide by zero.
764 if (DivIsSigned && DivRHS->isAllOnesValue())
765 return 0; // The overflow computation also screws up here
766 if (DivRHS->isOne()) {
767 // This eliminates some funny cases with INT_MIN.
768 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
772 // Compute Prod = CI * DivRHS. We are essentially solving an equation
773 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
774 // C2 (CI). By solving for X we can turn this into a range check
775 // instead of computing a divide.
776 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
778 // Determine if the product overflows by seeing if the product is
779 // not equal to the divide. Make sure we do the same kind of divide
780 // as in the LHS instruction that we're folding.
781 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
782 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
784 // Get the ICmp opcode
785 ICmpInst::Predicate Pred = ICI.getPredicate();
787 /// If the division is known to be exact, then there is no remainder from the
788 /// divide, so the covered range size is unit, otherwise it is the divisor.
789 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
791 // Figure out the interval that is being checked. For example, a comparison
792 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
793 // Compute this interval based on the constants involved and the signedness of
794 // the compare/divide. This computes a half-open interval, keeping track of
795 // whether either value in the interval overflows. After analysis each
796 // overflow variable is set to 0 if it's corresponding bound variable is valid
797 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
798 int LoOverflow = 0, HiOverflow = 0;
799 Constant *LoBound = 0, *HiBound = 0;
801 if (!DivIsSigned) { // udiv
802 // e.g. X/5 op 3 --> [15, 20)
804 HiOverflow = LoOverflow = ProdOV;
806 // If this is not an exact divide, then many values in the range collapse
807 // to the same result value.
808 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
811 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
812 if (CmpRHSV == 0) { // (X / pos) op 0
813 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
814 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
816 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
817 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
818 HiOverflow = LoOverflow = ProdOV;
820 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
821 } else { // (X / pos) op neg
822 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
823 HiBound = AddOne(Prod);
824 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
826 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
827 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
830 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
832 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
833 if (CmpRHSV == 0) { // (X / neg) op 0
834 // e.g. X/-5 op 0 --> [-4, 5)
835 LoBound = AddOne(RangeSize);
836 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
837 if (HiBound == DivRHS) { // -INTMIN = INTMIN
838 HiOverflow = 1; // [INTMIN+1, overflow)
839 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
841 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
842 // e.g. X/-5 op 3 --> [-19, -14)
843 HiBound = AddOne(Prod);
844 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
846 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
847 } else { // (X / neg) op neg
848 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
849 LoOverflow = HiOverflow = ProdOV;
851 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
854 // Dividing by a negative swaps the condition. LT <-> GT
855 Pred = ICmpInst::getSwappedPredicate(Pred);
858 Value *X = DivI->getOperand(0);
860 default: llvm_unreachable("Unhandled icmp opcode!");
861 case ICmpInst::ICMP_EQ:
862 if (LoOverflow && HiOverflow)
863 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
865 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
866 ICmpInst::ICMP_UGE, X, LoBound);
868 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
869 ICmpInst::ICMP_ULT, X, HiBound);
870 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
872 case ICmpInst::ICMP_NE:
873 if (LoOverflow && HiOverflow)
874 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
876 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
877 ICmpInst::ICMP_ULT, X, LoBound);
879 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
880 ICmpInst::ICMP_UGE, X, HiBound);
881 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
882 DivIsSigned, false));
883 case ICmpInst::ICMP_ULT:
884 case ICmpInst::ICMP_SLT:
885 if (LoOverflow == +1) // Low bound is greater than input range.
886 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
887 if (LoOverflow == -1) // Low bound is less than input range.
888 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
889 return new ICmpInst(Pred, X, LoBound);
890 case ICmpInst::ICMP_UGT:
891 case ICmpInst::ICMP_SGT:
892 if (HiOverflow == +1) // High bound greater than input range.
893 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
894 if (HiOverflow == -1) // High bound less than input range.
895 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
896 if (Pred == ICmpInst::ICMP_UGT)
897 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
898 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
902 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
903 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
904 ConstantInt *ShAmt) {
905 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
907 // Check that the shift amount is in range. If not, don't perform
908 // undefined shifts. When the shift is visited it will be
910 uint32_t TypeBits = CmpRHSV.getBitWidth();
911 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
912 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
915 if (!ICI.isEquality()) {
916 // If we have an unsigned comparison and an ashr, we can't simplify this.
917 // Similarly for signed comparisons with lshr.
918 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
921 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
922 // by a power of 2. Since we already have logic to simplify these,
923 // transform to div and then simplify the resultant comparison.
924 if (Shr->getOpcode() == Instruction::AShr &&
925 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
928 // Revisit the shift (to delete it).
932 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
935 Shr->getOpcode() == Instruction::AShr ?
936 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
937 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
939 ICI.setOperand(0, Tmp);
941 // If the builder folded the binop, just return it.
942 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
946 // Otherwise, fold this div/compare.
947 assert(TheDiv->getOpcode() == Instruction::SDiv ||
948 TheDiv->getOpcode() == Instruction::UDiv);
950 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
951 assert(Res && "This div/cst should have folded!");
956 // If we are comparing against bits always shifted out, the
957 // comparison cannot succeed.
958 APInt Comp = CmpRHSV << ShAmtVal;
959 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
960 if (Shr->getOpcode() == Instruction::LShr)
961 Comp = Comp.lshr(ShAmtVal);
963 Comp = Comp.ashr(ShAmtVal);
965 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
966 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
967 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
969 return ReplaceInstUsesWith(ICI, Cst);
972 // Otherwise, check to see if the bits shifted out are known to be zero.
973 // If so, we can compare against the unshifted value:
974 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
975 if (Shr->hasOneUse() && Shr->isExact())
976 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
978 if (Shr->hasOneUse()) {
979 // Otherwise strength reduce the shift into an and.
980 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
981 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
983 Value *And = Builder->CreateAnd(Shr->getOperand(0),
984 Mask, Shr->getName()+".mask");
985 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
991 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
993 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
996 const APInt &RHSV = RHS->getValue();
998 switch (LHSI->getOpcode()) {
999 case Instruction::Trunc:
1000 if (ICI.isEquality() && LHSI->hasOneUse()) {
1001 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1002 // of the high bits truncated out of x are known.
1003 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1004 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1005 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
1006 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1007 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
1009 // If all the high bits are known, we can do this xform.
1010 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1011 // Pull in the high bits from known-ones set.
1012 APInt NewRHS = RHS->getValue().zext(SrcBits);
1014 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1015 ConstantInt::get(ICI.getContext(), NewRHS));
1020 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1021 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1022 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1024 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1025 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1026 Value *CompareVal = LHSI->getOperand(0);
1028 // If the sign bit of the XorCST is not set, there is no change to
1029 // the operation, just stop using the Xor.
