1 //===- InstCombineCompares.cpp --------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitICmp and visitFCmp functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/IntrinsicInst.h"
20 #include "llvm/Support/ConstantRange.h"
21 #include "llvm/Support/GetElementPtrTypeIterator.h"
22 #include "llvm/Support/PatternMatch.h"
23 #include "llvm/Target/TargetLibraryInfo.h"
25 using namespace PatternMatch;
27 static ConstantInt *getOne(Constant *C) {
28 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
31 /// AddOne - Add one to a ConstantInt
32 static Constant *AddOne(Constant *C) {
33 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
35 /// SubOne - Subtract one from a ConstantInt
36 static Constant *SubOne(Constant *C) {
37 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
40 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
41 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
44 static bool HasAddOverflow(ConstantInt *Result,
45 ConstantInt *In1, ConstantInt *In2,
48 return Result->getValue().ult(In1->getValue());
50 if (In2->isNegative())
51 return Result->getValue().sgt(In1->getValue());
52 return Result->getValue().slt(In1->getValue());
55 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
56 /// overflowed for this type.
57 static bool AddWithOverflow(Constant *&Result, Constant *In1,
58 Constant *In2, bool IsSigned = false) {
59 Result = ConstantExpr::getAdd(In1, In2);
61 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
62 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
63 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
64 if (HasAddOverflow(ExtractElement(Result, Idx),
65 ExtractElement(In1, Idx),
66 ExtractElement(In2, Idx),
73 return HasAddOverflow(cast<ConstantInt>(Result),
74 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
78 static bool HasSubOverflow(ConstantInt *Result,
79 ConstantInt *In1, ConstantInt *In2,
82 return Result->getValue().ugt(In1->getValue());
84 if (In2->isNegative())
85 return Result->getValue().slt(In1->getValue());
87 return Result->getValue().sgt(In1->getValue());
90 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
91 /// overflowed for this type.
92 static bool SubWithOverflow(Constant *&Result, Constant *In1,
93 Constant *In2, bool IsSigned = false) {
94 Result = ConstantExpr::getSub(In1, In2);
96 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
97 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
98 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
99 if (HasSubOverflow(ExtractElement(Result, Idx),
100 ExtractElement(In1, Idx),
101 ExtractElement(In2, Idx),
108 return HasSubOverflow(cast<ConstantInt>(Result),
109 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
113 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
114 /// comparison only checks the sign bit. If it only checks the sign bit, set
115 /// TrueIfSigned if the result of the comparison is true when the input value is
117 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
118 bool &TrueIfSigned) {
120 case ICmpInst::ICMP_SLT: // True if LHS s< 0
122 return RHS->isZero();
123 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
125 return RHS->isAllOnesValue();
126 case ICmpInst::ICMP_SGT: // True if LHS s> -1
127 TrueIfSigned = false;
128 return RHS->isAllOnesValue();
129 case ICmpInst::ICMP_UGT:
130 // True if LHS u> RHS and RHS == high-bit-mask - 1
132 return RHS->isMaxValue(true);
133 case ICmpInst::ICMP_UGE:
134 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
136 return RHS->getValue().isSignBit();
142 /// Returns true if the exploded icmp can be expressed as a signed comparison
143 /// to zero and updates the predicate accordingly.
144 /// The signedness of the comparison is preserved.
145 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
146 if (!ICmpInst::isSigned(pred))
150 return ICmpInst::isRelational(pred);
153 if (pred == ICmpInst::ICMP_SLT) {
154 pred = ICmpInst::ICMP_SLE;
157 } else if (RHS->isAllOnesValue()) {
158 if (pred == ICmpInst::ICMP_SGT) {
159 pred = ICmpInst::ICMP_SGE;
167 // isHighOnes - Return true if the constant is of the form 1+0+.
168 // This is the same as lowones(~X).
169 static bool isHighOnes(const ConstantInt *CI) {
170 return (~CI->getValue() + 1).isPowerOf2();
173 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
174 /// set of known zero and one bits, compute the maximum and minimum values that
175 /// could have the specified known zero and known one bits, returning them in
177 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
178 const APInt& KnownOne,
179 APInt& Min, APInt& Max) {
180 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
181 KnownZero.getBitWidth() == Min.getBitWidth() &&
182 KnownZero.getBitWidth() == Max.getBitWidth() &&
183 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
184 APInt UnknownBits = ~(KnownZero|KnownOne);
186 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
187 // bit if it is unknown.
189 Max = KnownOne|UnknownBits;
191 if (UnknownBits.isNegative()) { // Sign bit is unknown
192 Min.setBit(Min.getBitWidth()-1);
193 Max.clearBit(Max.getBitWidth()-1);
197 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
198 // a set of known zero and one bits, compute the maximum and minimum values that
199 // could have the specified known zero and known one bits, returning them in
201 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
202 const APInt &KnownOne,
203 APInt &Min, APInt &Max) {
204 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
205 KnownZero.getBitWidth() == Min.getBitWidth() &&
206 KnownZero.getBitWidth() == Max.getBitWidth() &&
207 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
208 APInt UnknownBits = ~(KnownZero|KnownOne);
210 // The minimum value is when the unknown bits are all zeros.
212 // The maximum value is when the unknown bits are all ones.
213 Max = KnownOne|UnknownBits;
218 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
219 /// cmp pred (load (gep GV, ...)), cmpcst
220 /// where GV is a global variable with a constant initializer. Try to simplify
221 /// this into some simple computation that does not need the load. For example
222 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
224 /// If AndCst is non-null, then the loaded value is masked with that constant
225 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
226 Instruction *InstCombiner::
227 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
228 CmpInst &ICI, ConstantInt *AndCst) {
229 // We need TD information to know the pointer size unless this is inbounds.
230 if (!GEP->isInBounds() && TD == 0)
233 Constant *Init = GV->getInitializer();
234 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
237 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
238 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays.
240 // There are many forms of this optimization we can handle, for now, just do
241 // the simple index into a single-dimensional array.
243 // Require: GEP GV, 0, i {{, constant indices}}
244 if (GEP->getNumOperands() < 3 ||
245 !isa<ConstantInt>(GEP->getOperand(1)) ||
246 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
247 isa<Constant>(GEP->getOperand(2)))
250 // Check that indices after the variable are constants and in-range for the
251 // type they index. Collect the indices. This is typically for arrays of
253 SmallVector<unsigned, 4> LaterIndices;
255 Type *EltTy = Init->getType()->getArrayElementType();
256 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
257 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
258 if (Idx == 0) return 0; // Variable index.
260 uint64_t IdxVal = Idx->getZExtValue();
261 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
263 if (StructType *STy = dyn_cast<StructType>(EltTy))
264 EltTy = STy->getElementType(IdxVal);
265 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
266 if (IdxVal >= ATy->getNumElements()) return 0;
267 EltTy = ATy->getElementType();
269 return 0; // Unknown type.
272 LaterIndices.push_back(IdxVal);
275 enum { Overdefined = -3, Undefined = -2 };
277 // Variables for our state machines.
279 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
280 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
281 // and 87 is the second (and last) index. FirstTrueElement is -2 when
282 // undefined, otherwise set to the first true element. SecondTrueElement is
283 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
284 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
286 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
287 // form "i != 47 & i != 87". Same state transitions as for true elements.
288 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
290 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
291 /// define a state machine that triggers for ranges of values that the index
292 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
293 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
294 /// index in the range (inclusive). We use -2 for undefined here because we
295 /// use relative comparisons and don't want 0-1 to match -1.
296 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
298 // MagicBitvector - This is a magic bitvector where we set a bit if the
299 // comparison is true for element 'i'. If there are 64 elements or less in
300 // the array, this will fully represent all the comparison results.
301 uint64_t MagicBitvector = 0;
304 // Scan the array and see if one of our patterns matches.
305 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
306 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
307 Constant *Elt = Init->getAggregateElement(i);
308 if (Elt == 0) return 0;
310 // If this is indexing an array of structures, get the structure element.
311 if (!LaterIndices.empty())
312 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
314 // If the element is masked, handle it.
315 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
317 // Find out if the comparison would be true or false for the i'th element.
318 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
319 CompareRHS, TD, TLI);
320 // If the result is undef for this element, ignore it.
321 if (isa<UndefValue>(C)) {
322 // Extend range state machines to cover this element in case there is an
323 // undef in the middle of the range.
324 if (TrueRangeEnd == (int)i-1)
326 if (FalseRangeEnd == (int)i-1)
331 // If we can't compute the result for any of the elements, we have to give
332 // up evaluating the entire conditional.
333 if (!isa<ConstantInt>(C)) return 0;
335 // Otherwise, we know if the comparison is true or false for this element,
336 // update our state machines.
337 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
339 // State machine for single/double/range index comparison.
341 // Update the TrueElement state machine.
342 if (FirstTrueElement == Undefined)
343 FirstTrueElement = TrueRangeEnd = i; // First true element.
345 // Update double-compare state machine.
346 if (SecondTrueElement == Undefined)
347 SecondTrueElement = i;
349 SecondTrueElement = Overdefined;
351 // Update range state machine.
352 if (TrueRangeEnd == (int)i-1)
355 TrueRangeEnd = Overdefined;
358 // Update the FalseElement state machine.
359 if (FirstFalseElement == Undefined)
360 FirstFalseElement = FalseRangeEnd = i; // First false element.
362 // Update double-compare state machine.
363 if (SecondFalseElement == Undefined)
364 SecondFalseElement = i;
366 SecondFalseElement = Overdefined;
368 // Update range state machine.
369 if (FalseRangeEnd == (int)i-1)
372 FalseRangeEnd = Overdefined;
377 // If this element is in range, update our magic bitvector.
378 if (i < 64 && IsTrueForElt)
379 MagicBitvector |= 1ULL << i;
381 // If all of our states become overdefined, bail out early. Since the
382 // predicate is expensive, only check it every 8 elements. This is only
383 // really useful for really huge arrays.
384 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
385 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
386 FalseRangeEnd == Overdefined)
390 // Now that we've scanned the entire array, emit our new comparison(s). We
391 // order the state machines in complexity of the generated code.
392 Value *Idx = GEP->getOperand(2);
394 // If the index is larger than the pointer size of the target, truncate the
395 // index down like the GEP would do implicitly. We don't have to do this for
396 // an inbounds GEP because the index can't be out of range.
397 if (!GEP->isInBounds()) {
398 Type *IntPtrTy = TD->getIntPtrType(GEP->getType());
399 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
400 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
401 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
404 // If the comparison is only true for one or two elements, emit direct
406 if (SecondTrueElement != Overdefined) {
407 // None true -> false.
408 if (FirstTrueElement == Undefined)
409 return ReplaceInstUsesWith(ICI, Builder->getFalse());
411 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
413 // True for one element -> 'i == 47'.
414 if (SecondTrueElement == Undefined)
415 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
417 // True for two elements -> 'i == 47 | i == 72'.
418 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
419 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
420 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
421 return BinaryOperator::CreateOr(C1, C2);
424 // If the comparison is only false for one or two elements, emit direct
426 if (SecondFalseElement != Overdefined) {
427 // None false -> true.
