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 return Result->getValue().ult(In1->getValue());
48 if (In2->isNegative())
49 return Result->getValue().sgt(In1->getValue());
50 return Result->getValue().slt(In1->getValue());
53 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
54 /// overflowed for this type.
55 static bool AddWithOverflow(Constant *&Result, Constant *In1,
56 Constant *In2, bool IsSigned = false) {
57 Result = ConstantExpr::getAdd(In1, In2);
59 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
60 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
61 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
62 if (HasAddOverflow(ExtractElement(Result, Idx),
63 ExtractElement(In1, Idx),
64 ExtractElement(In2, Idx),
71 return HasAddOverflow(cast<ConstantInt>(Result),
72 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
76 static bool HasSubOverflow(ConstantInt *Result,
77 ConstantInt *In1, ConstantInt *In2,
80 return Result->getValue().ugt(In1->getValue());
82 if (In2->isNegative())
83 return Result->getValue().slt(In1->getValue());
85 return Result->getValue().sgt(In1->getValue());
88 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
89 /// overflowed for this type.
90 static bool SubWithOverflow(Constant *&Result, Constant *In1,
91 Constant *In2, bool IsSigned = false) {
92 Result = ConstantExpr::getSub(In1, In2);
94 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
95 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
96 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
97 if (HasSubOverflow(ExtractElement(Result, Idx),
98 ExtractElement(In1, Idx),
99 ExtractElement(In2, Idx),
106 return HasSubOverflow(cast<ConstantInt>(Result),
107 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
111 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
112 /// comparison only checks the sign bit. If it only checks the sign bit, set
113 /// TrueIfSigned if the result of the comparison is true when the input value is
115 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
116 bool &TrueIfSigned) {
118 case ICmpInst::ICMP_SLT: // True if LHS s< 0
120 return RHS->isZero();
121 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
123 return RHS->isAllOnesValue();
124 case ICmpInst::ICMP_SGT: // True if LHS s> -1
125 TrueIfSigned = false;
126 return RHS->isAllOnesValue();
127 case ICmpInst::ICMP_UGT:
128 // True if LHS u> RHS and RHS == high-bit-mask - 1
130 return RHS->isMaxValue(true);
131 case ICmpInst::ICMP_UGE:
132 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
134 return RHS->getValue().isSignBit();
140 // isHighOnes - Return true if the constant is of the form 1+0+.
141 // This is the same as lowones(~X).
142 static bool isHighOnes(const ConstantInt *CI) {
143 return (~CI->getValue() + 1).isPowerOf2();
146 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
147 /// set of known zero and one bits, compute the maximum and minimum values that
148 /// could have the specified known zero and known one bits, returning them in
150 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
151 const APInt& KnownOne,
152 APInt& Min, APInt& Max) {
153 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
154 KnownZero.getBitWidth() == Min.getBitWidth() &&
155 KnownZero.getBitWidth() == Max.getBitWidth() &&
156 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
157 APInt UnknownBits = ~(KnownZero|KnownOne);
159 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
160 // bit if it is unknown.
162 Max = KnownOne|UnknownBits;
164 if (UnknownBits.isNegative()) { // Sign bit is unknown
165 Min.setBit(Min.getBitWidth()-1);
166 Max.clearBit(Max.getBitWidth()-1);
170 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
171 // a set of known zero and one bits, compute the maximum and minimum values that
172 // could have the specified known zero and known one bits, returning them in
174 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
175 const APInt &KnownOne,
176 APInt &Min, APInt &Max) {
177 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
178 KnownZero.getBitWidth() == Min.getBitWidth() &&
179 KnownZero.getBitWidth() == Max.getBitWidth() &&
180 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
181 APInt UnknownBits = ~(KnownZero|KnownOne);
183 // The minimum value is when the unknown bits are all zeros.
185 // The maximum value is when the unknown bits are all ones.
186 Max = KnownOne|UnknownBits;
191 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
192 /// cmp pred (load (gep GV, ...)), cmpcst
193 /// where GV is a global variable with a constant initializer. Try to simplify
194 /// this into some simple computation that does not need the load. For example
195 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
197 /// If AndCst is non-null, then the loaded value is masked with that constant
198 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
199 Instruction *InstCombiner::
200 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
201 CmpInst &ICI, ConstantInt *AndCst) {
202 // We need TD information to know the pointer size unless this is inbounds.
203 if (!GEP->isInBounds() && TD == 0) return 0;
205 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer());
206 if (Init == 0 || Init->getNumOperands() > 1024) return 0;
208 // There are many forms of this optimization we can handle, for now, just do
209 // the simple index into a single-dimensional array.
211 // Require: GEP GV, 0, i {{, constant indices}}
212 if (GEP->getNumOperands() < 3 ||
213 !isa<ConstantInt>(GEP->getOperand(1)) ||
214 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
215 isa<Constant>(GEP->getOperand(2)))
218 // Check that indices after the variable are constants and in-range for the
219 // type they index. Collect the indices. This is typically for arrays of
221 SmallVector<unsigned, 4> LaterIndices;
223 Type *EltTy = cast<ArrayType>(Init->getType())->getElementType();
224 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
225 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
226 if (Idx == 0) return 0; // Variable index.
228 uint64_t IdxVal = Idx->getZExtValue();
229 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
231 if (StructType *STy = dyn_cast<StructType>(EltTy))
232 EltTy = STy->getElementType(IdxVal);
233 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
234 if (IdxVal >= ATy->getNumElements()) return 0;
235 EltTy = ATy->getElementType();
237 return 0; // Unknown type.
240 LaterIndices.push_back(IdxVal);
243 enum { Overdefined = -3, Undefined = -2 };
245 // Variables for our state machines.
247 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
248 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
249 // and 87 is the second (and last) index. FirstTrueElement is -2 when
250 // undefined, otherwise set to the first true element. SecondTrueElement is
251 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
252 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
254 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
255 // form "i != 47 & i != 87". Same state transitions as for true elements.
256 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
258 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
259 /// define a state machine that triggers for ranges of values that the index
260 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
261 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
262 /// index in the range (inclusive). We use -2 for undefined here because we
263 /// use relative comparisons and don't want 0-1 to match -1.
264 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
266 // MagicBitvector - This is a magic bitvector where we set a bit if the
267 // comparison is true for element 'i'. If there are 64 elements or less in
268 // the array, this will fully represent all the comparison results.
269 uint64_t MagicBitvector = 0;
272 // Scan the array and see if one of our patterns matches.
273 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
274 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) {
275 Constant *Elt = Init->getOperand(i);
277 // If this is indexing an array of structures, get the structure element.
278 if (!LaterIndices.empty())
279 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
281 // If the element is masked, handle it.
282 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
284 // Find out if the comparison would be true or false for the i'th element.
285 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
287 // If the result is undef for this element, ignore it.
288 if (isa<UndefValue>(C)) {
289 // Extend range state machines to cover this element in case there is an
290 // undef in the middle of the range.
291 if (TrueRangeEnd == (int)i-1)
293 if (FalseRangeEnd == (int)i-1)
298 // If we can't compute the result for any of the elements, we have to give
299 // up evaluating the entire conditional.
300 if (!isa<ConstantInt>(C)) return 0;
302 // Otherwise, we know if the comparison is true or false for this element,
303 // update our state machines.
304 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
306 // State machine for single/double/range index comparison.
308 // Update the TrueElement state machine.
309 if (FirstTrueElement == Undefined)
310 FirstTrueElement = TrueRangeEnd = i; // First true element.
312 // Update double-compare state machine.
313 if (SecondTrueElement == Undefined)
314 SecondTrueElement = i;
316 SecondTrueElement = Overdefined;
318 // Update range state machine.
319 if (TrueRangeEnd == (int)i-1)
322 TrueRangeEnd = Overdefined;
325 // Update the FalseElement state machine.
326 if (FirstFalseElement == Undefined)
327 FirstFalseElement = FalseRangeEnd = i; // First false element.
329 // Update double-compare state machine.
330 if (SecondFalseElement == Undefined)
331 SecondFalseElement = i;
333 SecondFalseElement = Overdefined;
335 // Update range state machine.
336 if (FalseRangeEnd == (int)i-1)
339 FalseRangeEnd = Overdefined;
344 // If this element is in range, update our magic bitvector.
345 if (i < 64 && IsTrueForElt)
346 MagicBitvector |= 1ULL << i;
348 // If all of our states become overdefined, bail out early. Since the
349 // predicate is expensive, only check it every 8 elements. This is only
350 // really useful for really huge arrays.
351 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
352 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
353 FalseRangeEnd == Overdefined)
357 // Now that we've scanned the entire array, emit our new comparison(s). We
358 // order the state machines in complexity of the generated code.
359 Value *Idx = GEP->getOperand(2);
361 // If the index is larger than the pointer size of the target, truncate the
362 // index down like the GEP would do implicitly. We don't have to do this for
363 // an inbounds GEP because the index can't be out of range.