1030 if (!XorCST->getValue().isNegative()) {
1031 ICI.setOperand(0, CompareVal);
1036 // Was the old condition true if the operand is positive?
1037 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1039 // If so, the new one isn't.
1040 isTrueIfPositive ^= true;
1042 if (isTrueIfPositive)
1043 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1046 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1050 if (LHSI->hasOneUse()) {
1051 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1052 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1053 const APInt &SignBit = XorCST->getValue();
1054 ICmpInst::Predicate Pred = ICI.isSigned()
1055 ? ICI.getUnsignedPredicate()
1056 : ICI.getSignedPredicate();
1057 return new ICmpInst(Pred, LHSI->getOperand(0),
1058 ConstantInt::get(ICI.getContext(),
1062 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1063 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
1064 const APInt &NotSignBit = XorCST->getValue();
1065 ICmpInst::Predicate Pred = ICI.isSigned()
1066 ? ICI.getUnsignedPredicate()
1067 : ICI.getSignedPredicate();
1068 Pred = ICI.getSwappedPredicate(Pred);
1069 return new ICmpInst(Pred, LHSI->getOperand(0),
1070 ConstantInt::get(ICI.getContext(),
1071 RHSV ^ NotSignBit));
1076 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1077 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1078 LHSI->getOperand(0)->hasOneUse()) {
1079 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1081 // If the LHS is an AND of a truncating cast, we can widen the
1082 // and/compare to be the input width without changing the value
1083 // produced, eliminating a cast.
1084 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1085 // We can do this transformation if either the AND constant does not
1086 // have its sign bit set or if it is an equality comparison.
1087 // Extending a relational comparison when we're checking the sign
1088 // bit would not work.
1089 if (ICI.isEquality() ||
1090 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative())) {
1092 Builder->CreateAnd(Cast->getOperand(0),
1093 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1094 NewAnd->takeName(LHSI);
1095 return new ICmpInst(ICI.getPredicate(), NewAnd,
1096 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1100 // If the LHS is an AND of a zext, and we have an equality compare, we can
1101 // shrink the and/compare to the smaller type, eliminating the cast.
1102 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1103 const IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1104 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1105 // should fold the icmp to true/false in that case.
1106 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1108 Builder->CreateAnd(Cast->getOperand(0),
1109 ConstantExpr::getTrunc(AndCST, Ty));
1110 NewAnd->takeName(LHSI);
1111 return new ICmpInst(ICI.getPredicate(), NewAnd,
1112 ConstantExpr::getTrunc(RHS, Ty));
1116 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1117 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1118 // happens a LOT in code produced by the C front-end, for bitfield
1120 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1121 if (Shift && !Shift->isShift())
1125 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1126 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1127 const Type *AndTy = AndCST->getType(); // Type of the and.
1129 // We can fold this as long as we can't shift unknown bits
1130 // into the mask. This can only happen with signed shift
1131 // rights, as they sign-extend.
1133 bool CanFold = Shift->isLogicalShift();
1135 // To test for the bad case of the signed shr, see if any
1136 // of the bits shifted in could be tested after the mask.
1137 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1138 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1140 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1141 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1142 AndCST->getValue()) == 0)
1148 if (Shift->getOpcode() == Instruction::Shl)
1149 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1151 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1153 // Check to see if we are shifting out any of the bits being
1155 if (ConstantExpr::get(Shift->getOpcode(),
1156 NewCst, ShAmt) != RHS) {
1157 // If we shifted bits out, the fold is not going to work out.
1158 // As a special case, check to see if this means that the
1159 // result is always true or false now.
1160 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1161 return ReplaceInstUsesWith(ICI,
1162 ConstantInt::getFalse(ICI.getContext()));
1163 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1164 return ReplaceInstUsesWith(ICI,
1165 ConstantInt::getTrue(ICI.getContext()));
1167 ICI.setOperand(1, NewCst);
1168 Constant *NewAndCST;
1169 if (Shift->getOpcode() == Instruction::Shl)
1170 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1172 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1173 LHSI->setOperand(1, NewAndCST);
1174 LHSI->setOperand(0, Shift->getOperand(0));
1175 Worklist.Add(Shift); // Shift is dead.
1181 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1182 // preferable because it allows the C<<Y expression to be hoisted out
1183 // of a loop if Y is invariant and X is not.
1184 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1185 ICI.isEquality() && !Shift->isArithmeticShift() &&
1186 !isa<Constant>(Shift->getOperand(0))) {
1189 if (Shift->getOpcode() == Instruction::LShr) {
1190 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
1192 // Insert a logical shift.
1193 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
1196 // Compute X & (C << Y).
1198 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1200 ICI.setOperand(0, NewAnd);
1205 // Try to optimize things like "A[i]&42 == 0" to index computations.
1206 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1207 if (GetElementPtrInst *GEP =
1208 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1209 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1210 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1211 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1212 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1213 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1219 case Instruction::Or: {
1220 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1223 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1224 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1225 // -> and (icmp eq P, null), (icmp eq Q, null).
1226 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1227 Constant::getNullValue(P->getType()));
1228 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1229 Constant::getNullValue(Q->getType()));
1231 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1232 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1234 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1240 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1241 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1244 uint32_t TypeBits = RHSV.getBitWidth();
1246 // Check that the shift amount is in range. If not, don't perform
1247 // undefined shifts. When the shift is visited it will be
1249 if (ShAmt->uge(TypeBits))
1252 if (ICI.isEquality()) {
1253 // If we are comparing against bits always shifted out, the
1254 // comparison cannot succeed.
1256 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1258 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1259 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1261 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1262 return ReplaceInstUsesWith(ICI, Cst);
1265 // If the shift is NUW, then it is just shifting out zeros, no need for an
1267 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1268 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1269 ConstantExpr::getLShr(RHS, ShAmt));
1271 if (LHSI->hasOneUse()) {
1272 // Otherwise strength reduce the shift into an and.
1273 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1275 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1276 TypeBits-ShAmtVal));
1279 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1280 return new ICmpInst(ICI.getPredicate(), And,
1281 ConstantExpr::getLShr(RHS, ShAmt));
1285 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1286 bool TrueIfSigned = false;
1287 if (LHSI->hasOneUse() &&
1288 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1289 // (X << 31) <s 0 --> (X&1) != 0
1290 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1291 APInt::getOneBitSet(TypeBits,
1292 TypeBits-ShAmt->getZExtValue()-1));
1294 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1295 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1296 And, Constant::getNullValue(And->getType()));
1301 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1302 case Instruction::AShr: {
1303 // Handle equality comparisons of shift-by-constant.
1304 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1305 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1306 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1310 // Handle exact shr's.