428 if (FirstFalseElement == Undefined)
429 return ReplaceInstUsesWith(ICI, Builder->getTrue());
431 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
433 // False for one element -> 'i != 47'.
434 if (SecondFalseElement == Undefined)
435 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
437 // False for two elements -> 'i != 47 & i != 72'.
438 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
439 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
440 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
441 return BinaryOperator::CreateAnd(C1, C2);
444 // If the comparison can be replaced with a range comparison for the elements
445 // where it is true, emit the range check.
446 if (TrueRangeEnd != Overdefined) {
447 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
449 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
450 if (FirstTrueElement) {
451 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
452 Idx = Builder->CreateAdd(Idx, Offs);
455 Value *End = ConstantInt::get(Idx->getType(),
456 TrueRangeEnd-FirstTrueElement+1);
457 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
460 // False range check.
461 if (FalseRangeEnd != Overdefined) {
462 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
463 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
464 if (FirstFalseElement) {
465 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
466 Idx = Builder->CreateAdd(Idx, Offs);
469 Value *End = ConstantInt::get(Idx->getType(),
470 FalseRangeEnd-FirstFalseElement);
471 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
475 // If a magic bitvector captures the entire comparison state
476 // of this load, replace it with computation that does:
477 // ((magic_cst >> i) & 1) != 0
481 // Look for an appropriate type:
482 // - The type of Idx if the magic fits
483 // - The smallest fitting legal type if we have a DataLayout
485 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
488 Ty = TD->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
489 else if (ArrayElementCount <= 32)
490 Ty = Type::getInt32Ty(Init->getContext());
493 Value *V = Builder->CreateIntCast(Idx, Ty, false);
494 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
495 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
496 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
504 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
505 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
506 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
507 /// be complex, and scales are involved. The above expression would also be
508 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
509 /// This later form is less amenable to optimization though, and we are allowed
510 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
512 /// If we can't emit an optimized form for this expression, this returns null.
514 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
515 DataLayout &TD = *IC.getDataLayout();
516 gep_type_iterator GTI = gep_type_begin(GEP);
518 // Check to see if this gep only has a single variable index. If so, and if
519 // any constant indices are a multiple of its scale, then we can compute this
520 // in terms of the scale of the variable index. For example, if the GEP
521 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
522 // because the expression will cross zero at the same point.
523 unsigned i, e = GEP->getNumOperands();
525 for (i = 1; i != e; ++i, ++GTI) {
526 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
527 // Compute the aggregate offset of constant indices.
528 if (CI->isZero()) continue;
530 // Handle a struct index, which adds its field offset to the pointer.
531 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
532 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
534 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
535 Offset += Size*CI->getSExtValue();
538 // Found our variable index.
543 // If there are no variable indices, we must have a constant offset, just
544 // evaluate it the general way.
545 if (i == e) return 0;
547 Value *VariableIdx = GEP->getOperand(i);
548 // Determine the scale factor of the variable element. For example, this is
549 // 4 if the variable index is into an array of i32.
550 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
552 // Verify that there are no other variable indices. If so, emit the hard way.
553 for (++i, ++GTI; i != e; ++i, ++GTI) {
554 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
557 // Compute the aggregate offset of constant indices.
558 if (CI->isZero()) continue;
560 // Handle a struct index, which adds its field offset to the pointer.
561 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
562 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
564 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
565 Offset += Size*CI->getSExtValue();
571 // Okay, we know we have a single variable index, which must be a
572 // pointer/array/vector index. If there is no offset, life is simple, return
574 Type *IntPtrTy = TD.getIntPtrType(GEP->getOperand(0)->getType());
575 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
577 // Cast to intptrty in case a truncation occurs. If an extension is needed,
578 // we don't need to bother extending: the extension won't affect where the
579 // computation crosses zero.
580 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
581 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
586 // Otherwise, there is an index. The computation we will do will be modulo
587 // the pointer size, so get it.
588 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
590 Offset &= PtrSizeMask;
591 VariableScale &= PtrSizeMask;
593 // To do this transformation, any constant index must be a multiple of the
594 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
595 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
596 // multiple of the variable scale.
597 int64_t NewOffs = Offset / (int64_t)VariableScale;
598 if (Offset != NewOffs*(int64_t)VariableScale)
601 // Okay, we can do this evaluation. Start by converting the index to intptr.
602 if (VariableIdx->getType() != IntPtrTy)
603 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
605 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
606 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
609 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
610 /// else. At this point we know that the GEP is on the LHS of the comparison.
611 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
612 ICmpInst::Predicate Cond,
614 // Don't transform signed compares of GEPs into index compares. Even if the
615 // GEP is inbounds, the final add of the base pointer can have signed overflow
616 // and would change the result of the icmp.
617 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
618 // the maximum signed value for the pointer type.
619 if (ICmpInst::isSigned(Cond))
622 // Look through bitcasts.
623 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
624 RHS = BCI->getOperand(0);
626 Value *PtrBase = GEPLHS->getOperand(0);
627 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
628 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
629 // This transformation (ignoring the base and scales) is valid because we
630 // know pointers can't overflow since the gep is inbounds. See if we can
631 // output an optimized form.
632 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
634 // If not, synthesize the offset the hard way.
636 Offset = EmitGEPOffset(GEPLHS);
637 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
638 Constant::getNullValue(Offset->getType()));
639 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
640 // If the base pointers are different, but the indices are the same, just
641 // compare the base pointer.
642 if (PtrBase != GEPRHS->getOperand(0)) {
643 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
644 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
645 GEPRHS->getOperand(0)->getType();
647 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
648 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
649 IndicesTheSame = false;
653 // If all indices are the same, just compare the base pointers.
655 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
657 // If we're comparing GEPs with two base pointers that only differ in type
658 // and both GEPs have only constant indices or just one use, then fold
659 // the compare with the adjusted indices.
660 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
661 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
662 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
663 PtrBase->stripPointerCasts() ==
664 GEPRHS->getOperand(0)->stripPointerCasts()) {
665 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
666 EmitGEPOffset(GEPLHS),
667 EmitGEPOffset(GEPRHS));
668 return ReplaceInstUsesWith(I, Cmp);
671 // Otherwise, the base pointers are different and the indices are
672 // different, bail out.
676 // If one of the GEPs has all zero indices, recurse.
677 bool AllZeros = true;
678 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
679 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
680 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
685 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
686 ICmpInst::getSwappedPredicate(Cond), I);
688 // If the other GEP has all zero indices, recurse.
690 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
691 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
692 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
697 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
699 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
700 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
701 // If the GEPs only differ by one index, compare it.
702 unsigned NumDifferences = 0; // Keep track of # differences.
703 unsigned DiffOperand = 0; // The operand that differs.
704 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
705 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
706 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
707 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
708 // Irreconcilable differences.
712 if (NumDifferences++) break;
717 if (NumDifferences == 0) // SAME GEP?
718 return ReplaceInstUsesWith(I, // No comparison is needed here.
719 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
721 else if (NumDifferences == 1 && GEPsInBounds) {
722 Value *LHSV = GEPLHS->getOperand(DiffOperand);
723 Value *RHSV = GEPRHS->getOperand(DiffOperand);
724 // Make sure we do a signed comparison here.
725 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
729 // Only lower this if the icmp is the only user of the GEP or if we expect
730 // the result to fold to a constant!
733 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
734 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
735 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
736 Value *L = EmitGEPOffset(GEPLHS);
737 Value *R = EmitGEPOffset(GEPRHS);
738 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
744 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
745 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
746 Value *X, ConstantInt *CI,
747 ICmpInst::Predicate Pred) {
748 // If we have X+0, exit early (simplifying logic below) and let it get folded
749 // elsewhere. icmp X+0, X -> icmp X, X
751 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
752 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
755 // (X+4) == X -> false.
756 if (Pred == ICmpInst::ICMP_EQ)
757 return ReplaceInstUsesWith(ICI, Builder->getFalse());
759 // (X+4) != X -> true.
760 if (Pred == ICmpInst::ICMP_NE)
761 return ReplaceInstUsesWith(ICI, Builder->getTrue());
763 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
764 // so the values can never be equal. Similarly for all other "or equals"
767 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
768 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
769 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
770 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
772 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
773 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
776 // (X+1) >u X --> X <u (0-1) --> X != 255
777 // (X+2) >u X --> X <u (0-2) --> X <u 254
778 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
779 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
780 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
782 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
783 ConstantInt *SMax = ConstantInt::get(X->getContext(),
784 APInt::getSignedMaxValue(BitWidth));
786 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
787 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
788 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
789 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
790 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
791 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
792 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
793 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
795 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
796 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
797 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
798 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
799 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
800 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
802 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
803 Constant *C = Builder->getInt(CI->getValue()-1);
804 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
807 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
808 /// and CmpRHS are both known to be integer constants.
809 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
810 ConstantInt *DivRHS) {
811 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
812 const APInt &CmpRHSV = CmpRHS->getValue();
814 // FIXME: If the operand types don't match the type of the divide
815 // then don't attempt this transform. The code below doesn't have the
816 // logic to deal with a signed divide and an unsigned compare (and
817 // vice versa). This is because (x /s C1) <s C2 produces different
818 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
819 // (x /u C1) <u C2. Simply casting the operands and result won't
820 // work. :( The if statement below tests that condition and bails
822 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
823 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
825 if (DivRHS->isZero())
826 return 0; // The ProdOV computation fails on divide by zero.
827 if (DivIsSigned && DivRHS->isAllOnesValue())
828 return 0; // The overflow computation also screws up here
829 if (DivRHS->isOne()) {
830 // This eliminates some funny cases with INT_MIN.
831 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
835 // Compute Prod = CI * DivRHS. We are essentially solving an equation
836 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
837 // C2 (CI). By solving for X we can turn this into a range check
838 // instead of computing a divide.
839 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
841 // Determine if the product overflows by seeing if the product is
842 // not equal to the divide. Make sure we do the same kind of divide
843 // as in the LHS instruction that we're folding.
844 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
845 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
847 // Get the ICmp opcode
848 ICmpInst::Predicate Pred = ICI.getPredicate();
850 /// If the division is known to be exact, then there is no remainder from the
851 /// divide, so the covered range size is unit, otherwise it is the divisor.
852 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
854 // Figure out the interval that is being checked. For example, a comparison
855 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
856 // Compute this interval based on the constants involved and the signedness of
857 // the compare/divide. This computes a half-open interval, keeping track of
858 // whether either value in the interval overflows. After analysis each
859 // overflow variable is set to 0 if it's corresponding bound variable is valid
860 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
861 int LoOverflow = 0, HiOverflow = 0;
862 Constant *LoBound = 0, *HiBound = 0;
864 if (!DivIsSigned) { // udiv
865 // e.g. X/5 op 3 --> [15, 20)
867 HiOverflow = LoOverflow = ProdOV;
869 // If this is not an exact divide, then many values in the range collapse
870 // to the same result value.
871 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
874 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
875 if (CmpRHSV == 0) { // (X / pos) op 0
876 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
877 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
879 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
880 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
881 HiOverflow = LoOverflow = ProdOV;
883 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
884 } else { // (X / pos) op neg
885 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
886 HiBound = AddOne(Prod);
887 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
889 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
890 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
893 } else if (DivRHS->isNegative()) { // Divisor is < 0.