364 if (!GEP->isInBounds() &&
365 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
366 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
368 // If the comparison is only true for one or two elements, emit direct
370 if (SecondTrueElement != Overdefined) {
371 // None true -> false.
372 if (FirstTrueElement == Undefined)
373 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
375 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
377 // True for one element -> 'i == 47'.
378 if (SecondTrueElement == Undefined)
379 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
381 // True for two elements -> 'i == 47 | i == 72'.
382 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
383 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
384 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
385 return BinaryOperator::CreateOr(C1, C2);
388 // If the comparison is only false for one or two elements, emit direct
390 if (SecondFalseElement != Overdefined) {
391 // None false -> true.
392 if (FirstFalseElement == Undefined)
393 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
395 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
397 // False for one element -> 'i != 47'.
398 if (SecondFalseElement == Undefined)
399 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
401 // False for two elements -> 'i != 47 & i != 72'.
402 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
403 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
404 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
405 return BinaryOperator::CreateAnd(C1, C2);
408 // If the comparison can be replaced with a range comparison for the elements
409 // where it is true, emit the range check.
410 if (TrueRangeEnd != Overdefined) {
411 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
413 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
414 if (FirstTrueElement) {
415 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
416 Idx = Builder->CreateAdd(Idx, Offs);
419 Value *End = ConstantInt::get(Idx->getType(),
420 TrueRangeEnd-FirstTrueElement+1);
421 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
424 // False range check.
425 if (FalseRangeEnd != Overdefined) {
426 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
427 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
428 if (FirstFalseElement) {
429 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
430 Idx = Builder->CreateAdd(Idx, Offs);
433 Value *End = ConstantInt::get(Idx->getType(),
434 FalseRangeEnd-FirstFalseElement);
435 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
439 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
440 // of this load, replace it with computation that does:
441 // ((magic_cst >> i) & 1) != 0
442 if (Init->getNumOperands() <= 32 ||
443 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) {
445 if (Init->getNumOperands() <= 32)
446 Ty = Type::getInt32Ty(Init->getContext());
448 Ty = Type::getInt64Ty(Init->getContext());
449 Value *V = Builder->CreateIntCast(Idx, Ty, false);
450 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
451 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
452 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
459 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
460 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
461 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
462 /// be complex, and scales are involved. The above expression would also be
463 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
464 /// This later form is less amenable to optimization though, and we are allowed
465 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
467 /// If we can't emit an optimized form for this expression, this returns null.
469 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
470 TargetData &TD = *IC.getTargetData();
471 gep_type_iterator GTI = gep_type_begin(GEP);
473 // Check to see if this gep only has a single variable index. If so, and if
474 // any constant indices are a multiple of its scale, then we can compute this
475 // in terms of the scale of the variable index. For example, if the GEP
476 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
477 // because the expression will cross zero at the same point.
478 unsigned i, e = GEP->getNumOperands();
480 for (i = 1; i != e; ++i, ++GTI) {
481 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
482 // Compute the aggregate offset of constant indices.
483 if (CI->isZero()) continue;
485 // Handle a struct index, which adds its field offset to the pointer.
486 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
487 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
489 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
490 Offset += Size*CI->getSExtValue();
493 // Found our variable index.
498 // If there are no variable indices, we must have a constant offset, just
499 // evaluate it the general way.
500 if (i == e) return 0;
502 Value *VariableIdx = GEP->getOperand(i);
503 // Determine the scale factor of the variable element. For example, this is
504 // 4 if the variable index is into an array of i32.
505 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
507 // Verify that there are no other variable indices. If so, emit the hard way.
508 for (++i, ++GTI; i != e; ++i, ++GTI) {
509 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
512 // Compute the aggregate offset of constant indices.
513 if (CI->isZero()) continue;
515 // Handle a struct index, which adds its field offset to the pointer.
516 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
517 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
519 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
520 Offset += Size*CI->getSExtValue();
524 // Okay, we know we have a single variable index, which must be a
525 // pointer/array/vector index. If there is no offset, life is simple, return
527 unsigned IntPtrWidth = TD.getPointerSizeInBits();
529 // Cast to intptrty in case a truncation occurs. If an extension is needed,
530 // we don't need to bother extending: the extension won't affect where the
531 // computation crosses zero.
532 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
533 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
534 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
539 // Otherwise, there is an index. The computation we will do will be modulo
540 // the pointer size, so get it.
541 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
543 Offset &= PtrSizeMask;
544 VariableScale &= PtrSizeMask;
546 // To do this transformation, any constant index must be a multiple of the
547 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
548 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
549 // multiple of the variable scale.
550 int64_t NewOffs = Offset / (int64_t)VariableScale;
551 if (Offset != NewOffs*(int64_t)VariableScale)
554 // Okay, we can do this evaluation. Start by converting the index to intptr.
555 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
556 if (VariableIdx->getType() != IntPtrTy)
557 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
559 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
560 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
563 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
564 /// else. At this point we know that the GEP is on the LHS of the comparison.
565 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
566 ICmpInst::Predicate Cond,
568 // Look through bitcasts.
569 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
570 RHS = BCI->getOperand(0);
572 Value *PtrBase = GEPLHS->getOperand(0);
573 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
574 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
575 // This transformation (ignoring the base and scales) is valid because we
576 // know pointers can't overflow since the gep is inbounds. See if we can
577 // output an optimized form.
578 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
580 // If not, synthesize the offset the hard way.
582 Offset = EmitGEPOffset(GEPLHS);
583 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
584 Constant::getNullValue(Offset->getType()));
585 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
586 // If the base pointers are different, but the indices are the same, just
587 // compare the base pointer.
588 if (PtrBase != GEPRHS->getOperand(0)) {
589 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
590 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
591 GEPRHS->getOperand(0)->getType();
593 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
594 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
595 IndicesTheSame = false;
599 // If all indices are the same, just compare the base pointers.
601 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
602 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
604 // Otherwise, the base pointers are different and the indices are
605 // different, bail out.
609 // If one of the GEPs has all zero indices, recurse.
610 bool AllZeros = true;
611 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
612 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
613 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
618 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
619 ICmpInst::getSwappedPredicate(Cond), I);
621 // If the other GEP has all zero indices, recurse.
623 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
624 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
625 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
630 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
632 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
633 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
634 // If the GEPs only differ by one index, compare it.
635 unsigned NumDifferences = 0; // Keep track of # differences.
636 unsigned DiffOperand = 0; // The operand that differs.
637 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
638 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
639 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
640 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
641 // Irreconcilable differences.
645 if (NumDifferences++) break;
650 if (NumDifferences == 0) // SAME GEP?
651 return ReplaceInstUsesWith(I, // No comparison is needed here.
652 ConstantInt::get(Type::getInt1Ty(I.getContext()),
653 ICmpInst::isTrueWhenEqual(Cond)));
655 else if (NumDifferences == 1 && GEPsInBounds) {
656 Value *LHSV = GEPLHS->getOperand(DiffOperand);
657 Value *RHSV = GEPRHS->getOperand(DiffOperand);
658 // Make sure we do a signed comparison here.
659 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
663 // Only lower this if the icmp is the only user of the GEP or if we expect
664 // the result to fold to a constant!
667 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
668 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
669 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
670 Value *L = EmitGEPOffset(GEPLHS);
671 Value *R = EmitGEPOffset(GEPRHS);
672 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
678 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
679 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
680 Value *X, ConstantInt *CI,
681 ICmpInst::Predicate Pred,
683 // If we have X+0, exit early (simplifying logic below) and let it get folded
684 // elsewhere. icmp X+0, X -> icmp X, X
686 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
687 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
690 // (X+4) == X -> false.
691 if (Pred == ICmpInst::ICMP_EQ)
692 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
694 // (X+4) != X -> true.
695 if (Pred == ICmpInst::ICMP_NE)
696 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
698 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
699 // so the values can never be equal. Similarly for all other "or equals"
702 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
703 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
704 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
705 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
707 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
708 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
711 // (X+1) >u X --> X <u (0-1) --> X != 255
712 // (X+2) >u X --> X <u (0-2) --> X <u 254
713 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
714 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
715 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
717 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
718 ConstantInt *SMax = ConstantInt::get(X->getContext(),
719 APInt::getSignedMaxValue(BitWidth));
721 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
722 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
723 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
724 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
725 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
726 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
727 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
728 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
730 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
731 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
732 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
733 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
734 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
735 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
737 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
738 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
739 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
742 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
743 /// and CmpRHS are both known to be integer constants.
744 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
745 ConstantInt *DivRHS) {
746 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
747 const APInt &CmpRHSV = CmpRHS->getValue();
749 // FIXME: If the operand types don't match the type of the divide
750 // then don't attempt this transform. The code below doesn't have the
751 // logic to deal with a signed divide and an unsigned compare (and
752 // vice versa). This is because (x /s C1) <s C2 produces different
753 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
754 // (x /u C1) <u C2. Simply casting the operands and result won't
755 // work. :( The if statement below tests that condition and bails
757 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
758 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
760 if (DivRHS->isZero())
761 return 0; // The ProdOV computation fails on divide by zero.