1311 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1312 if (RHSV.isMinValue())
1313 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1318 case Instruction::SDiv:
1319 case Instruction::UDiv:
1320 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1321 // Fold this div into the comparison, producing a range check.
1322 // Determine, based on the divide type, what the range is being
1323 // checked. If there is an overflow on the low or high side, remember
1324 // it, otherwise compute the range [low, hi) bounding the new value.
1325 // See: InsertRangeTest above for the kinds of replacements possible.
1326 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1327 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1332 case Instruction::Add:
1333 // Fold: icmp pred (add X, C1), C2
1334 if (!ICI.isEquality()) {
1335 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1337 const APInt &LHSV = LHSC->getValue();
1339 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1342 if (ICI.isSigned()) {
1343 if (CR.getLower().isSignBit()) {
1344 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1345 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1346 } else if (CR.getUpper().isSignBit()) {
1347 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1348 ConstantInt::get(ICI.getContext(),CR.getLower()));
1351 if (CR.getLower().isMinValue()) {
1352 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1353 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1354 } else if (CR.getUpper().isMinValue()) {
1355 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1356 ConstantInt::get(ICI.getContext(),CR.getLower()));
1363 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1364 if (ICI.isEquality()) {
1365 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1367 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1368 // the second operand is a constant, simplify a bit.
1369 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1370 switch (BO->getOpcode()) {
1371 case Instruction::SRem:
1372 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1373 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1374 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1375 if (V.sgt(1) && V.isPowerOf2()) {
1377 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1379 return new ICmpInst(ICI.getPredicate(), NewRem,
1380 Constant::getNullValue(BO->getType()));
1384 case Instruction::Add:
1385 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1386 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1387 if (BO->hasOneUse())
1388 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1389 ConstantExpr::getSub(RHS, BOp1C));
1390 } else if (RHSV == 0) {
1391 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1392 // efficiently invertible, or if the add has just this one use.
1393 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1395 if (Value *NegVal = dyn_castNegVal(BOp1))
1396 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1397 if (Value *NegVal = dyn_castNegVal(BOp0))
1398 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1399 if (BO->hasOneUse()) {
1400 Value *Neg = Builder->CreateNeg(BOp1);
1402 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1406 case Instruction::Xor:
1407 // For the xor case, we can xor two constants together, eliminating
1408 // the explicit xor.
1409 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1410 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1411 ConstantExpr::getXor(RHS, BOC));
1412 } else if (RHSV == 0) {
1413 // Replace ((xor A, B) != 0) with (A != B)
1414 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1418 case Instruction::Sub:
1419 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1420 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1421 if (BO->hasOneUse())
1422 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1423 ConstantExpr::getSub(BOp0C, RHS));
1424 } else if (RHSV == 0) {
1425 // Replace ((sub A, B) != 0) with (A != B)
1426 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1430 case Instruction::Or:
1431 // If bits are being or'd in that are not present in the constant we
1432 // are comparing against, then the comparison could never succeed!
1433 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1434 Constant *NotCI = ConstantExpr::getNot(RHS);
1435 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1436 return ReplaceInstUsesWith(ICI,
1437 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1442 case Instruction::And:
1443 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1444 // If bits are being compared against that are and'd out, then the
1445 // comparison can never succeed!
1446 if ((RHSV & ~BOC->getValue()) != 0)
1447 return ReplaceInstUsesWith(ICI,
1448 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1451 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1452 if (RHS == BOC && RHSV.isPowerOf2())
1453 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1454 ICmpInst::ICMP_NE, LHSI,
1455 Constant::getNullValue(RHS->getType()));
1457 // Don't perform the following transforms if the AND has multiple uses
1458 if (!BO->hasOneUse())
1461 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1462 if (BOC->getValue().isSignBit()) {
1463 Value *X = BO->getOperand(0);
1464 Constant *Zero = Constant::getNullValue(X->getType());
1465 ICmpInst::Predicate pred = isICMP_NE ?
1466 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1467 return new ICmpInst(pred, X, Zero);
1470 // ((X & ~7) == 0) --> X < 8
1471 if (RHSV == 0 && isHighOnes(BOC)) {
1472 Value *X = BO->getOperand(0);
1473 Constant *NegX = ConstantExpr::getNeg(BOC);
1474 ICmpInst::Predicate pred = isICMP_NE ?
1475 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1476 return new ICmpInst(pred, X, NegX);
1481 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1482 // Handle icmp {eq|ne} <intrinsic>, intcst.
1483 switch (II->getIntrinsicID()) {
1484 case Intrinsic::bswap:
1486 ICI.setOperand(0, II->getArgOperand(0));
1487 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1489 case Intrinsic::ctlz:
1490 case Intrinsic::cttz:
1491 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1492 if (RHSV == RHS->getType()->getBitWidth()) {
1494 ICI.setOperand(0, II->getArgOperand(0));
1495 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1499 case Intrinsic::ctpop:
1500 // popcount(A) == 0 -> A == 0 and likewise for !=
1501 if (RHS->isZero()) {
1503 ICI.setOperand(0, II->getArgOperand(0));
1504 ICI.setOperand(1, RHS);
1516 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1517 /// We only handle extending casts so far.
1519 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1520 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1521 Value *LHSCIOp = LHSCI->getOperand(0);
1522 const Type *SrcTy = LHSCIOp->getType();
1523 const Type *DestTy = LHSCI->getType();
1526 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1527 // integer type is the same size as the pointer type.
1528 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1529 TD->getPointerSizeInBits() ==
1530 cast<IntegerType>(DestTy)->getBitWidth()) {
1532 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1533 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1534 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1535 RHSOp = RHSC->getOperand(0);
1536 // If the pointer types don't match, insert a bitcast.
1537 if (LHSCIOp->getType() != RHSOp->getType())
1538 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1542 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1545 // The code below only handles extension cast instructions, so far.
1547 if (LHSCI->getOpcode() != Instruction::ZExt &&
1548 LHSCI->getOpcode() != Instruction::SExt)
1551 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1552 bool isSignedCmp = ICI.isSigned();
1554 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1555 // Not an extension from the same type?
1556 RHSCIOp = CI->getOperand(0);
1557 if (RHSCIOp->getType() != LHSCIOp->getType())
1560 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1561 // and the other is a zext), then we can't handle this.
1562 if (CI->getOpcode() != LHSCI->getOpcode())
1565 // Deal with equality cases early.
1566 if (ICI.isEquality())
1567 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1569 // A signed comparison of sign extended values simplifies into a
1570 // signed comparison.
1571 if (isSignedCmp && isSignedExt)
1572 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1574 // The other three cases all fold into an unsigned comparison.
1575 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1578 // If we aren't dealing with a constant on the RHS, exit early
1579 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1583 // Compute the constant that would happen if we truncated to SrcTy then
1584 // reextended to DestTy.