895 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
896 if (CmpRHSV == 0) { // (X / neg) op 0
897 // e.g. X/-5 op 0 --> [-4, 5)
898 LoBound = AddOne(RangeSize);
899 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
900 if (HiBound == DivRHS) { // -INTMIN = INTMIN
901 HiOverflow = 1; // [INTMIN+1, overflow)
902 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
904 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
905 // e.g. X/-5 op 3 --> [-19, -14)
906 HiBound = AddOne(Prod);
907 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
909 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
910 } else { // (X / neg) op neg
911 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
912 LoOverflow = HiOverflow = ProdOV;
914 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
917 // Dividing by a negative swaps the condition. LT <-> GT
918 Pred = ICmpInst::getSwappedPredicate(Pred);
921 Value *X = DivI->getOperand(0);
923 default: llvm_unreachable("Unhandled icmp opcode!");
924 case ICmpInst::ICMP_EQ:
925 if (LoOverflow && HiOverflow)
926 return ReplaceInstUsesWith(ICI, Builder->getFalse());
928 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
929 ICmpInst::ICMP_UGE, X, LoBound);
931 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
932 ICmpInst::ICMP_ULT, X, HiBound);
933 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
935 case ICmpInst::ICMP_NE:
936 if (LoOverflow && HiOverflow)
937 return ReplaceInstUsesWith(ICI, Builder->getTrue());
939 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
940 ICmpInst::ICMP_ULT, X, LoBound);
942 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
943 ICmpInst::ICMP_UGE, X, HiBound);
944 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
945 DivIsSigned, false));
946 case ICmpInst::ICMP_ULT:
947 case ICmpInst::ICMP_SLT:
948 if (LoOverflow == +1) // Low bound is greater than input range.
949 return ReplaceInstUsesWith(ICI, Builder->getTrue());
950 if (LoOverflow == -1) // Low bound is less than input range.
951 return ReplaceInstUsesWith(ICI, Builder->getFalse());
952 return new ICmpInst(Pred, X, LoBound);
953 case ICmpInst::ICMP_UGT:
954 case ICmpInst::ICMP_SGT:
955 if (HiOverflow == +1) // High bound greater than input range.
956 return ReplaceInstUsesWith(ICI, Builder->getFalse());
957 if (HiOverflow == -1) // High bound less than input range.
958 return ReplaceInstUsesWith(ICI, Builder->getTrue());
959 if (Pred == ICmpInst::ICMP_UGT)
960 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
961 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
965 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
966 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
967 ConstantInt *ShAmt) {
968 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
970 // Check that the shift amount is in range. If not, don't perform
971 // undefined shifts. When the shift is visited it will be
973 uint32_t TypeBits = CmpRHSV.getBitWidth();
974 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
975 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
978 if (!ICI.isEquality()) {
979 // If we have an unsigned comparison and an ashr, we can't simplify this.
980 // Similarly for signed comparisons with lshr.
981 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
984 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
985 // by a power of 2. Since we already have logic to simplify these,
986 // transform to div and then simplify the resultant comparison.
987 if (Shr->getOpcode() == Instruction::AShr &&
988 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
991 // Revisit the shift (to delete it).
995 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
998 Shr->getOpcode() == Instruction::AShr ?
999 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
1000 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
1002 ICI.setOperand(0, Tmp);
1004 // If the builder folded the binop, just return it.
1005 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
1009 // Otherwise, fold this div/compare.
1010 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1011 TheDiv->getOpcode() == Instruction::UDiv);
1013 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1014 assert(Res && "This div/cst should have folded!");
1019 // If we are comparing against bits always shifted out, the
1020 // comparison cannot succeed.
1021 APInt Comp = CmpRHSV << ShAmtVal;
1022 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1023 if (Shr->getOpcode() == Instruction::LShr)
1024 Comp = Comp.lshr(ShAmtVal);
1026 Comp = Comp.ashr(ShAmtVal);
1028 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1029 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1030 Constant *Cst = Builder->getInt1(IsICMP_NE);
1031 return ReplaceInstUsesWith(ICI, Cst);
1034 // Otherwise, check to see if the bits shifted out are known to be zero.
1035 // If so, we can compare against the unshifted value:
1036 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1037 if (Shr->hasOneUse() && Shr->isExact())
1038 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1040 if (Shr->hasOneUse()) {
1041 // Otherwise strength reduce the shift into an and.
1042 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1043 Constant *Mask = Builder->getInt(Val);
1045 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1046 Mask, Shr->getName()+".mask");
1047 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1053 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1055 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1058 const APInt &RHSV = RHS->getValue();
1060 switch (LHSI->getOpcode()) {
1061 case Instruction::Trunc:
1062 if (ICI.isEquality() && LHSI->hasOneUse()) {
1063 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1064 // of the high bits truncated out of x are known.
1065 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1066 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1067 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1068 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1070 // If all the high bits are known, we can do this xform.
1071 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1072 // Pull in the high bits from known-ones set.
1073 APInt NewRHS = RHS->getValue().zext(SrcBits);
1074 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1075 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1076 Builder->getInt(NewRHS));
1081 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1082 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1083 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1085 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1086 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1087 Value *CompareVal = LHSI->getOperand(0);
1089 // If the sign bit of the XorCst is not set, there is no change to
1090 // the operation, just stop using the Xor.
1091 if (!XorCst->isNegative()) {
1092 ICI.setOperand(0, CompareVal);
1097 // Was the old condition true if the operand is positive?
1098 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1100 // If so, the new one isn't.
1101 isTrueIfPositive ^= true;
1103 if (isTrueIfPositive)
1104 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1107 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1111 if (LHSI->hasOneUse()) {
1112 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1113 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1114 const APInt &SignBit = XorCst->getValue();
1115 ICmpInst::Predicate Pred = ICI.isSigned()
1116 ? ICI.getUnsignedPredicate()
1117 : ICI.getSignedPredicate();
1118 return new ICmpInst(Pred, LHSI->getOperand(0),
1119 Builder->getInt(RHSV ^ SignBit));
1122 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1123 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1124 const APInt &NotSignBit = XorCst->getValue();
1125 ICmpInst::Predicate Pred = ICI.isSigned()
1126 ? ICI.getUnsignedPredicate()
1127 : ICI.getSignedPredicate();
1128 Pred = ICI.getSwappedPredicate(Pred);
1129 return new ICmpInst(Pred, LHSI->getOperand(0),
1130 Builder->getInt(RHSV ^ NotSignBit));
1134 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1135 // iff -C is a power of 2
1136 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1137 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1138 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1140 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1141 // iff -C is a power of 2
1142 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1143 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1144 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1147 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1148 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1149 LHSI->getOperand(0)->hasOneUse()) {
1150 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1152 // If the LHS is an AND of a truncating cast, we can widen the
1153 // and/compare to be the input width without changing the value
1154 // produced, eliminating a cast.
1155 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1156 // We can do this transformation if either the AND constant does not
1157 // have its sign bit set or if it is an equality comparison.
1158 // Extending a relational comparison when we're checking the sign
1159 // bit would not work.
1160 if (ICI.isEquality() ||
1161 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1163 Builder->CreateAnd(Cast->getOperand(0),
1164 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1165 NewAnd->takeName(LHSI);
1166 return new ICmpInst(ICI.getPredicate(), NewAnd,
1167 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1171 // If the LHS is an AND of a zext, and we have an equality compare, we can
1172 // shrink the and/compare to the smaller type, eliminating the cast.
1173 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1174 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1175 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1176 // should fold the icmp to true/false in that case.
1177 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1179 Builder->CreateAnd(Cast->getOperand(0),
1180 ConstantExpr::getTrunc(AndCst, Ty));
1181 NewAnd->takeName(LHSI);
1182 return new ICmpInst(ICI.getPredicate(), NewAnd,
1183 ConstantExpr::getTrunc(RHS, Ty));
1187 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1188 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1189 // happens a LOT in code produced by the C front-end, for bitfield
1191 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1192 if (Shift && !Shift->isShift())
1196 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1198 // This seemingly simple opportunity to fold away a shift turns out to
1199 // be rather complicated. See PR17827
1200 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1202 bool CanFold = false;
1203 unsigned ShiftOpcode = Shift->getOpcode();
1204 if (ShiftOpcode == Instruction::AShr) {
1205 // There may be some constraints that make this possible,
1206 // but nothing simple has been discovered yet.
1208 } else if (ShiftOpcode == Instruction::Shl) {
1209 // For a left shift, we can fold if the comparison is not signed.
1210 // We can also fold a signed comparison if the mask value and
1211 // comparison value are not negative. These constraints may not be
1212 // obvious, but we can prove that they are correct using an SMT
1213 // solver such as "Z3" :
1214 // http://rise4fun.com/Z3/DyMp
1215 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1217 } else if (ShiftOpcode == Instruction::LShr) {
1218 // For a logical right shift, we can fold if the comparison is not
1219 // signed. We can also fold a signed comparison if the shifted mask
1220 // value and the shifted comparison value are not negative.
1221 // These constraints may not be obvious, but we can prove that they
1222 // are correct using an SMT solver such as "Z3" :
1223 // http://rise4fun.com/Z3/Tslfh
1224 if (!ICI.isSigned())
1227 ConstantInt *ShiftedAndCst =
1228 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1229 ConstantInt *ShiftedRHSCst =
1230 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1232 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1239 if (ShiftOpcode == Instruction::Shl)
1240 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1242 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1244 // Check to see if we are shifting out any of the bits being
1246 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1247 // If we shifted bits out, the fold is not going to work out.
1248 // As a special case, check to see if this means that the
1249 // result is always true or false now.
1250 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1251 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1252 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1253 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1255 ICI.setOperand(1, NewCst);
1256 Constant *NewAndCst;
1257 if (ShiftOpcode == Instruction::Shl)
1258 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1260 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1261 LHSI->setOperand(1, NewAndCst);
1262 LHSI->setOperand(0, Shift->getOperand(0));
1263 Worklist.Add(Shift); // Shift is dead.
1269 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1270 // preferable because it allows the C<<Y expression to be hoisted out
1271 // of a loop if Y is invariant and X is not.
1272 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1273 ICI.isEquality() && !Shift->isArithmeticShift() &&
1274 !isa<Constant>(Shift->getOperand(0))) {
1277 if (Shift->getOpcode() == Instruction::LShr) {
1278 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1280 // Insert a logical shift.
1281 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1284 // Compute X & (C << Y).
1286 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1288 ICI.setOperand(0, NewAnd);
1292 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1293 // bit set in (X & AndCst) will produce a result greater than RHSV.
1294 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1295 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1296 if ((NTZ < AndCst->getBitWidth()) &&
1297 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1298 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1299 Constant::getNullValue(RHS->getType()));
1303 // Try to optimize things like "A[i]&42 == 0" to index computations.