762 if (DivIsSigned && DivRHS->isAllOnesValue())
763 return 0; // The overflow computation also screws up here
764 if (DivRHS->isOne()) {
765 // This eliminates some funny cases with INT_MIN.
766 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
770 // Compute Prod = CI * DivRHS. We are essentially solving an equation
771 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
772 // C2 (CI). By solving for X we can turn this into a range check
773 // instead of computing a divide.
774 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
776 // Determine if the product overflows by seeing if the product is
777 // not equal to the divide. Make sure we do the same kind of divide
778 // as in the LHS instruction that we're folding.
779 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
780 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
782 // Get the ICmp opcode
783 ICmpInst::Predicate Pred = ICI.getPredicate();
785 /// If the division is known to be exact, then there is no remainder from the
786 /// divide, so the covered range size is unit, otherwise it is the divisor.
787 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
789 // Figure out the interval that is being checked. For example, a comparison
790 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
791 // Compute this interval based on the constants involved and the signedness of
792 // the compare/divide. This computes a half-open interval, keeping track of
793 // whether either value in the interval overflows. After analysis each
794 // overflow variable is set to 0 if it's corresponding bound variable is valid
795 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
796 int LoOverflow = 0, HiOverflow = 0;
797 Constant *LoBound = 0, *HiBound = 0;
799 if (!DivIsSigned) { // udiv
800 // e.g. X/5 op 3 --> [15, 20)
802 HiOverflow = LoOverflow = ProdOV;
804 // If this is not an exact divide, then many values in the range collapse
805 // to the same result value.
806 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
809 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
810 if (CmpRHSV == 0) { // (X / pos) op 0
811 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
812 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
814 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
815 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
816 HiOverflow = LoOverflow = ProdOV;
818 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
819 } else { // (X / pos) op neg
820 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
821 HiBound = AddOne(Prod);
822 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
824 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
825 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
828 } else if (DivRHS->isNegative()) { // Divisor is < 0.
830 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
831 if (CmpRHSV == 0) { // (X / neg) op 0
832 // e.g. X/-5 op 0 --> [-4, 5)
833 LoBound = AddOne(RangeSize);
834 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
835 if (HiBound == DivRHS) { // -INTMIN = INTMIN
836 HiOverflow = 1; // [INTMIN+1, overflow)
837 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
839 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
840 // e.g. X/-5 op 3 --> [-19, -14)
841 HiBound = AddOne(Prod);
842 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
844 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
845 } else { // (X / neg) op neg
846 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
847 LoOverflow = HiOverflow = ProdOV;
849 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
852 // Dividing by a negative swaps the condition. LT <-> GT
853 Pred = ICmpInst::getSwappedPredicate(Pred);
856 Value *X = DivI->getOperand(0);
858 default: llvm_unreachable("Unhandled icmp opcode!");
859 case ICmpInst::ICMP_EQ:
860 if (LoOverflow && HiOverflow)
861 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
863 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
864 ICmpInst::ICMP_UGE, X, LoBound);
866 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
867 ICmpInst::ICMP_ULT, X, HiBound);
868 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
870 case ICmpInst::ICMP_NE:
871 if (LoOverflow && HiOverflow)
872 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
874 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
875 ICmpInst::ICMP_ULT, X, LoBound);
877 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
878 ICmpInst::ICMP_UGE, X, HiBound);
879 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
880 DivIsSigned, false));
881 case ICmpInst::ICMP_ULT:
882 case ICmpInst::ICMP_SLT:
883 if (LoOverflow == +1) // Low bound is greater than input range.
884 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
885 if (LoOverflow == -1) // Low bound is less than input range.
886 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
887 return new ICmpInst(Pred, X, LoBound);
888 case ICmpInst::ICMP_UGT:
889 case ICmpInst::ICMP_SGT:
890 if (HiOverflow == +1) // High bound greater than input range.
891 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
892 if (HiOverflow == -1) // High bound less than input range.
893 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
894 if (Pred == ICmpInst::ICMP_UGT)
895 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
896 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
900 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
901 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
902 ConstantInt *ShAmt) {
903 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
905 // Check that the shift amount is in range. If not, don't perform
906 // undefined shifts. When the shift is visited it will be
908 uint32_t TypeBits = CmpRHSV.getBitWidth();
909 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
910 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
913 if (!ICI.isEquality()) {
914 // If we have an unsigned comparison and an ashr, we can't simplify this.
915 // Similarly for signed comparisons with lshr.
916 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
919 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
920 // by a power of 2. Since we already have logic to simplify these,
921 // transform to div and then simplify the resultant comparison.
922 if (Shr->getOpcode() == Instruction::AShr &&
923 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
926 // Revisit the shift (to delete it).
930 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
933 Shr->getOpcode() == Instruction::AShr ?
934 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
935 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
937 ICI.setOperand(0, Tmp);
939 // If the builder folded the binop, just return it.
940 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
944 // Otherwise, fold this div/compare.
945 assert(TheDiv->getOpcode() == Instruction::SDiv ||
946 TheDiv->getOpcode() == Instruction::UDiv);
948 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
949 assert(Res && "This div/cst should have folded!");
954 // If we are comparing against bits always shifted out, the
955 // comparison cannot succeed.
956 APInt Comp = CmpRHSV << ShAmtVal;
957 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
958 if (Shr->getOpcode() == Instruction::LShr)
959 Comp = Comp.lshr(ShAmtVal);
961 Comp = Comp.ashr(ShAmtVal);
963 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
964 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
965 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
967 return ReplaceInstUsesWith(ICI, Cst);
970 // Otherwise, check to see if the bits shifted out are known to be zero.
971 // If so, we can compare against the unshifted value:
972 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
973 if (Shr->hasOneUse() && Shr->isExact())
974 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
976 if (Shr->hasOneUse()) {
977 // Otherwise strength reduce the shift into an and.
978 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
979 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
981 Value *And = Builder->CreateAnd(Shr->getOperand(0),
982 Mask, Shr->getName()+".mask");
983 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
989 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
991 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
994 const APInt &RHSV = RHS->getValue();
996 switch (LHSI->getOpcode()) {
997 case Instruction::Trunc:
998 if (ICI.isEquality() && LHSI->hasOneUse()) {
999 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1000 // of the high bits truncated out of x are known.
1001 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1002 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1003 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
1004 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1005 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
1007 // If all the high bits are known, we can do this xform.
1008 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1009 // Pull in the high bits from known-ones set.
1010 APInt NewRHS = RHS->getValue().zext(SrcBits);
1012 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1013 ConstantInt::get(ICI.getContext(), NewRHS));
1018 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1019 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1020 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1022 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1023 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1024 Value *CompareVal = LHSI->getOperand(0);
1026 // If the sign bit of the XorCST is not set, there is no change to
1027 // the operation, just stop using the Xor.
1028 if (!XorCST->isNegative()) {
1029 ICI.setOperand(0, CompareVal);
1034 // Was the old condition true if the operand is positive?
1035 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1037 // If so, the new one isn't.
1038 isTrueIfPositive ^= true;
1040 if (isTrueIfPositive)
1041 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1044 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1048 if (LHSI->hasOneUse()) {
1049 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1050 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1051 const APInt &SignBit = XorCST->getValue();
1052 ICmpInst::Predicate Pred = ICI.isSigned()
1053 ? ICI.getUnsignedPredicate()
1054 : ICI.getSignedPredicate();
1055 return new ICmpInst(Pred, LHSI->getOperand(0),
1056 ConstantInt::get(ICI.getContext(),
1060 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1061 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1062 const APInt &NotSignBit = XorCST->getValue();
1063 ICmpInst::Predicate Pred = ICI.isSigned()
1064 ? ICI.getUnsignedPredicate()
1065 : ICI.getSignedPredicate();
1066 Pred = ICI.getSwappedPredicate(Pred);
1067 return new ICmpInst(Pred, LHSI->getOperand(0),
1068 ConstantInt::get(ICI.getContext(),
1069 RHSV ^ NotSignBit));
1074 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1075 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1076 LHSI->getOperand(0)->hasOneUse()) {
1077 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1079 // If the LHS is an AND of a truncating cast, we can widen the
1080 // and/compare to be the input width without changing the value
1081 // produced, eliminating a cast.
1082 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1083 // We can do this transformation if either the AND constant does not
1084 // have its sign bit set or if it is an equality comparison.
1085 // Extending a relational comparison when we're checking the sign
1086 // bit would not work.
1087 if (ICI.isEquality() ||
1088 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1090 Builder->CreateAnd(Cast->getOperand(0),
1091 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1092 NewAnd->takeName(LHSI);
1093 return new ICmpInst(ICI.getPredicate(), NewAnd,
1094 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1098 // If the LHS is an AND of a zext, and we have an equality compare, we can
1099 // shrink the and/compare to the smaller type, eliminating the cast.