1585 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1586 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1589 // If the re-extended constant didn't change...
1591 // Deal with equality cases early.
1592 if (ICI.isEquality())
1593 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1595 // A signed comparison of sign extended values simplifies into a
1596 // signed comparison.
1597 if (isSignedExt && isSignedCmp)
1598 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1600 // The other three cases all fold into an unsigned comparison.
1601 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1604 // The re-extended constant changed so the constant cannot be represented
1605 // in the shorter type. Consequently, we cannot emit a simple comparison.
1606 // All the cases that fold to true or false will have already been handled
1607 // by SimplifyICmpInst, so only deal with the tricky case.
1609 if (isSignedCmp || !isSignedExt)
1612 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1613 // should have been folded away previously and not enter in here.
1615 // We're performing an unsigned comp with a sign extended value.
1616 // This is true if the input is >= 0. [aka >s -1]
1617 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1618 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1620 // Finally, return the value computed.
1621 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1622 return ReplaceInstUsesWith(ICI, Result);
1624 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1625 return BinaryOperator::CreateNot(Result);
1628 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1629 /// I = icmp ugt (add (add A, B), CI2), CI1
1630 /// If this is of the form:
1632 /// if (sum+128 >u 255)
1633 /// Then replace it with llvm.sadd.with.overflow.i8.
1635 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1636 ConstantInt *CI2, ConstantInt *CI1,
1638 // The transformation we're trying to do here is to transform this into an
1639 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1640 // with a narrower add, and discard the add-with-constant that is part of the
1641 // range check (if we can't eliminate it, this isn't profitable).
1643 // In order to eliminate the add-with-constant, the compare can be its only
1645 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1646 if (!AddWithCst->hasOneUse()) return 0;
1648 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1649 if (!CI2->getValue().isPowerOf2()) return 0;
1650 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1651 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1653 // The width of the new add formed is 1 more than the bias.
1656 // Check to see that CI1 is an all-ones value with NewWidth bits.
1657 if (CI1->getBitWidth() == NewWidth ||
1658 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1661 // In order to replace the original add with a narrower
1662 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1663 // and truncates that discard the high bits of the add. Verify that this is
1665 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1666 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1668 if (*UI == AddWithCst) continue;
1670 // Only accept truncates for now. We would really like a nice recursive
1671 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1672 // chain to see which bits of a value are actually demanded. If the
1673 // original add had another add which was then immediately truncated, we
1674 // could still do the transformation.
1675 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1677 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1680 // If the pattern matches, truncate the inputs to the narrower type and
1681 // use the sadd_with_overflow intrinsic to efficiently compute both the
1682 // result and the overflow bit.
1683 Module *M = I.getParent()->getParent()->getParent();
1685 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1686 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1689 InstCombiner::BuilderTy *Builder = IC.Builder;
1691 // Put the new code above the original add, in case there are any uses of the
1692 // add between the add and the compare.
1693 Builder->SetInsertPoint(OrigAdd);
1695 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1696 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1697 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1698 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1699 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1701 // The inner add was the result of the narrow add, zero extended to the
1702 // wider type. Replace it with the result computed by the intrinsic.
1703 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1705 // The original icmp gets replaced with the overflow value.
1706 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1709 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1711 // Don't bother doing this transformation for pointers, don't do it for
1713 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1715 // If the add is a constant expr, then we don't bother transforming it.
1716 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1717 if (OrigAdd == 0) return 0;
1719 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1721 // Put the new code above the original add, in case there are any uses of the
1722 // add between the add and the compare.
1723 InstCombiner::BuilderTy *Builder = IC.Builder;
1724 Builder->SetInsertPoint(OrigAdd);
1726 Module *M = I.getParent()->getParent()->getParent();
1727 Type *Ty = LHS->getType();
1728 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1729 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1730 Value *Add = Builder->CreateExtractValue(Call, 0);
1732 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1734 // The original icmp gets replaced with the overflow value.
1735 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1738 // DemandedBitsLHSMask - When performing a comparison against a constant,
1739 // it is possible that not all the bits in the LHS are demanded. This helper
1740 // method computes the mask that IS demanded.
1741 static APInt DemandedBitsLHSMask(ICmpInst &I,
1742 unsigned BitWidth, bool isSignCheck) {
1744 return APInt::getSignBit(BitWidth);
1746 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1747 if (!CI) return APInt::getAllOnesValue(BitWidth);
1748 const APInt &RHS = CI->getValue();
1750 switch (I.getPredicate()) {
1751 // For a UGT comparison, we don't care about any bits that
1752 // correspond to the trailing ones of the comparand. The value of these
1753 // bits doesn't impact the outcome of the comparison, because any value
1754 // greater than the RHS must differ in a bit higher than these due to carry.
1755 case ICmpInst::ICMP_UGT: {
1756 unsigned trailingOnes = RHS.countTrailingOnes();
1757 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1761 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1762 // Any value less than the RHS must differ in a higher bit because of carries.
1763 case ICmpInst::ICMP_ULT: {
1764 unsigned trailingZeros = RHS.countTrailingZeros();
1765 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1770 return APInt::getAllOnesValue(BitWidth);
1775 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1776 bool Changed = false;
1777 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1779 /// Orders the operands of the compare so that they are listed from most
1780 /// complex to least complex. This puts constants before unary operators,
1781 /// before binary operators.
1782 if (getComplexity(Op0) < getComplexity(Op1)) {
1784 std::swap(Op0, Op1);
1788 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1789 return ReplaceInstUsesWith(I, V);
1791 const Type *Ty = Op0->getType();
1793 // icmp's with boolean values can always be turned into bitwise operations
1794 if (Ty->isIntegerTy(1)) {
1795 switch (I.getPredicate()) {
1796 default: llvm_unreachable("Invalid icmp instruction!");
1797 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1798 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1799 return BinaryOperator::CreateNot(Xor);
1801 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1802 return BinaryOperator::CreateXor(Op0, Op1);
1804 case ICmpInst::ICMP_UGT:
1805 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1807 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1808 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1809 return BinaryOperator::CreateAnd(Not, Op1);
1811 case ICmpInst::ICMP_SGT:
1812 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1814 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1815 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1816 return BinaryOperator::CreateAnd(Not, Op0);
1818 case ICmpInst::ICMP_UGE:
1819 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1821 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1822 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1823 return BinaryOperator::CreateOr(Not, Op1);
1825 case ICmpInst::ICMP_SGE:
1826 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1828 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1829 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1830 return BinaryOperator::CreateOr(Not, Op0);
1835 unsigned BitWidth = 0;
1836 if (Ty->isIntOrIntVectorTy())
1837 BitWidth = Ty->getScalarSizeInBits();
1838 else if (TD) // Pointers require TD info to get their size.