1304 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1305 if (GetElementPtrInst *GEP =
1306 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1307 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1308 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1309 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1310 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1311 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1316 // X & -C == -C -> X > u ~C
1317 // X & -C != -C -> X <= u ~C
1318 // iff C is a power of 2
1319 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1320 return new ICmpInst(
1321 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1322 : ICmpInst::ICMP_ULE,
1323 LHSI->getOperand(0), SubOne(RHS));
1326 case Instruction::Or: {
1327 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1330 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1331 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1332 // -> and (icmp eq P, null), (icmp eq Q, null).
1333 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1334 Constant::getNullValue(P->getType()));
1335 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1336 Constant::getNullValue(Q->getType()));
1338 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1339 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1341 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1347 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1348 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1351 // If this is a signed comparison to 0 and the mul is sign preserving,
1352 // use the mul LHS operand instead.
1353 ICmpInst::Predicate pred = ICI.getPredicate();
1354 if (isSignTest(pred, RHS) && !Val->isZero() &&
1355 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1356 return new ICmpInst(Val->isNegative() ?
1357 ICmpInst::getSwappedPredicate(pred) : pred,
1358 LHSI->getOperand(0),
1359 Constant::getNullValue(RHS->getType()));
1364 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1365 uint32_t TypeBits = RHSV.getBitWidth();
1366 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1369 // (1 << X) pred P2 -> X pred Log2(P2)
1370 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1371 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1372 ICmpInst::Predicate Pred = ICI.getPredicate();
1373 if (ICI.isUnsigned()) {
1374 if (!RHSVIsPowerOf2) {
1375 // (1 << X) < 30 -> X <= 4
1376 // (1 << X) <= 30 -> X <= 4
1377 // (1 << X) >= 30 -> X > 4
1378 // (1 << X) > 30 -> X > 4
1379 if (Pred == ICmpInst::ICMP_ULT)
1380 Pred = ICmpInst::ICMP_ULE;
1381 else if (Pred == ICmpInst::ICMP_UGE)
1382 Pred = ICmpInst::ICMP_UGT;
1384 unsigned RHSLog2 = RHSV.logBase2();
1386 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1387 // (1 << X) > 2147483648 -> X > 31 -> false
1388 // (1 << X) <= 2147483648 -> X <= 31 -> true
1389 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1390 if (RHSLog2 == TypeBits-1) {
1391 if (Pred == ICmpInst::ICMP_UGE)
1392 Pred = ICmpInst::ICMP_EQ;
1393 else if (Pred == ICmpInst::ICMP_UGT)
1394 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1395 else if (Pred == ICmpInst::ICMP_ULE)
1396 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1397 else if (Pred == ICmpInst::ICMP_ULT)
1398 Pred = ICmpInst::ICMP_NE;
1401 return new ICmpInst(Pred, X,
1402 ConstantInt::get(RHS->getType(), RHSLog2));
1403 } else if (ICI.isSigned()) {
1404 if (RHSV.isAllOnesValue()) {
1405 // (1 << X) <= -1 -> X == 31
1406 if (Pred == ICmpInst::ICMP_SLE)
1407 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1408 ConstantInt::get(RHS->getType(), TypeBits-1));
1410 // (1 << X) > -1 -> X != 31
1411 if (Pred == ICmpInst::ICMP_SGT)
1412 return new ICmpInst(ICmpInst::ICMP_NE, X,
1413 ConstantInt::get(RHS->getType(), TypeBits-1));
1415 // (1 << X) < 0 -> X == 31
1416 // (1 << X) <= 0 -> X == 31
1417 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1418 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1419 ConstantInt::get(RHS->getType(), TypeBits-1));
1421 // (1 << X) >= 0 -> X != 31
1422 // (1 << X) > 0 -> X != 31
1423 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1424 return new ICmpInst(ICmpInst::ICMP_NE, X,
1425 ConstantInt::get(RHS->getType(), TypeBits-1));
1427 } else if (ICI.isEquality()) {
1429 return new ICmpInst(
1430 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1432 return ReplaceInstUsesWith(
1433 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse()
1434 : Builder->getTrue());
1440 // Check that the shift amount is in range. If not, don't perform
1441 // undefined shifts. When the shift is visited it will be
1443 if (ShAmt->uge(TypeBits))
1446 if (ICI.isEquality()) {
1447 // If we are comparing against bits always shifted out, the
1448 // comparison cannot succeed.
1450 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1452 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1453 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1454 Constant *Cst = Builder->getInt1(IsICMP_NE);
1455 return ReplaceInstUsesWith(ICI, Cst);
1458 // If the shift is NUW, then it is just shifting out zeros, no need for an
1460 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1461 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1462 ConstantExpr::getLShr(RHS, ShAmt));
1464 // If the shift is NSW and we compare to 0, then it is just shifting out
1465 // sign bits, no need for an AND either.
1466 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1467 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1468 ConstantExpr::getLShr(RHS, ShAmt));
1470 if (LHSI->hasOneUse()) {
1471 // Otherwise strength reduce the shift into an and.
1472 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1473 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1474 TypeBits - ShAmtVal));
1477 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1478 return new ICmpInst(ICI.getPredicate(), And,
1479 ConstantExpr::getLShr(RHS, ShAmt));
1483 // If this is a signed comparison to 0 and the shift is sign preserving,
1484 // use the shift LHS operand instead.
1485 ICmpInst::Predicate pred = ICI.getPredicate();
1486 if (isSignTest(pred, RHS) &&
1487 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1488 return new ICmpInst(pred,
1489 LHSI->getOperand(0),
1490 Constant::getNullValue(RHS->getType()));
1492 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1493 bool TrueIfSigned = false;
1494 if (LHSI->hasOneUse() &&
1495 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1496 // (X << 31) <s 0 --> (X&1) != 0
1497 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1498 APInt::getOneBitSet(TypeBits,
1499 TypeBits-ShAmt->getZExtValue()-1));
1501 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1502 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1503 And, Constant::getNullValue(And->getType()));
1506 // Transform (icmp pred iM (shl iM %v, N), CI)
1507 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1508 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1509 // This enables to get rid of the shift in favor of a trunc which can be
1510 // free on the target. It has the additional benefit of comparing to a
1511 // smaller constant, which will be target friendly.
1512 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1513 if (LHSI->hasOneUse() &&
1514 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1515 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1516 Constant *NCI = ConstantExpr::getTrunc(
1517 ConstantExpr::getAShr(RHS,
1518 ConstantInt::get(RHS->getType(), Amt)),
1520 return new ICmpInst(ICI.getPredicate(),
1521 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1528 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1529 case Instruction::AShr: {
1530 // Handle equality comparisons of shift-by-constant.
1531 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1532 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1533 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1537 // Handle exact shr's.
1538 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1539 if (RHSV.isMinValue())
1540 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1545 case Instruction::SDiv:
1546 case Instruction::UDiv:
1547 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1548 // Fold this div into the comparison, producing a range check.
1549 // Determine, based on the divide type, what the range is being
1550 // checked. If there is an overflow on the low or high side, remember
1551 // it, otherwise compute the range [low, hi) bounding the new value.
1552 // See: InsertRangeTest above for the kinds of replacements possible.
1553 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1554 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1559 case Instruction::Sub: {
1560 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1562 const APInt &LHSV = LHSC->getValue();
1564 // C1-X <u C2 -> (X|(C2-1)) == C1
1565 // iff C1 & (C2-1) == C2-1
1566 // C2 is a power of 2
1567 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1568 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1569 return new ICmpInst(ICmpInst::ICMP_EQ,
1570 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1573 // C1-X >u C2 -> (X|C2) != C1
1574 // iff C1 & C2 == C2
1575 // C2+1 is a power of 2
1576 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1577 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1578 return new ICmpInst(ICmpInst::ICMP_NE,
1579 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1583 case Instruction::Add:
1584 // Fold: icmp pred (add X, C1), C2
1585 if (!ICI.isEquality()) {
1586 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1588 const APInt &LHSV = LHSC->getValue();
1590 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1593 if (ICI.isSigned()) {
1594 if (CR.getLower().isSignBit()) {
1595 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1596 Builder->getInt(CR.getUpper()));
1597 } else if (CR.getUpper().isSignBit()) {
1598 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1599 Builder->getInt(CR.getLower()));
1602 if (CR.getLower().isMinValue()) {
1603 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1604 Builder->getInt(CR.getUpper()));
1605 } else if (CR.getUpper().isMinValue()) {
1606 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1607 Builder->getInt(CR.getLower()));
1611 // X-C1 <u C2 -> (X & -C2) == C1
1612 // iff C1 & (C2-1) == 0
1613 // C2 is a power of 2
1614 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1615 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1616 return new ICmpInst(ICmpInst::ICMP_EQ,
1617 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1618 ConstantExpr::getNeg(LHSC));
1620 // X-C1 >u C2 -> (X & ~C2) != C1
1622 // C2+1 is a power of 2
1623 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1624 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1625 return new ICmpInst(ICmpInst::ICMP_NE,
1626 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1627 ConstantExpr::getNeg(LHSC));
1632 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1633 if (ICI.isEquality()) {
1634 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1636 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1637 // the second operand is a constant, simplify a bit.
1638 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1639 switch (BO->getOpcode()) {
1640 case Instruction::SRem:
1641 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1642 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1643 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1644 if (V.sgt(1) && V.isPowerOf2()) {
1646 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1648 return new ICmpInst(ICI.getPredicate(), NewRem,
1649 Constant::getNullValue(BO->getType()));
1653 case Instruction::Add:
1654 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1655 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1656 if (BO->hasOneUse())
1657 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1658 ConstantExpr::getSub(RHS, BOp1C));
1659 } else if (RHSV == 0) {
1660 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1661 // efficiently invertible, or if the add has just this one use.
1662 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1664 if (Value *NegVal = dyn_castNegVal(BOp1))
1665 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1666 if (Value *NegVal = dyn_castNegVal(BOp0))
1667 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1668 if (BO->hasOneUse()) {
1669 Value *Neg = Builder->CreateNeg(BOp1);
1671 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1675 case Instruction::Xor:
1676 // For the xor case, we can xor two constants together, eliminating
1677 // the explicit xor.
1678 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1679 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1680 ConstantExpr::getXor(RHS, BOC));
1681 } else if (RHSV == 0) {
1682 // Replace ((xor A, B) != 0) with (A != B)
1683 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1687 case Instruction::Sub:
1688 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1689 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1690 if (BO->hasOneUse())
1691 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1692 ConstantExpr::getSub(BOp0C, RHS));
1693 } else if (RHSV == 0) {
1694 // Replace ((sub A, B) != 0) with (A != B)
1695 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1699 case Instruction::Or:
1700 // If bits are being or'd in that are not present in the constant we
1701 // are comparing against, then the comparison could never succeed!
1702 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1703 Constant *NotCI = ConstantExpr::getNot(RHS);
1704 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1705 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1709 case Instruction::And:
1710 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1711 // If bits are being compared against that are and'd out, then the
1712 // comparison can never succeed!