1100 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1101 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1102 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1103 // should fold the icmp to true/false in that case.
1104 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1106 Builder->CreateAnd(Cast->getOperand(0),
1107 ConstantExpr::getTrunc(AndCST, Ty));
1108 NewAnd->takeName(LHSI);
1109 return new ICmpInst(ICI.getPredicate(), NewAnd,
1110 ConstantExpr::getTrunc(RHS, Ty));
1114 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1115 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1116 // happens a LOT in code produced by the C front-end, for bitfield
1118 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1119 if (Shift && !Shift->isShift())
1123 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1124 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1125 Type *AndTy = AndCST->getType(); // Type of the and.
1127 // We can fold this as long as we can't shift unknown bits
1128 // into the mask. This can only happen with signed shift
1129 // rights, as they sign-extend.
1131 bool CanFold = Shift->isLogicalShift();
1133 // To test for the bad case of the signed shr, see if any
1134 // of the bits shifted in could be tested after the mask.
1135 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1136 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1138 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1139 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1140 AndCST->getValue()) == 0)
1146 if (Shift->getOpcode() == Instruction::Shl)
1147 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1149 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1151 // Check to see if we are shifting out any of the bits being
1153 if (ConstantExpr::get(Shift->getOpcode(),
1154 NewCst, ShAmt) != RHS) {
1155 // If we shifted bits out, the fold is not going to work out.
1156 // As a special case, check to see if this means that the
1157 // result is always true or false now.
1158 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1159 return ReplaceInstUsesWith(ICI,
1160 ConstantInt::getFalse(ICI.getContext()));
1161 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1162 return ReplaceInstUsesWith(ICI,
1163 ConstantInt::getTrue(ICI.getContext()));
1165 ICI.setOperand(1, NewCst);
1166 Constant *NewAndCST;
1167 if (Shift->getOpcode() == Instruction::Shl)
1168 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1170 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1171 LHSI->setOperand(1, NewAndCST);
1172 LHSI->setOperand(0, Shift->getOperand(0));
1173 Worklist.Add(Shift); // Shift is dead.
1179 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1180 // preferable because it allows the C<<Y expression to be hoisted out
1181 // of a loop if Y is invariant and X is not.
1182 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1183 ICI.isEquality() && !Shift->isArithmeticShift() &&
1184 !isa<Constant>(Shift->getOperand(0))) {
1187 if (Shift->getOpcode() == Instruction::LShr) {
1188 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
1190 // Insert a logical shift.
1191 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
1194 // Compute X & (C << Y).
1196 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1198 ICI.setOperand(0, NewAnd);
1203 // Try to optimize things like "A[i]&42 == 0" to index computations.
1204 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1205 if (GetElementPtrInst *GEP =
1206 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1207 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1208 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1209 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1210 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1211 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1217 case Instruction::Or: {
1218 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1221 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1222 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1223 // -> and (icmp eq P, null), (icmp eq Q, null).
1224 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1225 Constant::getNullValue(P->getType()));
1226 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1227 Constant::getNullValue(Q->getType()));
1229 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1230 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1232 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1238 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1239 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1242 uint32_t TypeBits = RHSV.getBitWidth();
1244 // Check that the shift amount is in range. If not, don't perform
1245 // undefined shifts. When the shift is visited it will be
1247 if (ShAmt->uge(TypeBits))
1250 if (ICI.isEquality()) {
1251 // If we are comparing against bits always shifted out, the
1252 // comparison cannot succeed.
1254 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1256 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1257 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1259 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1260 return ReplaceInstUsesWith(ICI, Cst);
1263 // If the shift is NUW, then it is just shifting out zeros, no need for an
1265 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1266 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1267 ConstantExpr::getLShr(RHS, ShAmt));
1269 if (LHSI->hasOneUse()) {
1270 // Otherwise strength reduce the shift into an and.
1271 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1273 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1274 TypeBits-ShAmtVal));
1277 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1278 return new ICmpInst(ICI.getPredicate(), And,
1279 ConstantExpr::getLShr(RHS, ShAmt));
1283 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1284 bool TrueIfSigned = false;
1285 if (LHSI->hasOneUse() &&
1286 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1287 // (X << 31) <s 0 --> (X&1) != 0
1288 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1289 APInt::getOneBitSet(TypeBits,
1290 TypeBits-ShAmt->getZExtValue()-1));
1292 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1293 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1294 And, Constant::getNullValue(And->getType()));
1299 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1300 case Instruction::AShr: {
1301 // Handle equality comparisons of shift-by-constant.
1302 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1303 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1304 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1308 // Handle exact shr's.
1309 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1310 if (RHSV.isMinValue())
1311 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1316 case Instruction::SDiv:
1317 case Instruction::UDiv:
1318 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1319 // Fold this div into the comparison, producing a range check.
1320 // Determine, based on the divide type, what the range is being
1321 // checked. If there is an overflow on the low or high side, remember
1322 // it, otherwise compute the range [low, hi) bounding the new value.
1323 // See: InsertRangeTest above for the kinds of replacements possible.
1324 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1325 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1330 case Instruction::Add:
1331 // Fold: icmp pred (add X, C1), C2
1332 if (!ICI.isEquality()) {
1333 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1335 const APInt &LHSV = LHSC->getValue();
1337 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1340 if (ICI.isSigned()) {
1341 if (CR.getLower().isSignBit()) {
1342 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1343 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1344 } else if (CR.getUpper().isSignBit()) {
1345 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1346 ConstantInt::get(ICI.getContext(),CR.getLower()));
1349 if (CR.getLower().isMinValue()) {
1350 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1351 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1352 } else if (CR.getUpper().isMinValue()) {
1353 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1354 ConstantInt::get(ICI.getContext(),CR.getLower()));
1361 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1362 if (ICI.isEquality()) {
1363 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1365 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1366 // the second operand is a constant, simplify a bit.
1367 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1368 switch (BO->getOpcode()) {
1369 case Instruction::SRem:
1370 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1371 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1372 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1373 if (V.sgt(1) && V.isPowerOf2()) {
1375 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1377 return new ICmpInst(ICI.getPredicate(), NewRem,
1378 Constant::getNullValue(BO->getType()));
1382 case Instruction::Add:
1383 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1384 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1385 if (BO->hasOneUse())
1386 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1387 ConstantExpr::getSub(RHS, BOp1C));
1388 } else if (RHSV == 0) {
1389 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1390 // efficiently invertible, or if the add has just this one use.
1391 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1393 if (Value *NegVal = dyn_castNegVal(BOp1))
1394 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1395 if (Value *NegVal = dyn_castNegVal(BOp0))
1396 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1397 if (BO->hasOneUse()) {
1398 Value *Neg = Builder->CreateNeg(BOp1);
1400 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1404 case Instruction::Xor:
1405 // For the xor case, we can xor two constants together, eliminating
1406 // the explicit xor.
1407 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1408 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1409 ConstantExpr::getXor(RHS, BOC));
1410 } else if (RHSV == 0) {
1411 // Replace ((xor A, B) != 0) with (A != B)
1412 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1416 case Instruction::Sub:
1417 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1418 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1419 if (BO->hasOneUse())
1420 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1421 ConstantExpr::getSub(BOp0C, RHS));
1422 } else if (RHSV == 0) {
1423 // Replace ((sub A, B) != 0) with (A != B)
1424 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1428 case Instruction::Or:
1429 // If bits are being or'd in that are not present in the constant we
1430 // are comparing against, then the comparison could never succeed!
1431 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1432 Constant *NotCI = ConstantExpr::getNot(RHS);
1433 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1434 return ReplaceInstUsesWith(ICI,
1435 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1440 case Instruction::And:
1441 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1442 // If bits are being compared against that are and'd out, then the
1443 // comparison can never succeed!
1444 if ((RHSV & ~BOC->getValue()) != 0)
1445 return ReplaceInstUsesWith(ICI,
1446 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1449 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1450 if (RHS == BOC && RHSV.isPowerOf2())
1451 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1452 ICmpInst::ICMP_NE, LHSI,
1453 Constant::getNullValue(RHS->getType()));
1455 // Don't perform the following transforms if the AND has multiple uses
1456 if (!BO->hasOneUse())
1459 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1460 if (BOC->getValue().isSignBit()) {
1461 Value *X = BO->getOperand(0);
1462 Constant *Zero = Constant::getNullValue(X->getType());
1463 ICmpInst::Predicate pred = isICMP_NE ?
1464 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1465 return new ICmpInst(pred, X, Zero);
1468 // ((X & ~7) == 0) --> X < 8
1469 if (RHSV == 0 && isHighOnes(BOC)) {
1470 Value *X = BO->getOperand(0);
1471 Constant *NegX = ConstantExpr::getNeg(BOC);
1472 ICmpInst::Predicate pred = isICMP_NE ?
1473 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1474 return new ICmpInst(pred, X, NegX);
1479 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1480 // Handle icmp {eq|ne} <intrinsic>, intcst.