1839 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1841 bool isSignBit = false;
1843 // See if we are doing a comparison with a constant.
1844 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1845 Value *A = 0, *B = 0;
1847 // Match the following pattern, which is a common idiom when writing
1848 // overflow-safe integer arithmetic function. The source performs an
1849 // addition in wider type, and explicitly checks for overflow using
1850 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1851 // sadd_with_overflow intrinsic.
1853 // TODO: This could probably be generalized to handle other overflow-safe
1854 // operations if we worked out the formulas to compute the appropriate
1858 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1860 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1861 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1862 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1863 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1867 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1868 if (I.isEquality() && CI->isZero() &&
1869 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1870 // (icmp cond A B) if cond is equality
1871 return new ICmpInst(I.getPredicate(), A, B);
1874 // If we have an icmp le or icmp ge instruction, turn it into the
1875 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1876 // them being folded in the code below. The SimplifyICmpInst code has
1877 // already handled the edge cases for us, so we just assert on them.
1878 switch (I.getPredicate()) {
1880 case ICmpInst::ICMP_ULE:
1881 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1882 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1883 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1884 case ICmpInst::ICMP_SLE:
1885 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1886 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1887 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1888 case ICmpInst::ICMP_UGE:
1889 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1890 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1891 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1892 case ICmpInst::ICMP_SGE:
1893 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1894 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1895 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1898 // If this comparison is a normal comparison, it demands all
1899 // bits, if it is a sign bit comparison, it only demands the sign bit.
1901 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1904 // See if we can fold the comparison based on range information we can get
1905 // by checking whether bits are known to be zero or one in the input.
1906 if (BitWidth != 0) {
1907 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1908 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1910 if (SimplifyDemandedBits(I.getOperandUse(0),
1911 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1912 Op0KnownZero, Op0KnownOne, 0))
1914 if (SimplifyDemandedBits(I.getOperandUse(1),
1915 APInt::getAllOnesValue(BitWidth),
1916 Op1KnownZero, Op1KnownOne, 0))
1919 // Given the known and unknown bits, compute a range that the LHS could be
1920 // in. Compute the Min, Max and RHS values based on the known bits. For the
1921 // EQ and NE we use unsigned values.
1922 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1923 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1925 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1927 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1930 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1932 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1936 // If Min and Max are known to be the same, then SimplifyDemandedBits
1937 // figured out that the LHS is a constant. Just constant fold this now so
1938 // that code below can assume that Min != Max.
1939 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1940 return new ICmpInst(I.getPredicate(),
1941 ConstantInt::get(Op0->getType(), Op0Min), Op1);
1942 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1943 return new ICmpInst(I.getPredicate(), Op0,
1944 ConstantInt::get(Op1->getType(), Op1Min));
1946 // Based on the range information we know about the LHS, see if we can
1947 // simplify this comparison. For example, (x&4) < 8 is always true.
1948 switch (I.getPredicate()) {
1949 default: llvm_unreachable("Unknown icmp opcode!");
1950 case ICmpInst::ICMP_EQ: {
1951 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1952 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
1954 // If all bits are known zero except for one, then we know at most one
1955 // bit is set. If the comparison is against zero, then this is a check
1956 // to see if *that* bit is set.
1957 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1958 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1959 // If the LHS is an AND with the same constant, look through it.
1961 ConstantInt *LHSC = 0;
1962 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1963 LHSC->getValue() != Op0KnownZeroInverted)
1966 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1967 // then turn "((1 << x)&8) == 0" into "x != 3".
1969 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1970 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1971 return new ICmpInst(ICmpInst::ICMP_NE, X,
1972 ConstantInt::get(X->getType(), CmpVal));
1975 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1976 // then turn "((8 >>u x)&1) == 0" into "x != 3".
1978 if (Op0KnownZeroInverted == 1 &&
1979 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1980 return new ICmpInst(ICmpInst::ICMP_NE, X,
1981 ConstantInt::get(X->getType(),
1982 CI->countTrailingZeros()));
1987 case ICmpInst::ICMP_NE: {
1988 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1989 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
1991 // If all bits are known zero except for one, then we know at most one
1992 // bit is set. If the comparison is against zero, then this is a check
1993 // to see if *that* bit is set.
1994 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1995 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1996 // If the LHS is an AND with the same constant, look through it.
1998 ConstantInt *LHSC = 0;
1999 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2000 LHSC->getValue() != Op0KnownZeroInverted)
2003 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2004 // then turn "((1 << x)&8) != 0" into "x == 3".
2006 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2007 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2008 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2009 ConstantInt::get(X->getType(), CmpVal));
2012 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2013 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2015 if (Op0KnownZeroInverted == 1 &&
2016 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2017 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2018 ConstantInt::get(X->getType(),
2019 CI->countTrailingZeros()));
2024 case ICmpInst::ICMP_ULT:
2025 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2026 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2027 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2028 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2029 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2030 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2031 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2032 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2033 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2034 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2036 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2037 if (CI->isMinValue(true))
2038 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2039 Constant::getAllOnesValue(Op0->getType()));
2042 case ICmpInst::ICMP_UGT:
2043 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2044 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2045 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2046 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2048 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2049 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2050 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2051 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2052 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2053 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2055 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2056 if (CI->isMaxValue(true))
2057 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2058 Constant::getNullValue(Op0->getType()));
2061 case ICmpInst::ICMP_SLT:
2062 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2063 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2064 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2065 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2066 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2067 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2068 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2069 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2070 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2071 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2074 case ICmpInst::ICMP_SGT:
2075 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2076 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2077 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2078 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2080 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2081 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2082 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2083 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2084 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2085 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2088 case ICmpInst::ICMP_SGE:
2089 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2090 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2091 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2092 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2093 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2095 case ICmpInst::ICMP_SLE:
2096 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2097 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2098 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2099 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2100 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2102 case ICmpInst::ICMP_UGE:
2103 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2104 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2105 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2106 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2107 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2109 case ICmpInst::ICMP_ULE:
2110 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2111 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2112 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2113 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2114 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2118 // Turn a signed comparison into an unsigned one if both operands
2119 // are known to have the same sign.
2121 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2122 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2123 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2126 // Test if the ICmpInst instruction is used exclusively by a select as
2127 // part of a minimum or maximum operation. If so, refrain from doing
2128 // any other folding. This helps out other analyses which understand
2129 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2130 // and CodeGen. And in this case, at least one of the comparison
2131 // operands has at least one user besides the compare (the select),
2132 // which would often largely negate the benefit of folding anyway.
2134 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2135 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2136 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2139 // See if we are doing a comparison between a constant and an instruction that
2140 // can be folded into the comparison.