1713 if ((RHSV & ~BOC->getValue()) != 0)
1714 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1716 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1717 if (RHS == BOC && RHSV.isPowerOf2())
1718 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1719 ICmpInst::ICMP_NE, LHSI,
1720 Constant::getNullValue(RHS->getType()));
1722 // Don't perform the following transforms if the AND has multiple uses
1723 if (!BO->hasOneUse())
1726 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1727 if (BOC->getValue().isSignBit()) {
1728 Value *X = BO->getOperand(0);
1729 Constant *Zero = Constant::getNullValue(X->getType());
1730 ICmpInst::Predicate pred = isICMP_NE ?
1731 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1732 return new ICmpInst(pred, X, Zero);
1735 // ((X & ~7) == 0) --> X < 8
1736 if (RHSV == 0 && isHighOnes(BOC)) {
1737 Value *X = BO->getOperand(0);
1738 Constant *NegX = ConstantExpr::getNeg(BOC);
1739 ICmpInst::Predicate pred = isICMP_NE ?
1740 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1741 return new ICmpInst(pred, X, NegX);
1745 case Instruction::Mul:
1746 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1747 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1748 // The trivial case (mul X, 0) is handled by InstSimplify
1749 // General case : (mul X, C) != 0 iff X != 0
1750 // (mul X, C) == 0 iff X == 0
1752 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1753 Constant::getNullValue(RHS->getType()));
1759 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1760 // Handle icmp {eq|ne} <intrinsic>, intcst.
1761 switch (II->getIntrinsicID()) {
1762 case Intrinsic::bswap:
1764 ICI.setOperand(0, II->getArgOperand(0));
1765 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1767 case Intrinsic::ctlz:
1768 case Intrinsic::cttz:
1769 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1770 if (RHSV == RHS->getType()->getBitWidth()) {
1772 ICI.setOperand(0, II->getArgOperand(0));
1773 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1777 case Intrinsic::ctpop:
1778 // popcount(A) == 0 -> A == 0 and likewise for !=
1779 if (RHS->isZero()) {
1781 ICI.setOperand(0, II->getArgOperand(0));
1782 ICI.setOperand(1, RHS);
1794 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1795 /// We only handle extending casts so far.
1797 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1798 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1799 Value *LHSCIOp = LHSCI->getOperand(0);
1800 Type *SrcTy = LHSCIOp->getType();
1801 Type *DestTy = LHSCI->getType();
1804 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1805 // integer type is the same size as the pointer type.
1806 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1807 TD->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1809 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1810 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1811 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1812 RHSOp = RHSC->getOperand(0);
1813 // If the pointer types don't match, insert a bitcast.
1814 if (LHSCIOp->getType() != RHSOp->getType())
1815 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1819 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1822 // The code below only handles extension cast instructions, so far.
1824 if (LHSCI->getOpcode() != Instruction::ZExt &&
1825 LHSCI->getOpcode() != Instruction::SExt)
1828 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1829 bool isSignedCmp = ICI.isSigned();
1831 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1832 // Not an extension from the same type?
1833 RHSCIOp = CI->getOperand(0);
1834 if (RHSCIOp->getType() != LHSCIOp->getType())
1837 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1838 // and the other is a zext), then we can't handle this.
1839 if (CI->getOpcode() != LHSCI->getOpcode())
1842 // Deal with equality cases early.
1843 if (ICI.isEquality())
1844 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1846 // A signed comparison of sign extended values simplifies into a
1847 // signed comparison.
1848 if (isSignedCmp && isSignedExt)
1849 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1851 // The other three cases all fold into an unsigned comparison.
1852 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1855 // If we aren't dealing with a constant on the RHS, exit early
1856 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1860 // Compute the constant that would happen if we truncated to SrcTy then
1861 // reextended to DestTy.
1862 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1863 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1866 // If the re-extended constant didn't change...
1868 // Deal with equality cases early.
1869 if (ICI.isEquality())
1870 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1872 // A signed comparison of sign extended values simplifies into a
1873 // signed comparison.
1874 if (isSignedExt && isSignedCmp)
1875 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1877 // The other three cases all fold into an unsigned comparison.
1878 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1881 // The re-extended constant changed so the constant cannot be represented
1882 // in the shorter type. Consequently, we cannot emit a simple comparison.
1883 // All the cases that fold to true or false will have already been handled
1884 // by SimplifyICmpInst, so only deal with the tricky case.
1886 if (isSignedCmp || !isSignedExt)
1889 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1890 // should have been folded away previously and not enter in here.
1892 // We're performing an unsigned comp with a sign extended value.
1893 // This is true if the input is >= 0. [aka >s -1]
1894 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1895 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1897 // Finally, return the value computed.
1898 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1899 return ReplaceInstUsesWith(ICI, Result);
1901 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1902 return BinaryOperator::CreateNot(Result);
1905 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1906 /// I = icmp ugt (add (add A, B), CI2), CI1
1907 /// If this is of the form:
1909 /// if (sum+128 >u 255)
1910 /// Then replace it with llvm.sadd.with.overflow.i8.
1912 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1913 ConstantInt *CI2, ConstantInt *CI1,
1915 // The transformation we're trying to do here is to transform this into an
1916 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1917 // with a narrower add, and discard the add-with-constant that is part of the
1918 // range check (if we can't eliminate it, this isn't profitable).
1920 // In order to eliminate the add-with-constant, the compare can be its only
1922 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1923 if (!AddWithCst->hasOneUse()) return 0;
1925 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1926 if (!CI2->getValue().isPowerOf2()) return 0;
1927 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1928 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1930 // The width of the new add formed is 1 more than the bias.
1933 // Check to see that CI1 is an all-ones value with NewWidth bits.
1934 if (CI1->getBitWidth() == NewWidth ||
1935 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1938 // This is only really a signed overflow check if the inputs have been
1939 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1940 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1941 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1942 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1943 IC.ComputeNumSignBits(B) < NeededSignBits)
1946 // In order to replace the original add with a narrower
1947 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1948 // and truncates that discard the high bits of the add. Verify that this is
1950 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1951 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1953 if (*UI == AddWithCst) continue;
1955 // Only accept truncates for now. We would really like a nice recursive
1956 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1957 // chain to see which bits of a value are actually demanded. If the
1958 // original add had another add which was then immediately truncated, we
1959 // could still do the transformation.
1960 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1962 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1965 // If the pattern matches, truncate the inputs to the narrower type and
1966 // use the sadd_with_overflow intrinsic to efficiently compute both the
1967 // result and the overflow bit.
1968 Module *M = I.getParent()->getParent()->getParent();
1970 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1971 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1974 InstCombiner::BuilderTy *Builder = IC.Builder;
1976 // Put the new code above the original add, in case there are any uses of the
1977 // add between the add and the compare.
1978 Builder->SetInsertPoint(OrigAdd);
1980 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1981 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1982 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1983 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1984 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1986 // The inner add was the result of the narrow add, zero extended to the
1987 // wider type. Replace it with the result computed by the intrinsic.
1988 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1990 // The original icmp gets replaced with the overflow value.
1991 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1994 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1996 // Don't bother doing this transformation for pointers, don't do it for
1998 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
2000 // If the add is a constant expr, then we don't bother transforming it.
2001 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
2002 if (OrigAdd == 0) return 0;
2004 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
2006 // Put the new code above the original add, in case there are any uses of the
2007 // add between the add and the compare.
2008 InstCombiner::BuilderTy *Builder = IC.Builder;
2009 Builder->SetInsertPoint(OrigAdd);
2011 Module *M = I.getParent()->getParent()->getParent();
2012 Type *Ty = LHS->getType();
2013 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
2014 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
2015 Value *Add = Builder->CreateExtractValue(Call, 0);
2017 IC.ReplaceInstUsesWith(*OrigAdd, Add);
2019 // The original icmp gets replaced with the overflow value.
2020 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2023 // DemandedBitsLHSMask - When performing a comparison against a constant,
2024 // it is possible that not all the bits in the LHS are demanded. This helper
2025 // method computes the mask that IS demanded.
2026 static APInt DemandedBitsLHSMask(ICmpInst &I,
2027 unsigned BitWidth, bool isSignCheck) {
2029 return APInt::getSignBit(BitWidth);
2031 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2032 if (!CI) return APInt::getAllOnesValue(BitWidth);
2033 const APInt &RHS = CI->getValue();
2035 switch (I.getPredicate()) {
2036 // For a UGT comparison, we don't care about any bits that
2037 // correspond to the trailing ones of the comparand. The value of these
2038 // bits doesn't impact the outcome of the comparison, because any value
2039 // greater than the RHS must differ in a bit higher than these due to carry.
2040 case ICmpInst::ICMP_UGT: {
2041 unsigned trailingOnes = RHS.countTrailingOnes();
2042 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2046 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2047 // Any value less than the RHS must differ in a higher bit because of carries.
2048 case ICmpInst::ICMP_ULT: {
2049 unsigned trailingZeros = RHS.countTrailingZeros();
2050 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2055 return APInt::getAllOnesValue(BitWidth);
2060 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2061 /// should be swapped.
2062 /// The descision is based on how many times these two operands are reused
2063 /// as subtract operands and their positions in those instructions.
2064 /// The rational is that several architectures use the same instruction for
2065 /// both subtract and cmp, thus it is better if the order of those operands
2067 /// \return true if Op0 and Op1 should be swapped.
2068 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2069 const Value * Op1) {
2070 // Filter out pointer value as those cannot appears directly in subtract.
2071 // FIXME: we may want to go through inttoptrs or bitcasts.
2072 if (Op0->getType()->isPointerTy())
2074 // Count every uses of both Op0 and Op1 in a subtract.
2075 // Each time Op0 is the first operand, count -1: swapping is bad, the
2076 // subtract has already the same layout as the compare.
2077 // Each time Op0 is the second operand, count +1: swapping is good, the
2078 // subtract has a diffrent layout as the compare.
2079 // At the end, if the benefit is greater than 0, Op0 should come second to
2080 // expose more CSE opportunities.
2081 int GlobalSwapBenefits = 0;
2082 for (Value::const_use_iterator UI = Op0->use_begin(), UIEnd = Op0->use_end(); UI != UIEnd; ++UI) {
2083 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(*UI);
2084 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2086 // If Op0 is the first argument, this is not beneficial to swap the
2088 int LocalSwapBenefits = -1;
2089 unsigned Op1Idx = 1;
2090 if (BinOp->getOperand(Op1Idx) == Op0) {
2092 LocalSwapBenefits = 1;
2094 if (BinOp->getOperand(Op1Idx) != Op1)
2096 GlobalSwapBenefits += LocalSwapBenefits;
2098 return GlobalSwapBenefits > 0;
2101 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2102 bool Changed = false;
2103 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2104 unsigned Op0Cplxity = getComplexity(Op0);
2105 unsigned Op1Cplxity = getComplexity(Op1);
2107 /// Orders the operands of the compare so that they are listed from most
2108 /// complex to least complex. This puts constants before unary operators,
2109 /// before binary operators.