1481 switch (II->getIntrinsicID()) {
1482 case Intrinsic::bswap:
1484 ICI.setOperand(0, II->getArgOperand(0));
1485 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1487 case Intrinsic::ctlz:
1488 case Intrinsic::cttz:
1489 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1490 if (RHSV == RHS->getType()->getBitWidth()) {
1492 ICI.setOperand(0, II->getArgOperand(0));
1493 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1497 case Intrinsic::ctpop:
1498 // popcount(A) == 0 -> A == 0 and likewise for !=
1499 if (RHS->isZero()) {
1501 ICI.setOperand(0, II->getArgOperand(0));
1502 ICI.setOperand(1, RHS);
1514 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1515 /// We only handle extending casts so far.
1517 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1518 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1519 Value *LHSCIOp = LHSCI->getOperand(0);
1520 Type *SrcTy = LHSCIOp->getType();
1521 Type *DestTy = LHSCI->getType();
1524 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1525 // integer type is the same size as the pointer type.
1526 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1527 TD->getPointerSizeInBits() ==
1528 cast<IntegerType>(DestTy)->getBitWidth()) {
1530 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1531 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1532 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1533 RHSOp = RHSC->getOperand(0);
1534 // If the pointer types don't match, insert a bitcast.
1535 if (LHSCIOp->getType() != RHSOp->getType())
1536 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1540 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1543 // The code below only handles extension cast instructions, so far.
1545 if (LHSCI->getOpcode() != Instruction::ZExt &&
1546 LHSCI->getOpcode() != Instruction::SExt)
1549 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1550 bool isSignedCmp = ICI.isSigned();
1552 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1553 // Not an extension from the same type?
1554 RHSCIOp = CI->getOperand(0);
1555 if (RHSCIOp->getType() != LHSCIOp->getType())
1558 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1559 // and the other is a zext), then we can't handle this.
1560 if (CI->getOpcode() != LHSCI->getOpcode())
1563 // Deal with equality cases early.
1564 if (ICI.isEquality())
1565 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1567 // A signed comparison of sign extended values simplifies into a
1568 // signed comparison.
1569 if (isSignedCmp && isSignedExt)
1570 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1572 // The other three cases all fold into an unsigned comparison.
1573 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1576 // If we aren't dealing with a constant on the RHS, exit early
1577 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1581 // Compute the constant that would happen if we truncated to SrcTy then
1582 // reextended to DestTy.
1583 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1584 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1587 // If the re-extended constant didn't change...
1589 // Deal with equality cases early.
1590 if (ICI.isEquality())
1591 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1593 // A signed comparison of sign extended values simplifies into a
1594 // signed comparison.
1595 if (isSignedExt && isSignedCmp)
1596 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1598 // The other three cases all fold into an unsigned comparison.
1599 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1602 // The re-extended constant changed so the constant cannot be represented
1603 // in the shorter type. Consequently, we cannot emit a simple comparison.
1604 // All the cases that fold to true or false will have already been handled
1605 // by SimplifyICmpInst, so only deal with the tricky case.
1607 if (isSignedCmp || !isSignedExt)
1610 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1611 // should have been folded away previously and not enter in here.
1613 // We're performing an unsigned comp with a sign extended value.
1614 // This is true if the input is >= 0. [aka >s -1]
1615 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1616 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1618 // Finally, return the value computed.
1619 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1620 return ReplaceInstUsesWith(ICI, Result);
1622 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1623 return BinaryOperator::CreateNot(Result);
1626 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1627 /// I = icmp ugt (add (add A, B), CI2), CI1
1628 /// If this is of the form:
1630 /// if (sum+128 >u 255)
1631 /// Then replace it with llvm.sadd.with.overflow.i8.
1633 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1634 ConstantInt *CI2, ConstantInt *CI1,
1636 // The transformation we're trying to do here is to transform this into an
1637 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1638 // with a narrower add, and discard the add-with-constant that is part of the
1639 // range check (if we can't eliminate it, this isn't profitable).
1641 // In order to eliminate the add-with-constant, the compare can be its only
1643 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1644 if (!AddWithCst->hasOneUse()) return 0;
1646 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1647 if (!CI2->getValue().isPowerOf2()) return 0;
1648 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1649 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1651 // The width of the new add formed is 1 more than the bias.
1654 // Check to see that CI1 is an all-ones value with NewWidth bits.
1655 if (CI1->getBitWidth() == NewWidth ||
1656 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1659 // In order to replace the original add with a narrower
1660 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1661 // and truncates that discard the high bits of the add. Verify that this is
1663 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1664 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1666 if (*UI == AddWithCst) continue;
1668 // Only accept truncates for now. We would really like a nice recursive
1669 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1670 // chain to see which bits of a value are actually demanded. If the
1671 // original add had another add which was then immediately truncated, we
1672 // could still do the transformation.
1673 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1675 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1678 // If the pattern matches, truncate the inputs to the narrower type and
1679 // use the sadd_with_overflow intrinsic to efficiently compute both the
1680 // result and the overflow bit.
1681 Module *M = I.getParent()->getParent()->getParent();
1683 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1684 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1687 InstCombiner::BuilderTy *Builder = IC.Builder;
1689 // Put the new code above the original add, in case there are any uses of the
1690 // add between the add and the compare.
1691 Builder->SetInsertPoint(OrigAdd);
1693 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1694 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1695 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1696 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1697 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1699 // The inner add was the result of the narrow add, zero extended to the
1700 // wider type. Replace it with the result computed by the intrinsic.
1701 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1703 // The original icmp gets replaced with the overflow value.
1704 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1707 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1709 // Don't bother doing this transformation for pointers, don't do it for
1711 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1713 // If the add is a constant expr, then we don't bother transforming it.
1714 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1715 if (OrigAdd == 0) return 0;
1717 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1719 // Put the new code above the original add, in case there are any uses of the
1720 // add between the add and the compare.
1721 InstCombiner::BuilderTy *Builder = IC.Builder;
1722 Builder->SetInsertPoint(OrigAdd);
1724 Module *M = I.getParent()->getParent()->getParent();
1725 Type *Ty = LHS->getType();
1726 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1727 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1728 Value *Add = Builder->CreateExtractValue(Call, 0);
1730 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1732 // The original icmp gets replaced with the overflow value.
1733 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1736 // DemandedBitsLHSMask - When performing a comparison against a constant,
1737 // it is possible that not all the bits in the LHS are demanded. This helper
1738 // method computes the mask that IS demanded.
1739 static APInt DemandedBitsLHSMask(ICmpInst &I,
1740 unsigned BitWidth, bool isSignCheck) {
1742 return APInt::getSignBit(BitWidth);
1744 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1745 if (!CI) return APInt::getAllOnesValue(BitWidth);
1746 const APInt &RHS = CI->getValue();
1748 switch (I.getPredicate()) {
1749 // For a UGT comparison, we don't care about any bits that
1750 // correspond to the trailing ones of the comparand. The value of these
1751 // bits doesn't impact the outcome of the comparison, because any value
1752 // greater than the RHS must differ in a bit higher than these due to carry.
1753 case ICmpInst::ICMP_UGT: {
1754 unsigned trailingOnes = RHS.countTrailingOnes();
1755 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1759 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1760 // Any value less than the RHS must differ in a higher bit because of carries.
1761 case ICmpInst::ICMP_ULT: {
1762 unsigned trailingZeros = RHS.countTrailingZeros();
1763 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1768 return APInt::getAllOnesValue(BitWidth);
1773 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1774 bool Changed = false;
1775 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1777 /// Orders the operands of the compare so that they are listed from most
1778 /// complex to least complex. This puts constants before unary operators,
1779 /// before binary operators.
1780 if (getComplexity(Op0) < getComplexity(Op1)) {
1782 std::swap(Op0, Op1);
1786 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1787 return ReplaceInstUsesWith(I, V);
1789 Type *Ty = Op0->getType();
1791 // icmp's with boolean values can always be turned into bitwise operations
1792 if (Ty->isIntegerTy(1)) {
1793 switch (I.getPredicate()) {
1794 default: llvm_unreachable("Invalid icmp instruction!");
1795 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1796 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1797 return BinaryOperator::CreateNot(Xor);
1799 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1800 return BinaryOperator::CreateXor(Op0, Op1);
1802 case ICmpInst::ICMP_UGT:
1803 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1805 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1806 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1807 return BinaryOperator::CreateAnd(Not, Op1);
1809 case ICmpInst::ICMP_SGT:
1810 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1812 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1813 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1814 return BinaryOperator::CreateAnd(Not, Op0);
1816 case ICmpInst::ICMP_UGE:
1817 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1819 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1820 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1821 return BinaryOperator::CreateOr(Not, Op1);
1823 case ICmpInst::ICMP_SGE:
1824 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1826 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1827 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1828 return BinaryOperator::CreateOr(Not, Op0);
1833 unsigned BitWidth = 0;
1834 if (Ty->isIntOrIntVectorTy())
1835 BitWidth = Ty->getScalarSizeInBits();
1836 else if (TD) // Pointers require TD info to get their size.