2141 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2142 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2143 // instruction, see if that instruction also has constants so that the
2144 // instruction can be folded into the icmp
2145 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2146 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2150 // Handle icmp with constant (but not simple integer constant) RHS
2151 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2152 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2153 switch (LHSI->getOpcode()) {
2154 case Instruction::GetElementPtr:
2155 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2156 if (RHSC->isNullValue() &&
2157 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2158 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2159 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2161 case Instruction::PHI:
2162 // Only fold icmp into the PHI if the phi and icmp are in the same
2163 // block. If in the same block, we're encouraging jump threading. If
2164 // not, we are just pessimizing the code by making an i1 phi.
2165 if (LHSI->getParent() == I.getParent())
2166 if (Instruction *NV = FoldOpIntoPhi(I))
2169 case Instruction::Select: {
2170 // If either operand of the select is a constant, we can fold the
2171 // comparison into the select arms, which will cause one to be
2172 // constant folded and the select turned into a bitwise or.
2173 Value *Op1 = 0, *Op2 = 0;
2174 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2175 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2176 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2177 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2179 // We only want to perform this transformation if it will not lead to
2180 // additional code. This is true if either both sides of the select
2181 // fold to a constant (in which case the icmp is replaced with a select
2182 // which will usually simplify) or this is the only user of the
2183 // select (in which case we are trading a select+icmp for a simpler
2185 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2187 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2190 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2192 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2196 case Instruction::IntToPtr:
2197 // icmp pred inttoptr(X), null -> icmp pred X, 0
2198 if (RHSC->isNullValue() && TD &&
2199 TD->getIntPtrType(RHSC->getContext()) ==
2200 LHSI->getOperand(0)->getType())
2201 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2202 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2205 case Instruction::Load:
2206 // Try to optimize things like "A[i] > 4" to index computations.
2207 if (GetElementPtrInst *GEP =
2208 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2209 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2210 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2211 !cast<LoadInst>(LHSI)->isVolatile())
2212 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2219 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2220 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2221 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2223 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2224 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2225 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2228 // Test to see if the operands of the icmp are casted versions of other
2229 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2231 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2232 if (Op0->getType()->isPointerTy() &&
2233 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2234 // We keep moving the cast from the left operand over to the right
2235 // operand, where it can often be eliminated completely.
2236 Op0 = CI->getOperand(0);
2238 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2239 // so eliminate it as well.
2240 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2241 Op1 = CI2->getOperand(0);
2243 // If Op1 is a constant, we can fold the cast into the constant.
2244 if (Op0->getType() != Op1->getType()) {
2245 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2246 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2248 // Otherwise, cast the RHS right before the icmp
2249 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2252 return new ICmpInst(I.getPredicate(), Op0, Op1);
2256 if (isa<CastInst>(Op0)) {
2257 // Handle the special case of: icmp (cast bool to X), <cst>
2258 // This comes up when you have code like
2261 // For generality, we handle any zero-extension of any operand comparison
2262 // with a constant or another cast from the same type.
2263 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2264 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2268 // Special logic for binary operators.
2269 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2270 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2272 CmpInst::Predicate Pred = I.getPredicate();
2273 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2274 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2275 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2276 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2277 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2278 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2279 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2280 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2281 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2283 // Analyze the case when either Op0 or Op1 is an add instruction.
2284 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2285 Value *A = 0, *B = 0, *C = 0, *D = 0;
2286 if (BO0 && BO0->getOpcode() == Instruction::Add)
2287 A = BO0->getOperand(0), B = BO0->getOperand(1);
2288 if (BO1 && BO1->getOpcode() == Instruction::Add)
2289 C = BO1->getOperand(0), D = BO1->getOperand(1);
2291 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2292 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2293 return new ICmpInst(Pred, A == Op1 ? B : A,
2294 Constant::getNullValue(Op1->getType()));
2296 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2297 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2298 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2301 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2302 if (A && C && (A == C || A == D || B == C || B == D) &&
2303 NoOp0WrapProblem && NoOp1WrapProblem &&
2304 // Try not to increase register pressure.
2305 BO0->hasOneUse() && BO1->hasOneUse()) {
2306 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2307 Value *Y = (A == C || A == D) ? B : A;
2308 Value *Z = (C == A || C == B) ? D : C;
2309 return new ICmpInst(Pred, Y, Z);
2312 // Analyze the case when either Op0 or Op1 is a sub instruction.
2313 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2314 A = 0; B = 0; C = 0; D = 0;
2315 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2316 A = BO0->getOperand(0), B = BO0->getOperand(1);
2317 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2318 C = BO1->getOperand(0), D = BO1->getOperand(1);
2320 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2321 if (A == Op1 && NoOp0WrapProblem)
2322 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2324 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2325 if (C == Op0 && NoOp1WrapProblem)
2326 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2328 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2329 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2330 // Try not to increase register pressure.
2331 BO0->hasOneUse() && BO1->hasOneUse())
2332 return new ICmpInst(Pred, A, C);
2334 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2335 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2336 // Try not to increase register pressure.
2337 BO0->hasOneUse() && BO1->hasOneUse())
2338 return new ICmpInst(Pred, D, B);
2340 BinaryOperator *SRem = NULL;
2341 // icmp (srem X, Y), Y
2342 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2343 Op1 == BO0->getOperand(1))
2345 // icmp Y, (srem X, Y)
2346 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2347 Op0 == BO1->getOperand(1))
2350 // We don't check hasOneUse to avoid increasing register pressure because
2351 // the value we use is the same value this instruction was already using.
2352 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2354 case ICmpInst::ICMP_EQ:
2355 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2356 case ICmpInst::ICMP_NE:
2357 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2358 case ICmpInst::ICMP_SGT:
2359 case ICmpInst::ICMP_SGE:
2360 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2361 Constant::getAllOnesValue(SRem->getType()));
2362 case ICmpInst::ICMP_SLT:
2363 case ICmpInst::ICMP_SLE:
2364 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2365 Constant::getNullValue(SRem->getType()));
2369 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2370 BO0->hasOneUse() && BO1->hasOneUse() &&
2371 BO0->getOperand(1) == BO1->getOperand(1)) {
2372 switch (BO0->getOpcode()) {
2374 case Instruction::Add:
2375 case Instruction::Sub:
2376 case Instruction::Xor:
2377 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2378 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2379 BO1->getOperand(0));
2380 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2381 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2382 if (CI->getValue().isSignBit()) {
2383 ICmpInst::Predicate Pred = I.isSigned()
2384 ? I.getUnsignedPredicate()
2385 : I.getSignedPredicate();
2386 return new ICmpInst(Pred, BO0->getOperand(0),
2387 BO1->getOperand(0));
2390 if (CI->getValue().isMaxSignedValue()) {
2391 ICmpInst::Predicate Pred = I.isSigned()
2392 ? I.getUnsignedPredicate()
2393 : I.getSignedPredicate();
2394 Pred = I.getSwappedPredicate(Pred);
2395 return new ICmpInst(Pred, BO0->getOperand(0),
2396 BO1->getOperand(0));
2400 case Instruction::Mul:
2401 if (!I.isEquality())
2404 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2405 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2406 // Mask = -1 >> count-trailing-zeros(Cst).