2110 if (Op0Cplxity < Op1Cplxity ||
2111 (Op0Cplxity == Op1Cplxity &&
2112 swapMayExposeCSEOpportunities(Op0, Op1))) {
2114 std::swap(Op0, Op1);
2118 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
2119 return ReplaceInstUsesWith(I, V);
2121 // comparing -val or val with non-zero is the same as just comparing val
2122 // ie, abs(val) != 0 -> val != 0
2123 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2125 Value *Cond, *SelectTrue, *SelectFalse;
2126 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2127 m_Value(SelectFalse)))) {
2128 if (Value *V = dyn_castNegVal(SelectTrue)) {
2129 if (V == SelectFalse)
2130 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2132 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2133 if (V == SelectTrue)
2134 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2139 Type *Ty = Op0->getType();
2141 // icmp's with boolean values can always be turned into bitwise operations
2142 if (Ty->isIntegerTy(1)) {
2143 switch (I.getPredicate()) {
2144 default: llvm_unreachable("Invalid icmp instruction!");
2145 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2146 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2147 return BinaryOperator::CreateNot(Xor);
2149 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2150 return BinaryOperator::CreateXor(Op0, Op1);
2152 case ICmpInst::ICMP_UGT:
2153 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2155 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2156 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2157 return BinaryOperator::CreateAnd(Not, Op1);
2159 case ICmpInst::ICMP_SGT:
2160 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2162 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2163 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2164 return BinaryOperator::CreateAnd(Not, Op0);
2166 case ICmpInst::ICMP_UGE:
2167 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2169 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2170 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2171 return BinaryOperator::CreateOr(Not, Op1);
2173 case ICmpInst::ICMP_SGE:
2174 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2176 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2177 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2178 return BinaryOperator::CreateOr(Not, Op0);
2183 unsigned BitWidth = 0;
2184 if (Ty->isIntOrIntVectorTy())
2185 BitWidth = Ty->getScalarSizeInBits();
2186 else if (TD) // Pointers require TD info to get their size.
2187 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
2189 bool isSignBit = false;
2191 // See if we are doing a comparison with a constant.
2192 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2193 Value *A = 0, *B = 0;
2195 // Match the following pattern, which is a common idiom when writing
2196 // overflow-safe integer arithmetic function. The source performs an
2197 // addition in wider type, and explicitly checks for overflow using
2198 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2199 // sadd_with_overflow intrinsic.
2201 // TODO: This could probably be generalized to handle other overflow-safe
2202 // operations if we worked out the formulas to compute the appropriate
2206 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2208 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2209 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2210 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2211 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2215 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2216 if (I.isEquality() && CI->isZero() &&
2217 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2218 // (icmp cond A B) if cond is equality
2219 return new ICmpInst(I.getPredicate(), A, B);
2222 // If we have an icmp le or icmp ge instruction, turn it into the
2223 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2224 // them being folded in the code below. The SimplifyICmpInst code has
2225 // already handled the edge cases for us, so we just assert on them.
2226 switch (I.getPredicate()) {
2228 case ICmpInst::ICMP_ULE:
2229 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2230 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2231 Builder->getInt(CI->getValue()+1));
2232 case ICmpInst::ICMP_SLE:
2233 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2234 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2235 Builder->getInt(CI->getValue()+1));
2236 case ICmpInst::ICMP_UGE:
2237 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2238 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2239 Builder->getInt(CI->getValue()-1));
2240 case ICmpInst::ICMP_SGE:
2241 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2242 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2243 Builder->getInt(CI->getValue()-1));
2246 // If this comparison is a normal comparison, it demands all
2247 // bits, if it is a sign bit comparison, it only demands the sign bit.
2249 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2252 // See if we can fold the comparison based on range information we can get
2253 // by checking whether bits are known to be zero or one in the input.
2254 if (BitWidth != 0) {
2255 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2256 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2258 if (SimplifyDemandedBits(I.getOperandUse(0),
2259 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2260 Op0KnownZero, Op0KnownOne, 0))
2262 if (SimplifyDemandedBits(I.getOperandUse(1),
2263 APInt::getAllOnesValue(BitWidth),
2264 Op1KnownZero, Op1KnownOne, 0))
2267 // Given the known and unknown bits, compute a range that the LHS could be
2268 // in. Compute the Min, Max and RHS values based on the known bits. For the
2269 // EQ and NE we use unsigned values.
2270 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2271 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2273 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2275 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2278 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2280 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2284 // If Min and Max are known to be the same, then SimplifyDemandedBits
2285 // figured out that the LHS is a constant. Just constant fold this now so
2286 // that code below can assume that Min != Max.
2287 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2288 return new ICmpInst(I.getPredicate(),
2289 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2290 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2291 return new ICmpInst(I.getPredicate(), Op0,
2292 ConstantInt::get(Op1->getType(), Op1Min));
2294 // Based on the range information we know about the LHS, see if we can
2295 // simplify this comparison. For example, (x&4) < 8 is always true.
2296 switch (I.getPredicate()) {
2297 default: llvm_unreachable("Unknown icmp opcode!");
2298 case ICmpInst::ICMP_EQ: {
2299 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2300 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2302 // If all bits are known zero except for one, then we know at most one
2303 // bit is set. If the comparison is against zero, then this is a check
2304 // to see if *that* bit is set.
2305 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2306 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2307 // If the LHS is an AND with the same constant, look through it.
2309 ConstantInt *LHSC = 0;
2310 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2311 LHSC->getValue() != Op0KnownZeroInverted)
2314 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2315 // then turn "((1 << x)&8) == 0" into "x != 3".
2317 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2318 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2319 return new ICmpInst(ICmpInst::ICMP_NE, X,
2320 ConstantInt::get(X->getType(), CmpVal));
2323 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2324 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2326 if (Op0KnownZeroInverted == 1 &&
2327 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2328 return new ICmpInst(ICmpInst::ICMP_NE, X,
2329 ConstantInt::get(X->getType(),
2330 CI->countTrailingZeros()));
2335 case ICmpInst::ICMP_NE: {
2336 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2337 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2339 // If all bits are known zero except for one, then we know at most one
2340 // bit is set. If the comparison is against zero, then this is a check
2341 // to see if *that* bit is set.
2342 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2343 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2344 // If the LHS is an AND with the same constant, look through it.
2346 ConstantInt *LHSC = 0;
2347 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2348 LHSC->getValue() != Op0KnownZeroInverted)
2351 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2352 // then turn "((1 << x)&8) != 0" into "x == 3".
2354 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2355 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2356 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2357 ConstantInt::get(X->getType(), CmpVal));
2360 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2361 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2363 if (Op0KnownZeroInverted == 1 &&
2364 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2365 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2366 ConstantInt::get(X->getType(),
2367 CI->countTrailingZeros()));
2372 case ICmpInst::ICMP_ULT:
2373 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2374 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2375 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2376 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2377 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2378 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2379 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2380 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2381 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2382 Builder->getInt(CI->getValue()-1));
2384 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2385 if (CI->isMinValue(true))
2386 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2387 Constant::getAllOnesValue(Op0->getType()));
2390 case ICmpInst::ICMP_UGT:
2391 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2392 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2393 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2394 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2396 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2397 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2398 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2399 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2400 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2401 Builder->getInt(CI->getValue()+1));
2403 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2404 if (CI->isMaxValue(true))
2405 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2406 Constant::getNullValue(Op0->getType()));
2409 case ICmpInst::ICMP_SLT:
2410 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2411 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2412 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2413 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2414 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2415 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2416 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2417 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2418 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2419 Builder->getInt(CI->getValue()-1));
2422 case ICmpInst::ICMP_SGT:
2423 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2424 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2425 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2426 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2428 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2429 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2430 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2431 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2432 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2433 Builder->getInt(CI->getValue()+1));
2436 case ICmpInst::ICMP_SGE:
2437 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2438 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2439 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2440 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2441 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2443 case ICmpInst::ICMP_SLE:
2444 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2445 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2446 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2447 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2448 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2450 case ICmpInst::ICMP_UGE:
2451 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2452 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2453 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2454 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2455 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2457 case ICmpInst::ICMP_ULE:
2458 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2459 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2460 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2461 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2462 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2466 // Turn a signed comparison into an unsigned one if both operands
2467 // are known to have the same sign.
2469 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2470 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2471 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2474 // Test if the ICmpInst instruction is used exclusively by a select as
2475 // part of a minimum or maximum operation. If so, refrain from doing
2476 // any other folding. This helps out other analyses which understand
2477 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2478 // and CodeGen. And in this case, at least one of the comparison
2479 // operands has at least one user besides the compare (the select),
2480 // which would often largely negate the benefit of folding anyway.
2482 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2483 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2484 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2487 // See if we are doing a comparison between a constant and an instruction that
2488 // can be folded into the comparison.
2489 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2490 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2491 // instruction, see if that instruction also has constants so that the
2492 // instruction can be folded into the icmp
2493 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2494 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2498 // Handle icmp with constant (but not simple integer constant) RHS
2499 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2500 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2501 switch (LHSI->getOpcode()) {
2502 case Instruction::GetElementPtr:
2503 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2504 if (RHSC->isNullValue() &&
2505 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2506 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2507 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2509 case Instruction::PHI:
2510 // Only fold icmp into the PHI if the phi and icmp are in the same
2511 // block. If in the same block, we're encouraging jump threading. If
2512 // not, we are just pessimizing the code by making an i1 phi.
2513 if (LHSI->getParent() == I.getParent())
2514 if (Instruction *NV = FoldOpIntoPhi(I))
2517 case Instruction::Select: {
2518 // If either operand of the select is a constant, we can fold the
2519 // comparison into the select arms, which will cause one to be
2520 // constant folded and the select turned into a bitwise or.
2521 Value *Op1 = 0, *Op2 = 0;
2522 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2523 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2524 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2525 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2527 // We only want to perform this transformation if it will not lead to
2528 // additional code. This is true if either both sides of the select
2529 // fold to a constant (in which case the icmp is replaced with a select
2530 // which will usually simplify) or this is the only user of the
2531 // select (in which case we are trading a select+icmp for a simpler
2533 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2535 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2538 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2540 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2544 case Instruction::IntToPtr:
2545 // icmp pred inttoptr(X), null -> icmp pred X, 0
2546 if (RHSC->isNullValue() && TD &&
2547 TD->getIntPtrType(RHSC->getType()) ==
2548 LHSI->getOperand(0)->getType())
2549 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2550 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2553 case Instruction::Load:
2554 // Try to optimize things like "A[i] > 4" to index computations.
2555 if (GetElementPtrInst *GEP =
2556 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2557 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2558 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2559 !cast<LoadInst>(LHSI)->isVolatile())
2560 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2567 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2568 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2569 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2571 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2572 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2573 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2576 // Test to see if the operands of the icmp are casted versions of other
2577 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2579 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2580 if (Op0->getType()->isPointerTy() &&
2581 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2582 // We keep moving the cast from the left operand over to the right
2583 // operand, where it can often be eliminated completely.
2584 Op0 = CI->getOperand(0);
2586 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2587 // so eliminate it as well.
2588 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2589 Op1 = CI2->getOperand(0);
2591 // If Op1 is a constant, we can fold the cast into the constant.
2592 if (Op0->getType() != Op1->getType()) {
2593 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2594 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2596 // Otherwise, cast the RHS right before the icmp
2597 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2600 return new ICmpInst(I.getPredicate(), Op0, Op1);
2604 if (isa<CastInst>(Op0)) {
2605 // Handle the special case of: icmp (cast bool to X), <cst>
2606 // This comes up when you have code like
2609 // For generality, we handle any zero-extension of any operand comparison
2610 // with a constant or another cast from the same type.
