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 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
399 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
401 // If the comparison is only true for one or two elements, emit direct
403 if (SecondTrueElement != Overdefined) {
404 // None true -> false.
405 if (FirstTrueElement == Undefined)
406 return ReplaceInstUsesWith(ICI, Builder->getFalse());
408 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
410 // True for one element -> 'i == 47'.
411 if (SecondTrueElement == Undefined)
412 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
414 // True for two elements -> 'i == 47 | i == 72'.
415 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
416 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
417 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
418 return BinaryOperator::CreateOr(C1, C2);
421 // If the comparison is only false for one or two elements, emit direct
423 if (SecondFalseElement != Overdefined) {
424 // None false -> true.
425 if (FirstFalseElement == Undefined)
426 return ReplaceInstUsesWith(ICI, Builder->getTrue());
428 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
430 // False for one element -> 'i != 47'.
431 if (SecondFalseElement == Undefined)
432 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
434 // False for two elements -> 'i != 47 & i != 72'.
435 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
436 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
437 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
438 return BinaryOperator::CreateAnd(C1, C2);
441 // If the comparison can be replaced with a range comparison for the elements
442 // where it is true, emit the range check.
443 if (TrueRangeEnd != Overdefined) {
444 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
446 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
447 if (FirstTrueElement) {
448 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
449 Idx = Builder->CreateAdd(Idx, Offs);
452 Value *End = ConstantInt::get(Idx->getType(),
453 TrueRangeEnd-FirstTrueElement+1);
454 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
457 // False range check.
458 if (FalseRangeEnd != Overdefined) {
459 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
460 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
461 if (FirstFalseElement) {
462 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
463 Idx = Builder->CreateAdd(Idx, Offs);
466 Value *End = ConstantInt::get(Idx->getType(),
467 FalseRangeEnd-FirstFalseElement);
468 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
472 // If a magic bitvector captures the entire comparison state
473 // of this load, replace it with computation that does:
474 // ((magic_cst >> i) & 1) != 0
478 // Look for an appropriate type:
479 // - The type of Idx if the magic fits
480 // - The smallest fitting legal type if we have a DataLayout
482 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
485 Ty = TD->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
486 else if (ArrayElementCount <= 32)
487 Ty = Type::getInt32Ty(Init->getContext());
490 Value *V = Builder->CreateIntCast(Idx, Ty, false);
491 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
492 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
493 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
501 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
502 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
503 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
504 /// be complex, and scales are involved. The above expression would also be
505 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
506 /// This later form is less amenable to optimization though, and we are allowed
507 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
509 /// If we can't emit an optimized form for this expression, this returns null.
511 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
512 DataLayout &TD = *IC.getDataLayout();
513 gep_type_iterator GTI = gep_type_begin(GEP);
515 // Check to see if this gep only has a single variable index. If so, and if
516 // any constant indices are a multiple of its scale, then we can compute this
517 // in terms of the scale of the variable index. For example, if the GEP
518 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
519 // because the expression will cross zero at the same point.
520 unsigned i, e = GEP->getNumOperands();
522 for (i = 1; i != e; ++i, ++GTI) {
523 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
524 // Compute the aggregate offset of constant indices.
525 if (CI->isZero()) continue;
527 // Handle a struct index, which adds its field offset to the pointer.
528 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
529 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
531 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
532 Offset += Size*CI->getSExtValue();
535 // Found our variable index.
540 // If there are no variable indices, we must have a constant offset, just
541 // evaluate it the general way.
542 if (i == e) return 0;
544 Value *VariableIdx = GEP->getOperand(i);
545 // Determine the scale factor of the variable element. For example, this is
546 // 4 if the variable index is into an array of i32.
547 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
549 // Verify that there are no other variable indices. If so, emit the hard way.
550 for (++i, ++GTI; i != e; ++i, ++GTI) {
551 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
554 // Compute the aggregate offset of constant indices.
555 if (CI->isZero()) continue;
557 // Handle a struct index, which adds its field offset to the pointer.