2407 if (!CI->isZero() && !CI->isOne()) {
2408 const APInt &AP = CI->getValue();
2409 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2410 APInt::getLowBitsSet(AP.getBitWidth(),
2412 AP.countTrailingZeros()));
2413 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2414 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2415 return new ICmpInst(I.getPredicate(), And1, And2);
2419 case Instruction::UDiv:
2420 case Instruction::LShr:
2424 case Instruction::SDiv:
2425 case Instruction::AShr:
2426 if (!BO0->isExact() || !BO1->isExact())
2428 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2429 BO1->getOperand(0));
2430 case Instruction::Shl: {
2431 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2432 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2435 if (!NSW && I.isSigned())
2437 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2438 BO1->getOperand(0));
2445 // ~x < ~y --> y < x
2446 // ~x < cst --> ~cst < x
2447 if (match(Op0, m_Not(m_Value(A)))) {
2448 if (match(Op1, m_Not(m_Value(B))))
2449 return new ICmpInst(I.getPredicate(), B, A);
2450 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2451 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2454 // (a+b) <u a --> llvm.uadd.with.overflow.
2455 // (a+b) <u b --> llvm.uadd.with.overflow.
2456 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2457 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2458 (Op1 == A || Op1 == B))
2459 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2462 // a >u (a+b) --> llvm.uadd.with.overflow.
2463 // b >u (a+b) --> llvm.uadd.with.overflow.
2464 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2465 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2466 (Op0 == A || Op0 == B))
2467 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2471 if (I.isEquality()) {
2472 Value *A, *B, *C, *D;
2474 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2475 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2476 Value *OtherVal = A == Op1 ? B : A;
2477 return new ICmpInst(I.getPredicate(), OtherVal,
2478 Constant::getNullValue(A->getType()));
2481 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2482 // A^c1 == C^c2 --> A == C^(c1^c2)
2483 ConstantInt *C1, *C2;
2484 if (match(B, m_ConstantInt(C1)) &&
2485 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2486 Constant *NC = ConstantInt::get(I.getContext(),
2487 C1->getValue() ^ C2->getValue());
2488 Value *Xor = Builder->CreateXor(C, NC, "tmp");
2489 return new ICmpInst(I.getPredicate(), A, Xor);
2492 // A^B == A^D -> B == D
2493 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2494 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2495 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2496 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2500 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2501 (A == Op0 || B == Op0)) {
2502 // A == (A^B) -> B == 0
2503 Value *OtherVal = A == Op0 ? B : A;
2504 return new ICmpInst(I.getPredicate(), OtherVal,
2505 Constant::getNullValue(A->getType()));
2508 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2509 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2510 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2511 Value *X = 0, *Y = 0, *Z = 0;
2514 X = B; Y = D; Z = A;
2515 } else if (A == D) {
2516 X = B; Y = C; Z = A;
2517 } else if (B == C) {
2518 X = A; Y = D; Z = B;
2519 } else if (B == D) {
2520 X = A; Y = C; Z = B;
2523 if (X) { // Build (X^Y) & Z
2524 Op1 = Builder->CreateXor(X, Y, "tmp");
2525 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
2526 I.setOperand(0, Op1);
2527 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2532 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2533 // "icmp (and X, mask), cst"
2536 if (Op0->hasOneUse() &&
2537 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2538 m_ConstantInt(ShAmt))))) &&
2539 match(Op1, m_ConstantInt(Cst1)) &&
2540 // Only do this when A has multiple uses. This is most important to do
2541 // when it exposes other optimizations.
2543 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2545 if (ShAmt < ASize) {
2547 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2550 APInt CmpV = Cst1->getValue().zext(ASize);
2553 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2554 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2560 Value *X; ConstantInt *Cst;
2562 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2563 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2566 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2567 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2569 return Changed ? &I : 0;
2577 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2579 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2582 if (!isa<ConstantFP>(RHSC)) return 0;
2583 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2585 // Get the width of the mantissa. We don't want to hack on conversions that
2586 // might lose information from the integer, e.g. "i64 -> float"
2587 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2588 if (MantissaWidth == -1) return 0; // Unknown.
2590 // Check to see that the input is converted from an integer type that is small
2591 // enough that preserves all bits. TODO: check here for "known" sign bits.
2592 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2593 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2595 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2596 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2600 // If the conversion would lose info, don't hack on this.
2601 if ((int)InputSize > MantissaWidth)
2604 // Otherwise, we can potentially simplify the comparison. We know that it
2605 // will always come through as an integer value and we know the constant is
2606 // not a NAN (it would have been previously simplified).
2607 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2609 ICmpInst::Predicate Pred;
2610 switch (I.getPredicate()) {
2611 default: llvm_unreachable("Unexpected predicate!");
2612 case FCmpInst::FCMP_UEQ:
2613 case FCmpInst::FCMP_OEQ:
2614 Pred = ICmpInst::ICMP_EQ;
2616 case FCmpInst::FCMP_UGT:
2617 case FCmpInst::FCMP_OGT:
2618 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2620 case FCmpInst::FCMP_UGE:
2621 case FCmpInst::FCMP_OGE:
2622 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2624 case FCmpInst::FCMP_ULT:
2625 case FCmpInst::FCMP_OLT:
2626 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2628 case FCmpInst::FCMP_ULE:
2629 case FCmpInst::FCMP_OLE:
2630 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2632 case FCmpInst::FCMP_UNE:
2633 case FCmpInst::FCMP_ONE:
2634 Pred = ICmpInst::ICMP_NE;
2636 case FCmpInst::FCMP_ORD:
2637 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2638 case FCmpInst::FCMP_UNO:
2639 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2642 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2644 // Now we know that the APFloat is a normal number, zero or inf.
2646 // See if the FP constant is too large for the integer. For example,
2647 // comparing an i8 to 300.0.
2648 unsigned IntWidth = IntTy->getScalarSizeInBits();
2651 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2652 // and large values.
2653 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2654 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2655 APFloat::rmNearestTiesToEven);
2656 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2657 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2658 Pred == ICmpInst::ICMP_SLE)
2659 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2660 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2663 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2664 // +INF and large values.
2665 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2666 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2667 APFloat::rmNearestTiesToEven);
2668 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2669 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2670 Pred == ICmpInst::ICMP_ULE)
2671 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2672 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2677 // See if the RHS value is < SignedMin.