1837 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1839 bool isSignBit = false;
1841 // See if we are doing a comparison with a constant.
1842 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1843 Value *A = 0, *B = 0;
1845 // Match the following pattern, which is a common idiom when writing
1846 // overflow-safe integer arithmetic function. The source performs an
1847 // addition in wider type, and explicitly checks for overflow using
1848 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1849 // sadd_with_overflow intrinsic.
1851 // TODO: This could probably be generalized to handle other overflow-safe
1852 // operations if we worked out the formulas to compute the appropriate
1856 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1858 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1859 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1860 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1861 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1865 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1866 if (I.isEquality() && CI->isZero() &&
1867 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1868 // (icmp cond A B) if cond is equality
1869 return new ICmpInst(I.getPredicate(), A, B);
1872 // If we have an icmp le or icmp ge instruction, turn it into the
1873 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1874 // them being folded in the code below. The SimplifyICmpInst code has
1875 // already handled the edge cases for us, so we just assert on them.
1876 switch (I.getPredicate()) {
1878 case ICmpInst::ICMP_ULE:
1879 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1880 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1881 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1882 case ICmpInst::ICMP_SLE:
1883 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1884 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1885 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1886 case ICmpInst::ICMP_UGE:
1887 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1888 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1889 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1890 case ICmpInst::ICMP_SGE:
1891 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1892 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1893 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1896 // If this comparison is a normal comparison, it demands all
1897 // bits, if it is a sign bit comparison, it only demands the sign bit.
1899 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1902 // See if we can fold the comparison based on range information we can get
1903 // by checking whether bits are known to be zero or one in the input.
1904 if (BitWidth != 0) {
1905 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1906 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1908 if (SimplifyDemandedBits(I.getOperandUse(0),
1909 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1910 Op0KnownZero, Op0KnownOne, 0))
1912 if (SimplifyDemandedBits(I.getOperandUse(1),
1913 APInt::getAllOnesValue(BitWidth),
1914 Op1KnownZero, Op1KnownOne, 0))
1917 // Given the known and unknown bits, compute a range that the LHS could be
1918 // in. Compute the Min, Max and RHS values based on the known bits. For the
1919 // EQ and NE we use unsigned values.
1920 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1921 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1923 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1925 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1928 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1930 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1934 // If Min and Max are known to be the same, then SimplifyDemandedBits
1935 // figured out that the LHS is a constant. Just constant fold this now so
1936 // that code below can assume that Min != Max.
1937 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1938 return new ICmpInst(I.getPredicate(),
1939 ConstantInt::get(Op0->getType(), Op0Min), Op1);
1940 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1941 return new ICmpInst(I.getPredicate(), Op0,
1942 ConstantInt::get(Op1->getType(), Op1Min));
1944 // Based on the range information we know about the LHS, see if we can
1945 // simplify this comparison. For example, (x&4) < 8 is always true.
1946 switch (I.getPredicate()) {
1947 default: llvm_unreachable("Unknown icmp opcode!");
1948 case ICmpInst::ICMP_EQ: {
1949 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1950 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
1952 // If all bits are known zero except for one, then we know at most one
1953 // bit is set. If the comparison is against zero, then this is a check
1954 // to see if *that* bit is set.
1955 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1956 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1957 // If the LHS is an AND with the same constant, look through it.
1959 ConstantInt *LHSC = 0;
1960 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1961 LHSC->getValue() != Op0KnownZeroInverted)
1964 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1965 // then turn "((1 << x)&8) == 0" into "x != 3".
1967 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1968 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1969 return new ICmpInst(ICmpInst::ICMP_NE, X,
1970 ConstantInt::get(X->getType(), CmpVal));
1973 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1974 // then turn "((8 >>u x)&1) == 0" into "x != 3".
1976 if (Op0KnownZeroInverted == 1 &&
1977 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1978 return new ICmpInst(ICmpInst::ICMP_NE, X,
1979 ConstantInt::get(X->getType(),
1980 CI->countTrailingZeros()));
1985 case ICmpInst::ICMP_NE: {
1986 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1987 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
1989 // If all bits are known zero except for one, then we know at most one
1990 // bit is set. If the comparison is against zero, then this is a check
1991 // to see if *that* bit is set.
1992 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1993 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1994 // If the LHS is an AND with the same constant, look through it.
1996 ConstantInt *LHSC = 0;
1997 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1998 LHSC->getValue() != Op0KnownZeroInverted)
2001 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2002 // then turn "((1 << x)&8) != 0" into "x == 3".
2004 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2005 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2006 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2007 ConstantInt::get(X->getType(), CmpVal));
2010 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2011 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2013 if (Op0KnownZeroInverted == 1 &&
2014 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2015 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2016 ConstantInt::get(X->getType(),
2017 CI->countTrailingZeros()));
2022 case ICmpInst::ICMP_ULT:
2023 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2024 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2025 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2026 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2027 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2028 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2029 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2030 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2031 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2032 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2034 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2035 if (CI->isMinValue(true))
2036 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2037 Constant::getAllOnesValue(Op0->getType()));
2040 case ICmpInst::ICMP_UGT:
2041 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2042 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2043 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2044 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2046 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2047 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2048 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2049 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2050 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2051 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2053 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2054 if (CI->isMaxValue(true))
2055 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2056 Constant::getNullValue(Op0->getType()));
2059 case ICmpInst::ICMP_SLT:
2060 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2061 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2062 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2063 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2064 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2065 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2066 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2067 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2068 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2069 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2072 case ICmpInst::ICMP_SGT:
2073 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2074 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2075 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2076 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2078 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2079 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2080 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2081 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2082 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2083 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2086 case ICmpInst::ICMP_SGE:
2087 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2088 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2089 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2090 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2091 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2093 case ICmpInst::ICMP_SLE:
2094 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2095 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2096 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2097 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2098 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2100 case ICmpInst::ICMP_UGE:
2101 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2102 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2103 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2104 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2105 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2107 case ICmpInst::ICMP_ULE:
2108 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2109 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2110 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2111 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2112 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2116 // Turn a signed comparison into an unsigned one if both operands
2117 // are known to have the same sign.
2119 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2120 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2121 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2124 // Test if the ICmpInst instruction is used exclusively by a select as
2125 // part of a minimum or maximum operation. If so, refrain from doing
2126 // any other folding. This helps out other analyses which understand
2127 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2128 // and CodeGen. And in this case, at least one of the comparison
2129 // operands has at least one user besides the compare (the select),
2130 // which would often largely negate the benefit of folding anyway.
2132 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2133 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2134 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2137 // See if we are doing a comparison between a constant and an instruction that
2138 // can be folded into the comparison.
2139 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2140 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2141 // instruction, see if that instruction also has constants so that the
2142 // instruction can be folded into the icmp
2143 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2144 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2148 // Handle icmp with constant (but not simple integer constant) RHS
2149 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2150 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2151 switch (LHSI->getOpcode()) {
2152 case Instruction::GetElementPtr:
2153 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2154 if (RHSC->isNullValue() &&
2155 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2156 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2157 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2159 case Instruction::PHI:
2160 // Only fold icmp into the PHI if the phi and icmp are in the same
2161 // block. If in the same block, we're encouraging jump threading. If
2162 // not, we are just pessimizing the code by making an i1 phi.
2163 if (LHSI->getParent() == I.getParent())
2164 if (Instruction *NV = FoldOpIntoPhi(I))
2167 case Instruction::Select: {
2168 // If either operand of the select is a constant, we can fold the
2169 // comparison into the select arms, which will cause one to be
2170 // constant folded and the select turned into a bitwise or.
2171 Value *Op1 = 0, *Op2 = 0;
2172 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2173 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2174 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2175 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2177 // We only want to perform this transformation if it will not lead to
2178 // additional code. This is true if either both sides of the select
2179 // fold to a constant (in which case the icmp is replaced with a select
2180 // which will usually simplify) or this is the only user of the
2181 // select (in which case we are trading a select+icmp for a simpler
2183 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2185 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2188 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2190 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2194 case Instruction::IntToPtr:
2195 // icmp pred inttoptr(X), null -> icmp pred X, 0
2196 if (RHSC->isNullValue() && TD &&
2197 TD->getIntPtrType(RHSC->getContext()) ==
2198 LHSI->getOperand(0)->getType())
2199 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2200 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2203 case Instruction::Load:
2204 // Try to optimize things like "A[i] > 4" to index computations.
2205 if (GetElementPtrInst *GEP =
2206 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2207 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2208 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2209 !cast<LoadInst>(LHSI)->isVolatile())
2210 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2217 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2218 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2219 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2221 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2222 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2223 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2226 // Test to see if the operands of the icmp are casted versions of other
2227 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2229 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2230 if (Op0->getType()->isPointerTy() &&
2231 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2232 // We keep moving the cast from the left operand over to the right
2233 // operand, where it can often be eliminated completely.