2611 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2612 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2616 // Special logic for binary operators.
2617 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2618 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2620 CmpInst::Predicate Pred = I.getPredicate();
2621 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2622 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2623 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2624 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2625 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2626 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2627 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2628 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2629 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2631 // Analyze the case when either Op0 or Op1 is an add instruction.
2632 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2633 Value *A = 0, *B = 0, *C = 0, *D = 0;
2634 if (BO0 && BO0->getOpcode() == Instruction::Add)
2635 A = BO0->getOperand(0), B = BO0->getOperand(1);
2636 if (BO1 && BO1->getOpcode() == Instruction::Add)
2637 C = BO1->getOperand(0), D = BO1->getOperand(1);
2639 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2640 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2641 return new ICmpInst(Pred, A == Op1 ? B : A,
2642 Constant::getNullValue(Op1->getType()));
2644 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2645 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2646 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2649 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2650 if (A && C && (A == C || A == D || B == C || B == D) &&
2651 NoOp0WrapProblem && NoOp1WrapProblem &&
2652 // Try not to increase register pressure.
2653 BO0->hasOneUse() && BO1->hasOneUse()) {
2654 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2657 // C + B == C + D -> B == D
2660 } else if (A == D) {
2661 // D + B == C + D -> B == C
2664 } else if (B == C) {
2665 // A + C == C + D -> A == D
2670 // A + D == C + D -> A == C
2674 return new ICmpInst(Pred, Y, Z);
2677 // icmp slt (X + -1), Y -> icmp sle X, Y
2678 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2679 match(B, m_AllOnes()))
2680 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2682 // icmp sge (X + -1), Y -> icmp sgt X, Y
2683 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2684 match(B, m_AllOnes()))
2685 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2687 // icmp sle (X + 1), Y -> icmp slt X, Y
2688 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
2690 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2692 // icmp sgt (X + 1), Y -> icmp sge X, Y
2693 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
2695 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2697 // if C1 has greater magnitude than C2:
2698 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2699 // s.t. C3 = C1 - C2
2701 // if C2 has greater magnitude than C1:
2702 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2703 // s.t. C3 = C2 - C1
2704 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2705 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2706 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2707 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2708 const APInt &AP1 = C1->getValue();
2709 const APInt &AP2 = C2->getValue();
2710 if (AP1.isNegative() == AP2.isNegative()) {
2711 APInt AP1Abs = C1->getValue().abs();
2712 APInt AP2Abs = C2->getValue().abs();
2713 if (AP1Abs.uge(AP2Abs)) {
2714 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2715 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2716 return new ICmpInst(Pred, NewAdd, C);
2718 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2719 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2720 return new ICmpInst(Pred, A, NewAdd);
2726 // Analyze the case when either Op0 or Op1 is a sub instruction.
2727 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2728 A = 0; B = 0; C = 0; D = 0;
2729 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2730 A = BO0->getOperand(0), B = BO0->getOperand(1);
2731 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2732 C = BO1->getOperand(0), D = BO1->getOperand(1);
2734 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2735 if (A == Op1 && NoOp0WrapProblem)
2736 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2738 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2739 if (C == Op0 && NoOp1WrapProblem)
2740 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2742 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2743 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2744 // Try not to increase register pressure.
2745 BO0->hasOneUse() && BO1->hasOneUse())
2746 return new ICmpInst(Pred, A, C);
2748 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2749 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2750 // Try not to increase register pressure.
2751 BO0->hasOneUse() && BO1->hasOneUse())
2752 return new ICmpInst(Pred, D, B);
2754 BinaryOperator *SRem = NULL;
2755 // icmp (srem X, Y), Y
2756 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2757 Op1 == BO0->getOperand(1))
2759 // icmp Y, (srem X, Y)
2760 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2761 Op0 == BO1->getOperand(1))
2764 // We don't check hasOneUse to avoid increasing register pressure because
2765 // the value we use is the same value this instruction was already using.
2766 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2768 case ICmpInst::ICMP_EQ:
2769 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2770 case ICmpInst::ICMP_NE:
2771 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2772 case ICmpInst::ICMP_SGT:
2773 case ICmpInst::ICMP_SGE:
2774 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2775 Constant::getAllOnesValue(SRem->getType()));
2776 case ICmpInst::ICMP_SLT:
2777 case ICmpInst::ICMP_SLE:
2778 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2779 Constant::getNullValue(SRem->getType()));
2783 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2784 BO0->hasOneUse() && BO1->hasOneUse() &&
2785 BO0->getOperand(1) == BO1->getOperand(1)) {
2786 switch (BO0->getOpcode()) {
2788 case Instruction::Add:
2789 case Instruction::Sub:
2790 case Instruction::Xor:
2791 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2792 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2793 BO1->getOperand(0));
2794 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2795 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2796 if (CI->getValue().isSignBit()) {
2797 ICmpInst::Predicate Pred = I.isSigned()
2798 ? I.getUnsignedPredicate()
2799 : I.getSignedPredicate();
2800 return new ICmpInst(Pred, BO0->getOperand(0),
2801 BO1->getOperand(0));
2804 if (CI->isMaxValue(true)) {
2805 ICmpInst::Predicate Pred = I.isSigned()
2806 ? I.getUnsignedPredicate()
2807 : I.getSignedPredicate();
2808 Pred = I.getSwappedPredicate(Pred);
2809 return new ICmpInst(Pred, BO0->getOperand(0),
2810 BO1->getOperand(0));
2814 case Instruction::Mul:
2815 if (!I.isEquality())
2818 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2819 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2820 // Mask = -1 >> count-trailing-zeros(Cst).
2821 if (!CI->isZero() && !CI->isOne()) {
2822 const APInt &AP = CI->getValue();
2823 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2824 APInt::getLowBitsSet(AP.getBitWidth(),
2826 AP.countTrailingZeros()));
2827 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2828 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2829 return new ICmpInst(I.getPredicate(), And1, And2);
2833 case Instruction::UDiv:
2834 case Instruction::LShr:
2838 case Instruction::SDiv:
2839 case Instruction::AShr:
2840 if (!BO0->isExact() || !BO1->isExact())
2842 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2843 BO1->getOperand(0));
2844 case Instruction::Shl: {
2845 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2846 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2849 if (!NSW && I.isSigned())
2851 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2852 BO1->getOperand(0));
2859 // Transform (A & ~B) == 0 --> (A & B) != 0
2860 // and (A & ~B) != 0 --> (A & B) == 0
2861 // if A is a power of 2.
2862 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2863 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
2864 return new ICmpInst(I.getInversePredicate(),
2865 Builder->CreateAnd(A, B),
2868 // ~x < ~y --> y < x
2869 // ~x < cst --> ~cst < x
2870 if (match(Op0, m_Not(m_Value(A)))) {
2871 if (match(Op1, m_Not(m_Value(B))))
2872 return new ICmpInst(I.getPredicate(), B, A);
2873 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2874 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2877 // (a+b) <u a --> llvm.uadd.with.overflow.
2878 // (a+b) <u b --> llvm.uadd.with.overflow.
2879 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2880 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2881 (Op1 == A || Op1 == B))
2882 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2885 // a >u (a+b) --> llvm.uadd.with.overflow.
2886 // b >u (a+b) --> llvm.uadd.with.overflow.
2887 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2888 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2889 (Op0 == A || Op0 == B))
2890 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2894 if (I.isEquality()) {
2895 Value *A, *B, *C, *D;
2897 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2898 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2899 Value *OtherVal = A == Op1 ? B : A;
2900 return new ICmpInst(I.getPredicate(), OtherVal,
2901 Constant::getNullValue(A->getType()));
2904 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2905 // A^c1 == C^c2 --> A == C^(c1^c2)
2906 ConstantInt *C1, *C2;
2907 if (match(B, m_ConstantInt(C1)) &&
2908 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2909 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
2910 Value *Xor = Builder->CreateXor(C, NC);
2911 return new ICmpInst(I.getPredicate(), A, Xor);
2914 // A^B == A^D -> B == D
2915 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2916 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2917 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2918 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2922 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2923 (A == Op0 || B == Op0)) {
2924 // A == (A^B) -> B == 0
2925 Value *OtherVal = A == Op0 ? B : A;
2926 return new ICmpInst(I.getPredicate(), OtherVal,
2927 Constant::getNullValue(A->getType()));
2930 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2931 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2932 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2933 Value *X = 0, *Y = 0, *Z = 0;
2936 X = B; Y = D; Z = A;
2937 } else if (A == D) {
2938 X = B; Y = C; Z = A;
2939 } else if (B == C) {
2940 X = A; Y = D; Z = B;
2941 } else if (B == D) {
2942 X = A; Y = C; Z = B;
2945 if (X) { // Build (X^Y) & Z
2946 Op1 = Builder->CreateXor(X, Y);
2947 Op1 = Builder->CreateAnd(Op1, Z);
2948 I.setOperand(0, Op1);
2949 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2954 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2955 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2957 if ((Op0->hasOneUse() &&
2958 match(Op0, m_ZExt(m_Value(A))) &&
2959 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2960 (Op1->hasOneUse() &&
2961 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2962 match(Op1, m_ZExt(m_Value(A))))) {
2963 APInt Pow2 = Cst1->getValue() + 1;
2964 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2965 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2966 return new ICmpInst(I.getPredicate(), A,
2967 Builder->CreateTrunc(B, A->getType()));
2970 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
2971 // For lshr and ashr pairs.
2972 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
2973 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
2974 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
2975 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
2976 unsigned TypeBits = Cst1->getBitWidth();
2977 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
2978 if (ShAmt < TypeBits && ShAmt != 0) {
2979 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
2980 ? ICmpInst::ICMP_UGE
2981 : ICmpInst::ICMP_ULT;
2982 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
2983 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
2984 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
2988 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2989 // "icmp (and X, mask), cst"
2991 if (Op0->hasOneUse() &&
2992 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2993 m_ConstantInt(ShAmt))))) &&
2994 match(Op1, m_ConstantInt(Cst1)) &&
2995 // Only do this when A has multiple uses. This is most important to do
2996 // when it exposes other optimizations.
2998 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3000 if (ShAmt < ASize) {
3002 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3005 APInt CmpV = Cst1->getValue().zext(ASize);
3008 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3009 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3015 Value *X; ConstantInt *Cst;
3017 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3018 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3021 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3022 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3024 return Changed ? &I : 0;
3027 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3029 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3032 if (!isa<ConstantFP>(RHSC)) return 0;
3033 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3035 // Get the width of the mantissa. We don't want to hack on conversions that
3036 // might lose information from the integer, e.g. "i64 -> float"
3037 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3038 if (MantissaWidth == -1) return 0; // Unknown.
3040 // Check to see that the input is converted from an integer type that is small
3041 // enough that preserves all bits. TODO: check here for "known" sign bits.
3042 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3043 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
3045 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3046 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3050 // If the conversion would lose info, don't hack on this.