2678 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2679 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2680 APFloat::rmNearestTiesToEven);
2681 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2682 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2683 Pred == ICmpInst::ICMP_SGE)
2684 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2685 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2689 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2690 // [0, UMAX], but it may still be fractional. See if it is fractional by
2691 // casting the FP value to the integer value and back, checking for equality.
2692 // Don't do this for zero, because -0.0 is not fractional.
2693 Constant *RHSInt = LHSUnsigned
2694 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2695 : ConstantExpr::getFPToSI(RHSC, IntTy);
2696 if (!RHS.isZero()) {
2697 bool Equal = LHSUnsigned
2698 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2699 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2701 // If we had a comparison against a fractional value, we have to adjust
2702 // the compare predicate and sometimes the value. RHSC is rounded towards
2703 // zero at this point.
2705 default: llvm_unreachable("Unexpected integer comparison!");
2706 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2707 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2708 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2709 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2710 case ICmpInst::ICMP_ULE:
2711 // (float)int <= 4.4 --> int <= 4
2712 // (float)int <= -4.4 --> false
2713 if (RHS.isNegative())
2714 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2716 case ICmpInst::ICMP_SLE:
2717 // (float)int <= 4.4 --> int <= 4
2718 // (float)int <= -4.4 --> int < -4
2719 if (RHS.isNegative())
2720 Pred = ICmpInst::ICMP_SLT;
2722 case ICmpInst::ICMP_ULT:
2723 // (float)int < -4.4 --> false
2724 // (float)int < 4.4 --> int <= 4
2725 if (RHS.isNegative())
2726 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2727 Pred = ICmpInst::ICMP_ULE;
2729 case ICmpInst::ICMP_SLT:
2730 // (float)int < -4.4 --> int < -4
2731 // (float)int < 4.4 --> int <= 4
2732 if (!RHS.isNegative())
2733 Pred = ICmpInst::ICMP_SLE;
2735 case ICmpInst::ICMP_UGT:
2736 // (float)int > 4.4 --> int > 4
2737 // (float)int > -4.4 --> true
2738 if (RHS.isNegative())
2739 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2741 case ICmpInst::ICMP_SGT:
2742 // (float)int > 4.4 --> int > 4
2743 // (float)int > -4.4 --> int >= -4
2744 if (RHS.isNegative())
2745 Pred = ICmpInst::ICMP_SGE;
2747 case ICmpInst::ICMP_UGE:
2748 // (float)int >= -4.4 --> true
2749 // (float)int >= 4.4 --> int > 4
2750 if (!RHS.isNegative())
2751 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2752 Pred = ICmpInst::ICMP_UGT;
2754 case ICmpInst::ICMP_SGE:
2755 // (float)int >= -4.4 --> int >= -4
2756 // (float)int >= 4.4 --> int > 4
2757 if (!RHS.isNegative())
2758 Pred = ICmpInst::ICMP_SGT;
2764 // Lower this FP comparison into an appropriate integer version of the
2766 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2769 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2770 bool Changed = false;
2772 /// Orders the operands of the compare so that they are listed from most
2773 /// complex to least complex. This puts constants before unary operators,
2774 /// before binary operators.
2775 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2780 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2782 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2783 return ReplaceInstUsesWith(I, V);
2785 // Simplify 'fcmp pred X, X'
2787 switch (I.getPredicate()) {
2788 default: llvm_unreachable("Unknown predicate!");
2789 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2790 case FCmpInst::FCMP_ULT: // True if unordered or less than
2791 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2792 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2793 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2794 I.setPredicate(FCmpInst::FCMP_UNO);
2795 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2798 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2799 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2800 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2801 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2802 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2803 I.setPredicate(FCmpInst::FCMP_ORD);
2804 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2809 // Handle fcmp with constant RHS
2810 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2811 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2812 switch (LHSI->getOpcode()) {
2813 case Instruction::FPExt: {
2814 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2815 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2816 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2820 // We can't convert a PPC double double.
2821 if (RHSF->getType()->isPPC_FP128Ty())
2824 const fltSemantics *Sem;
2825 // FIXME: This shouldn't be here.
2826 if (LHSExt->getSrcTy()->isFloatTy())
2827 Sem = &APFloat::IEEEsingle;
2828 else if (LHSExt->getSrcTy()->isDoubleTy())
2829 Sem = &APFloat::IEEEdouble;
2830 else if (LHSExt->getSrcTy()->isFP128Ty())
2831 Sem = &APFloat::IEEEquad;
2832 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2833 Sem = &APFloat::x87DoubleExtended;
2838 APFloat F = RHSF->getValueAPF();
2839 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2841 // Avoid lossy conversions and denormals.
2843 F.compare(APFloat::getSmallestNormalized(*Sem)) !=
2844 APFloat::cmpLessThan)
2845 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2846 ConstantFP::get(RHSC->getContext(), F));
2849 case Instruction::PHI:
2850 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2851 // block. If in the same block, we're encouraging jump threading. If
2852 // not, we are just pessimizing the code by making an i1 phi.
2853 if (LHSI->getParent() == I.getParent())
2854 if (Instruction *NV = FoldOpIntoPhi(I))
2857 case Instruction::SIToFP:
2858 case Instruction::UIToFP:
2859 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2862 case Instruction::Select: {
2863 // If either operand of the select is a constant, we can fold the
2864 // comparison into the select arms, which will cause one to be
2865 // constant folded and the select turned into a bitwise or.
2866 Value *Op1 = 0, *Op2 = 0;
2867 if (LHSI->hasOneUse()) {
2868 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2869 // Fold the known value into the constant operand.
2870 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2871 // Insert a new FCmp of the other select operand.
2872 Op2 = Builder->CreateFCmp(I.getPredicate(),
2873 LHSI->getOperand(2), RHSC, I.getName());
2874 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2875 // Fold the known value into the constant operand.
2876 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2877 // Insert a new FCmp of the other select operand.
2878 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2884 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2887 case Instruction::FSub: {
2888 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2890 if (match(LHSI, m_FNeg(m_Value(Op))))
2891 return new FCmpInst(I.getSwappedPredicate(), Op,
2892 ConstantExpr::getFNeg(RHSC));
2895 case Instruction::Load:
2896 if (GetElementPtrInst *GEP =
2897 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2898 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2899 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2900 !cast<LoadInst>(LHSI)->isVolatile())
2901 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2908 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
2910 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
2911 return new FCmpInst(I.getSwappedPredicate(), X, Y);
2913 // fcmp (fpext x), (fpext y) -> fcmp x, y
2914 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
2915 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
2916 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
2917 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2918 RHSExt->getOperand(0));
2920 return Changed ? &I : 0;