2234 Op0 = CI->getOperand(0);
2236 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2237 // so eliminate it as well.
2238 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2239 Op1 = CI2->getOperand(0);
2241 // If Op1 is a constant, we can fold the cast into the constant.
2242 if (Op0->getType() != Op1->getType()) {
2243 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2244 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2246 // Otherwise, cast the RHS right before the icmp
2247 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2250 return new ICmpInst(I.getPredicate(), Op0, Op1);
2254 if (isa<CastInst>(Op0)) {
2255 // Handle the special case of: icmp (cast bool to X), <cst>
2256 // This comes up when you have code like
2259 // For generality, we handle any zero-extension of any operand comparison
2260 // with a constant or another cast from the same type.
2261 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2262 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2266 // Special logic for binary operators.
2267 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2268 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2270 CmpInst::Predicate Pred = I.getPredicate();
2271 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2272 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2273 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2274 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2275 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2276 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2277 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2278 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2279 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2281 // Analyze the case when either Op0 or Op1 is an add instruction.
2282 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2283 Value *A = 0, *B = 0, *C = 0, *D = 0;
2284 if (BO0 && BO0->getOpcode() == Instruction::Add)
2285 A = BO0->getOperand(0), B = BO0->getOperand(1);
2286 if (BO1 && BO1->getOpcode() == Instruction::Add)
2287 C = BO1->getOperand(0), D = BO1->getOperand(1);
2289 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2290 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2291 return new ICmpInst(Pred, A == Op1 ? B : A,
2292 Constant::getNullValue(Op1->getType()));
2294 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2295 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2296 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2299 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2300 if (A && C && (A == C || A == D || B == C || B == D) &&
2301 NoOp0WrapProblem && NoOp1WrapProblem &&
2302 // Try not to increase register pressure.
2303 BO0->hasOneUse() && BO1->hasOneUse()) {
2304 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2305 Value *Y = (A == C || A == D) ? B : A;
2306 Value *Z = (C == A || C == B) ? D : C;
2307 return new ICmpInst(Pred, Y, Z);
2310 // Analyze the case when either Op0 or Op1 is a sub instruction.
2311 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2312 A = 0; B = 0; C = 0; D = 0;
2313 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2314 A = BO0->getOperand(0), B = BO0->getOperand(1);
2315 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2316 C = BO1->getOperand(0), D = BO1->getOperand(1);
2318 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2319 if (A == Op1 && NoOp0WrapProblem)
2320 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2322 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2323 if (C == Op0 && NoOp1WrapProblem)
2324 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2326 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2327 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2328 // Try not to increase register pressure.
2329 BO0->hasOneUse() && BO1->hasOneUse())
2330 return new ICmpInst(Pred, A, C);
2332 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2333 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2334 // Try not to increase register pressure.
2335 BO0->hasOneUse() && BO1->hasOneUse())
2336 return new ICmpInst(Pred, D, B);
2338 BinaryOperator *SRem = NULL;
2339 // icmp (srem X, Y), Y
2340 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2341 Op1 == BO0->getOperand(1))
2343 // icmp Y, (srem X, Y)
2344 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2345 Op0 == BO1->getOperand(1))
2348 // We don't check hasOneUse to avoid increasing register pressure because
2349 // the value we use is the same value this instruction was already using.
2350 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2352 case ICmpInst::ICMP_EQ:
2353 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2354 case ICmpInst::ICMP_NE:
2355 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2356 case ICmpInst::ICMP_SGT:
2357 case ICmpInst::ICMP_SGE:
2358 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2359 Constant::getAllOnesValue(SRem->getType()));
2360 case ICmpInst::ICMP_SLT:
2361 case ICmpInst::ICMP_SLE:
2362 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2363 Constant::getNullValue(SRem->getType()));
2367 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2368 BO0->hasOneUse() && BO1->hasOneUse() &&
2369 BO0->getOperand(1) == BO1->getOperand(1)) {
2370 switch (BO0->getOpcode()) {
2372 case Instruction::Add:
2373 case Instruction::Sub:
2374 case Instruction::Xor:
2375 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2376 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2377 BO1->getOperand(0));
2378 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2379 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2380 if (CI->getValue().isSignBit()) {
2381 ICmpInst::Predicate Pred = I.isSigned()
2382 ? I.getUnsignedPredicate()
2383 : I.getSignedPredicate();
2384 return new ICmpInst(Pred, BO0->getOperand(0),
2385 BO1->getOperand(0));
2388 if (CI->isMaxValue(true)) {
2389 ICmpInst::Predicate Pred = I.isSigned()
2390 ? I.getUnsignedPredicate()
2391 : I.getSignedPredicate();
2392 Pred = I.getSwappedPredicate(Pred);
2393 return new ICmpInst(Pred, BO0->getOperand(0),
2394 BO1->getOperand(0));
2398 case Instruction::Mul:
2399 if (!I.isEquality())
2402 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2403 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2404 // Mask = -1 >> count-trailing-zeros(Cst).
2405 if (!CI->isZero() && !CI->isOne()) {
2406 const APInt &AP = CI->getValue();
2407 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2408 APInt::getLowBitsSet(AP.getBitWidth(),
2410 AP.countTrailingZeros()));
2411 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2412 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2413 return new ICmpInst(I.getPredicate(), And1, And2);
2417 case Instruction::UDiv:
2418 case Instruction::LShr:
2422 case Instruction::SDiv:
2423 case Instruction::AShr:
2424 if (!BO0->isExact() || !BO1->isExact())
2426 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2427 BO1->getOperand(0));
2428 case Instruction::Shl: {
2429 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2430 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2433 if (!NSW && I.isSigned())
2435 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2436 BO1->getOperand(0));
2443 // ~x < ~y --> y < x
2444 // ~x < cst --> ~cst < x
2445 if (match(Op0, m_Not(m_Value(A)))) {
2446 if (match(Op1, m_Not(m_Value(B))))
2447 return new ICmpInst(I.getPredicate(), B, A);
2448 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2449 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2452 // (a+b) <u a --> llvm.uadd.with.overflow.
2453 // (a+b) <u b --> llvm.uadd.with.overflow.
2454 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2455 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2456 (Op1 == A || Op1 == B))
2457 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2460 // a >u (a+b) --> llvm.uadd.with.overflow.
2461 // b >u (a+b) --> llvm.uadd.with.overflow.
2462 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2463 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2464 (Op0 == A || Op0 == B))
2465 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2469 if (I.isEquality()) {
2470 Value *A, *B, *C, *D;
2472 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2473 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2474 Value *OtherVal = A == Op1 ? B : A;
2475 return new ICmpInst(I.getPredicate(), OtherVal,
2476 Constant::getNullValue(A->getType()));
2479 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2480 // A^c1 == C^c2 --> A == C^(c1^c2)
2481 ConstantInt *C1, *C2;
2482 if (match(B, m_ConstantInt(C1)) &&
2483 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2484 Constant *NC = ConstantInt::get(I.getContext(),
2485 C1->getValue() ^ C2->getValue());
2486 Value *Xor = Builder->CreateXor(C, NC, "tmp");
2487 return new ICmpInst(I.getPredicate(), A, Xor);
2490 // A^B == A^D -> B == D
2491 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2492 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2493 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2494 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2498 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2499 (A == Op0 || B == Op0)) {
2500 // A == (A^B) -> B == 0
2501 Value *OtherVal = A == Op0 ? B : A;
2502 return new ICmpInst(I.getPredicate(), OtherVal,
2503 Constant::getNullValue(A->getType()));
2506 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2507 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2508 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2509 Value *X = 0, *Y = 0, *Z = 0;
2512 X = B; Y = D; Z = A;
2513 } else if (A == D) {
2514 X = B; Y = C; Z = A;
2515 } else if (B == C) {
2516 X = A; Y = D; Z = B;
2517 } else if (B == D) {
2518 X = A; Y = C; Z = B;
2521 if (X) { // Build (X^Y) & Z
2522 Op1 = Builder->CreateXor(X, Y, "tmp");
2523 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
2524 I.setOperand(0, Op1);
2525 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2530 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2531 // "icmp (and X, mask), cst"
2534 if (Op0->hasOneUse() &&
2535 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2536 m_ConstantInt(ShAmt))))) &&
2537 match(Op1, m_ConstantInt(Cst1)) &&
2538 // Only do this when A has multiple uses. This is most important to do
2539 // when it exposes other optimizations.
2541 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2543 if (ShAmt < ASize) {
2545 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2548 APInt CmpV = Cst1->getValue().zext(ASize);
2551 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2552 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2558 Value *X; ConstantInt *Cst;
2560 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2561 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2564 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2565 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2567 return Changed ? &I : 0;
2575 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2577 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2580 if (!isa<ConstantFP>(RHSC)) return 0;
2581 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2583 // Get the width of the mantissa. We don't want to hack on conversions that
2584 // might lose information from the integer, e.g. "i64 -> float"
2585 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2586 if (MantissaWidth == -1) return 0; // Unknown.