3051 if ((int)InputSize > MantissaWidth)
3054 // Otherwise, we can potentially simplify the comparison. We know that it
3055 // will always come through as an integer value and we know the constant is
3056 // not a NAN (it would have been previously simplified).
3057 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3059 ICmpInst::Predicate Pred;
3060 switch (I.getPredicate()) {
3061 default: llvm_unreachable("Unexpected predicate!");
3062 case FCmpInst::FCMP_UEQ:
3063 case FCmpInst::FCMP_OEQ:
3064 Pred = ICmpInst::ICMP_EQ;
3066 case FCmpInst::FCMP_UGT:
3067 case FCmpInst::FCMP_OGT:
3068 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3070 case FCmpInst::FCMP_UGE:
3071 case FCmpInst::FCMP_OGE:
3072 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3074 case FCmpInst::FCMP_ULT:
3075 case FCmpInst::FCMP_OLT:
3076 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3078 case FCmpInst::FCMP_ULE:
3079 case FCmpInst::FCMP_OLE:
3080 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3082 case FCmpInst::FCMP_UNE:
3083 case FCmpInst::FCMP_ONE:
3084 Pred = ICmpInst::ICMP_NE;
3086 case FCmpInst::FCMP_ORD:
3087 return ReplaceInstUsesWith(I, Builder->getTrue());
3088 case FCmpInst::FCMP_UNO:
3089 return ReplaceInstUsesWith(I, Builder->getFalse());
3092 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3094 // Now we know that the APFloat is a normal number, zero or inf.
3096 // See if the FP constant is too large for the integer. For example,
3097 // comparing an i8 to 300.0.
3098 unsigned IntWidth = IntTy->getScalarSizeInBits();
3101 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3102 // and large values.
3103 APFloat SMax(RHS.getSemantics());
3104 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3105 APFloat::rmNearestTiesToEven);
3106 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3107 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3108 Pred == ICmpInst::ICMP_SLE)
3109 return ReplaceInstUsesWith(I, Builder->getTrue());
3110 return ReplaceInstUsesWith(I, Builder->getFalse());
3113 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3114 // +INF and large values.
3115 APFloat UMax(RHS.getSemantics());
3116 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3117 APFloat::rmNearestTiesToEven);
3118 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3119 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3120 Pred == ICmpInst::ICMP_ULE)
3121 return ReplaceInstUsesWith(I, Builder->getTrue());
3122 return ReplaceInstUsesWith(I, Builder->getFalse());
3127 // See if the RHS value is < SignedMin.
3128 APFloat SMin(RHS.getSemantics());
3129 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3130 APFloat::rmNearestTiesToEven);
3131 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3132 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3133 Pred == ICmpInst::ICMP_SGE)
3134 return ReplaceInstUsesWith(I, Builder->getTrue());
3135 return ReplaceInstUsesWith(I, Builder->getFalse());
3138 // See if the RHS value is < UnsignedMin.
3139 APFloat SMin(RHS.getSemantics());
3140 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3141 APFloat::rmNearestTiesToEven);
3142 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3143 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3144 Pred == ICmpInst::ICMP_UGE)
3145 return ReplaceInstUsesWith(I, Builder->getTrue());
3146 return ReplaceInstUsesWith(I, Builder->getFalse());
3150 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3151 // [0, UMAX], but it may still be fractional. See if it is fractional by
3152 // casting the FP value to the integer value and back, checking for equality.
3153 // Don't do this for zero, because -0.0 is not fractional.
3154 Constant *RHSInt = LHSUnsigned
3155 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3156 : ConstantExpr::getFPToSI(RHSC, IntTy);
3157 if (!RHS.isZero()) {
3158 bool Equal = LHSUnsigned
3159 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3160 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3162 // If we had a comparison against a fractional value, we have to adjust
3163 // the compare predicate and sometimes the value. RHSC is rounded towards
3164 // zero at this point.
3166 default: llvm_unreachable("Unexpected integer comparison!");
3167 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3168 return ReplaceInstUsesWith(I, Builder->getTrue());
3169 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3170 return ReplaceInstUsesWith(I, Builder->getFalse());
3171 case ICmpInst::ICMP_ULE:
3172 // (float)int <= 4.4 --> int <= 4
3173 // (float)int <= -4.4 --> false
3174 if (RHS.isNegative())
3175 return ReplaceInstUsesWith(I, Builder->getFalse());
3177 case ICmpInst::ICMP_SLE:
3178 // (float)int <= 4.4 --> int <= 4
3179 // (float)int <= -4.4 --> int < -4
3180 if (RHS.isNegative())
3181 Pred = ICmpInst::ICMP_SLT;
3183 case ICmpInst::ICMP_ULT:
3184 // (float)int < -4.4 --> false
3185 // (float)int < 4.4 --> int <= 4
3186 if (RHS.isNegative())
3187 return ReplaceInstUsesWith(I, Builder->getFalse());
3188 Pred = ICmpInst::ICMP_ULE;
3190 case ICmpInst::ICMP_SLT:
3191 // (float)int < -4.4 --> int < -4
3192 // (float)int < 4.4 --> int <= 4
3193 if (!RHS.isNegative())
3194 Pred = ICmpInst::ICMP_SLE;
3196 case ICmpInst::ICMP_UGT:
3197 // (float)int > 4.4 --> int > 4
3198 // (float)int > -4.4 --> true
3199 if (RHS.isNegative())
3200 return ReplaceInstUsesWith(I, Builder->getTrue());
3202 case ICmpInst::ICMP_SGT:
3203 // (float)int > 4.4 --> int > 4
3204 // (float)int > -4.4 --> int >= -4
3205 if (RHS.isNegative())
3206 Pred = ICmpInst::ICMP_SGE;
3208 case ICmpInst::ICMP_UGE:
3209 // (float)int >= -4.4 --> true
3210 // (float)int >= 4.4 --> int > 4
3211 if (RHS.isNegative())
3212 return ReplaceInstUsesWith(I, Builder->getTrue());
3213 Pred = ICmpInst::ICMP_UGT;
3215 case ICmpInst::ICMP_SGE:
3216 // (float)int >= -4.4 --> int >= -4
3217 // (float)int >= 4.4 --> int > 4
3218 if (!RHS.isNegative())
3219 Pred = ICmpInst::ICMP_SGT;
3225 // Lower this FP comparison into an appropriate integer version of the
3227 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3230 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3231 bool Changed = false;
3233 /// Orders the operands of the compare so that they are listed from most
3234 /// complex to least complex. This puts constants before unary operators,
3235 /// before binary operators.
3236 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3241 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3243 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
3244 return ReplaceInstUsesWith(I, V);
3246 // Simplify 'fcmp pred X, X'
3248 switch (I.getPredicate()) {
3249 default: llvm_unreachable("Unknown predicate!");
3250 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3251 case FCmpInst::FCMP_ULT: // True if unordered or less than
3252 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3253 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3254 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3255 I.setPredicate(FCmpInst::FCMP_UNO);
3256 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3259 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3260 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3261 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3262 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3263 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3264 I.setPredicate(FCmpInst::FCMP_ORD);
3265 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3270 // Handle fcmp with constant RHS
3271 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3272 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3273 switch (LHSI->getOpcode()) {
3274 case Instruction::FPExt: {
3275 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3276 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3277 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3281 const fltSemantics *Sem;
3282 // FIXME: This shouldn't be here.
3283 if (LHSExt->getSrcTy()->isHalfTy())
3284 Sem = &APFloat::IEEEhalf;
3285 else if (LHSExt->getSrcTy()->isFloatTy())
3286 Sem = &APFloat::IEEEsingle;
3287 else if (LHSExt->getSrcTy()->isDoubleTy())
3288 Sem = &APFloat::IEEEdouble;
3289 else if (LHSExt->getSrcTy()->isFP128Ty())
3290 Sem = &APFloat::IEEEquad;
3291 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3292 Sem = &APFloat::x87DoubleExtended;
3293 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3294 Sem = &APFloat::PPCDoubleDouble;
3299 APFloat F = RHSF->getValueAPF();
3300 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3302 // Avoid lossy conversions and denormals. Zero is a special case
3303 // that's OK to convert.
3307 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3308 APFloat::cmpLessThan) || Fabs.isZero()))
3310 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3311 ConstantFP::get(RHSC->getContext(), F));
3314 case Instruction::PHI:
3315 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3316 // block. If in the same block, we're encouraging jump threading. If
3317 // not, we are just pessimizing the code by making an i1 phi.
3318 if (LHSI->getParent() == I.getParent())
3319 if (Instruction *NV = FoldOpIntoPhi(I))
3322 case Instruction::SIToFP:
3323 case Instruction::UIToFP:
3324 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3327 case Instruction::Select: {
3328 // If either operand of the select is a constant, we can fold the
3329 // comparison into the select arms, which will cause one to be
3330 // constant folded and the select turned into a bitwise or.
3331 Value *Op1 = 0, *Op2 = 0;
3332 if (LHSI->hasOneUse()) {
3333 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3334 // Fold the known value into the constant operand.
3335 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3336 // Insert a new FCmp of the other select operand.
3337 Op2 = Builder->CreateFCmp(I.getPredicate(),
3338 LHSI->getOperand(2), RHSC, I.getName());
3339 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3340 // Fold the known value into the constant operand.
3341 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3342 // Insert a new FCmp of the other select operand.
3343 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
3349 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3352 case Instruction::FSub: {
3353 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3355 if (match(LHSI, m_FNeg(m_Value(Op))))
3356 return new FCmpInst(I.getSwappedPredicate(), Op,
3357 ConstantExpr::getFNeg(RHSC));
3360 case Instruction::Load:
3361 if (GetElementPtrInst *GEP =
3362 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3363 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3364 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3365 !cast<LoadInst>(LHSI)->isVolatile())
3366 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3370 case Instruction::Call: {
3371 CallInst *CI = cast<CallInst>(LHSI);
3373 // Various optimization for fabs compared with zero.
3374 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3375 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3377 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3378 Func == LibFunc::fabsl) {
3379 switch (I.getPredicate()) {
3381 // fabs(x) < 0 --> false
3382 case FCmpInst::FCMP_OLT:
3383 return ReplaceInstUsesWith(I, Builder->getFalse());
3384 // fabs(x) > 0 --> x != 0
3385 case FCmpInst::FCMP_OGT:
3386 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3388 // fabs(x) <= 0 --> x == 0
3389 case FCmpInst::FCMP_OLE:
3390 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3392 // fabs(x) >= 0 --> !isnan(x)
3393 case FCmpInst::FCMP_OGE:
3394 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3396 // fabs(x) == 0 --> x == 0
3397 // fabs(x) != 0 --> x != 0
3398 case FCmpInst::FCMP_OEQ:
3399 case FCmpInst::FCMP_UEQ:
3400 case FCmpInst::FCMP_ONE:
3401 case FCmpInst::FCMP_UNE:
3402 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3411 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3413 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3414 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3416 // fcmp (fpext x), (fpext y) -> fcmp x, y
3417 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3418 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3419 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3420 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3421 RHSExt->getOperand(0));
3423 return Changed ? &I : 0;