2588 // Check to see that the input is converted from an integer type that is small
2589 // enough that preserves all bits. TODO: check here for "known" sign bits.
2590 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2591 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2593 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2594 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2598 // If the conversion would lose info, don't hack on this.
2599 if ((int)InputSize > MantissaWidth)
2602 // Otherwise, we can potentially simplify the comparison. We know that it
2603 // will always come through as an integer value and we know the constant is
2604 // not a NAN (it would have been previously simplified).
2605 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2607 ICmpInst::Predicate Pred;
2608 switch (I.getPredicate()) {
2609 default: llvm_unreachable("Unexpected predicate!");
2610 case FCmpInst::FCMP_UEQ:
2611 case FCmpInst::FCMP_OEQ:
2612 Pred = ICmpInst::ICMP_EQ;
2614 case FCmpInst::FCMP_UGT:
2615 case FCmpInst::FCMP_OGT:
2616 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2618 case FCmpInst::FCMP_UGE:
2619 case FCmpInst::FCMP_OGE:
2620 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2622 case FCmpInst::FCMP_ULT:
2623 case FCmpInst::FCMP_OLT:
2624 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2626 case FCmpInst::FCMP_ULE:
2627 case FCmpInst::FCMP_OLE:
2628 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2630 case FCmpInst::FCMP_UNE:
2631 case FCmpInst::FCMP_ONE:
2632 Pred = ICmpInst::ICMP_NE;
2634 case FCmpInst::FCMP_ORD:
2635 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2636 case FCmpInst::FCMP_UNO:
2637 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2640 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2642 // Now we know that the APFloat is a normal number, zero or inf.
2644 // See if the FP constant is too large for the integer. For example,
2645 // comparing an i8 to 300.0.
2646 unsigned IntWidth = IntTy->getScalarSizeInBits();
2649 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2650 // and large values.
2651 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2652 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2653 APFloat::rmNearestTiesToEven);
2654 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2655 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2656 Pred == ICmpInst::ICMP_SLE)
2657 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2658 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2661 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2662 // +INF and large values.
2663 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2664 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2665 APFloat::rmNearestTiesToEven);
2666 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2667 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2668 Pred == ICmpInst::ICMP_ULE)
2669 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2670 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2675 // See if the RHS value is < SignedMin.
2676 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2677 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2678 APFloat::rmNearestTiesToEven);
2679 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2680 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2681 Pred == ICmpInst::ICMP_SGE)
2682 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2683 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2687 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2688 // [0, UMAX], but it may still be fractional. See if it is fractional by
2689 // casting the FP value to the integer value and back, checking for equality.
2690 // Don't do this for zero, because -0.0 is not fractional.
2691 Constant *RHSInt = LHSUnsigned
2692 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2693 : ConstantExpr::getFPToSI(RHSC, IntTy);
2694 if (!RHS.isZero()) {
2695 bool Equal = LHSUnsigned
2696 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2697 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2699 // If we had a comparison against a fractional value, we have to adjust
2700 // the compare predicate and sometimes the value. RHSC is rounded towards
2701 // zero at this point.
2703 default: llvm_unreachable("Unexpected integer comparison!");
2704 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2705 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2706 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2707 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2708 case ICmpInst::ICMP_ULE:
2709 // (float)int <= 4.4 --> int <= 4
2710 // (float)int <= -4.4 --> false
2711 if (RHS.isNegative())
2712 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2714 case ICmpInst::ICMP_SLE:
2715 // (float)int <= 4.4 --> int <= 4
2716 // (float)int <= -4.4 --> int < -4
2717 if (RHS.isNegative())
2718 Pred = ICmpInst::ICMP_SLT;
2720 case ICmpInst::ICMP_ULT:
2721 // (float)int < -4.4 --> false
2722 // (float)int < 4.4 --> int <= 4
2723 if (RHS.isNegative())
2724 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2725 Pred = ICmpInst::ICMP_ULE;
2727 case ICmpInst::ICMP_SLT:
2728 // (float)int < -4.4 --> int < -4
2729 // (float)int < 4.4 --> int <= 4
2730 if (!RHS.isNegative())
2731 Pred = ICmpInst::ICMP_SLE;
2733 case ICmpInst::ICMP_UGT:
2734 // (float)int > 4.4 --> int > 4
2735 // (float)int > -4.4 --> true
2736 if (RHS.isNegative())
2737 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2739 case ICmpInst::ICMP_SGT:
2740 // (float)int > 4.4 --> int > 4
2741 // (float)int > -4.4 --> int >= -4
2742 if (RHS.isNegative())
2743 Pred = ICmpInst::ICMP_SGE;
2745 case ICmpInst::ICMP_UGE:
2746 // (float)int >= -4.4 --> true
2747 // (float)int >= 4.4 --> int > 4
2748 if (!RHS.isNegative())
2749 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2750 Pred = ICmpInst::ICMP_UGT;
2752 case ICmpInst::ICMP_SGE:
2753 // (float)int >= -4.4 --> int >= -4
2754 // (float)int >= 4.4 --> int > 4
2755 if (!RHS.isNegative())
2756 Pred = ICmpInst::ICMP_SGT;
2762 // Lower this FP comparison into an appropriate integer version of the
2764 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2767 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2768 bool Changed = false;
2770 /// Orders the operands of the compare so that they are listed from most
2771 /// complex to least complex. This puts constants before unary operators,
2772 /// before binary operators.
2773 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2778 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2780 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2781 return ReplaceInstUsesWith(I, V);
2783 // Simplify 'fcmp pred X, X'
2785 switch (I.getPredicate()) {
2786 default: llvm_unreachable("Unknown predicate!");
2787 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2788 case FCmpInst::FCMP_ULT: // True if unordered or less than
2789 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2790 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2791 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2792 I.setPredicate(FCmpInst::FCMP_UNO);
2793 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2796 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2797 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2798 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2799 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2800 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2801 I.setPredicate(FCmpInst::FCMP_ORD);
2802 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2807 // Handle fcmp with constant RHS
2808 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2809 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2810 switch (LHSI->getOpcode()) {
2811 case Instruction::FPExt: {
2812 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2813 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2814 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2818 // We can't convert a PPC double double.
2819 if (RHSF->getType()->isPPC_FP128Ty())
2822 const fltSemantics *Sem;
2823 // FIXME: This shouldn't be here.
2824 if (LHSExt->getSrcTy()->isFloatTy())
2825 Sem = &APFloat::IEEEsingle;
2826 else if (LHSExt->getSrcTy()->isDoubleTy())
2827 Sem = &APFloat::IEEEdouble;
2828 else if (LHSExt->getSrcTy()->isFP128Ty())
2829 Sem = &APFloat::IEEEquad;
2830 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2831 Sem = &APFloat::x87DoubleExtended;
2836 APFloat F = RHSF->getValueAPF();
2837 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2839 // Avoid lossy conversions and denormals.
2841 F.compare(APFloat::getSmallestNormalized(*Sem)) !=
2842 APFloat::cmpLessThan)
2843 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2844 ConstantFP::get(RHSC->getContext(), F));
2847 case Instruction::PHI:
2848 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2849 // block. If in the same block, we're encouraging jump threading. If
2850 // not, we are just pessimizing the code by making an i1 phi.
2851 if (LHSI->getParent() == I.getParent())
2852 if (Instruction *NV = FoldOpIntoPhi(I))
2855 case Instruction::SIToFP:
2856 case Instruction::UIToFP:
2857 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2860 case Instruction::Select: {
2861 // If either operand of the select is a constant, we can fold the
2862 // comparison into the select arms, which will cause one to be
2863 // constant folded and the select turned into a bitwise or.
2864 Value *Op1 = 0, *Op2 = 0;
2865 if (LHSI->hasOneUse()) {
2866 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2867 // Fold the known value into the constant operand.
2868 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2869 // Insert a new FCmp of the other select operand.
2870 Op2 = Builder->CreateFCmp(I.getPredicate(),
2871 LHSI->getOperand(2), RHSC, I.getName());
2872 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2873 // Fold the known value into the constant operand.
2874 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2875 // Insert a new FCmp of the other select operand.
2876 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2882 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2885 case Instruction::FSub: {
2886 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2888 if (match(LHSI, m_FNeg(m_Value(Op))))
2889 return new FCmpInst(I.getSwappedPredicate(), Op,
2890 ConstantExpr::getFNeg(RHSC));
2893 case Instruction::Load:
2894 if (GetElementPtrInst *GEP =
2895 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2896 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2897 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2898 !cast<LoadInst>(LHSI)->isVolatile())
2899 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2906 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
2908 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
2909 return new FCmpInst(I.getSwappedPredicate(), X, Y);
2911 // fcmp (fpext x), (fpext y) -> fcmp x, y
2912 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
2913 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
2914 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
2915 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2916 RHSExt->getOperand(0));
2918 return Changed ? &I : 0;