558 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
559 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
561 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
562 Offset += Size*CI->getSExtValue();
568 // Okay, we know we have a single variable index, which must be a
569 // pointer/array/vector index. If there is no offset, life is simple, return
571 Type *IntPtrTy = TD.getIntPtrType(GEP->getOperand(0)->getType());
572 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
574 // Cast to intptrty in case a truncation occurs. If an extension is needed,
575 // we don't need to bother extending: the extension won't affect where the
576 // computation crosses zero.
577 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
578 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
583 // Otherwise, there is an index. The computation we will do will be modulo
584 // the pointer size, so get it.
585 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
587 Offset &= PtrSizeMask;
588 VariableScale &= PtrSizeMask;
590 // To do this transformation, any constant index must be a multiple of the
591 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
592 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
593 // multiple of the variable scale.
594 int64_t NewOffs = Offset / (int64_t)VariableScale;
595 if (Offset != NewOffs*(int64_t)VariableScale)
598 // Okay, we can do this evaluation. Start by converting the index to intptr.
599 if (VariableIdx->getType() != IntPtrTy)
600 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
602 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
603 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
606 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
607 /// else. At this point we know that the GEP is on the LHS of the comparison.
608 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
609 ICmpInst::Predicate Cond,
611 // Don't transform signed compares of GEPs into index compares. Even if the
612 // GEP is inbounds, the final add of the base pointer can have signed overflow
613 // and would change the result of the icmp.
614 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
615 // the maximum signed value for the pointer type.
616 if (ICmpInst::isSigned(Cond))
619 // Look through bitcasts.
620 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
621 RHS = BCI->getOperand(0);
623 Value *PtrBase = GEPLHS->getOperand(0);
624 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
625 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
626 // This transformation (ignoring the base and scales) is valid because we
627 // know pointers can't overflow since the gep is inbounds. See if we can
628 // output an optimized form.
629 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
631 // If not, synthesize the offset the hard way.
633 Offset = EmitGEPOffset(GEPLHS);
634 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
635 Constant::getNullValue(Offset->getType()));
636 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
637 // If the base pointers are different, but the indices are the same, just
638 // compare the base pointer.
639 if (PtrBase != GEPRHS->getOperand(0)) {
640 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
641 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
642 GEPRHS->getOperand(0)->getType();
644 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
645 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
646 IndicesTheSame = false;
650 // If all indices are the same, just compare the base pointers.
652 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
654 // If we're comparing GEPs with two base pointers that only differ in type
655 // and both GEPs have only constant indices or just one use, then fold
656 // the compare with the adjusted indices.
657 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
658 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
659 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
660 PtrBase->stripPointerCasts() ==
661 GEPRHS->getOperand(0)->stripPointerCasts()) {
662 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
663 EmitGEPOffset(GEPLHS),
664 EmitGEPOffset(GEPRHS));
665 return ReplaceInstUsesWith(I, Cmp);
668 // Otherwise, the base pointers are different and the indices are
669 // different, bail out.
673 // If one of the GEPs has all zero indices, recurse.
674 bool AllZeros = true;
675 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
676 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
677 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
682 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
683 ICmpInst::getSwappedPredicate(Cond), I);
685 // If the other GEP has all zero indices, recurse.
687 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
688 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
689 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
694 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
696 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
697 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
698 // If the GEPs only differ by one index, compare it.
699 unsigned NumDifferences = 0; // Keep track of # differences.
700 unsigned DiffOperand = 0; // The operand that differs.
701 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
702 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
703 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
704 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
705 // Irreconcilable differences.
709 if (NumDifferences++) break;
714 if (NumDifferences == 0) // SAME GEP?
715 return ReplaceInstUsesWith(I, // No comparison is needed here.
716 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
718 else if (NumDifferences == 1 && GEPsInBounds) {
719 Value *LHSV = GEPLHS->getOperand(DiffOperand);
720 Value *RHSV = GEPRHS->getOperand(DiffOperand);
721 // Make sure we do a signed comparison here.
722 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
726 // Only lower this if the icmp is the only user of the GEP or if we expect
727 // the result to fold to a constant!
730 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
731 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
732 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
733 Value *L = EmitGEPOffset(GEPLHS);
734 Value *R = EmitGEPOffset(GEPRHS);
735 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
741 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
742 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
743 Value *X, ConstantInt *CI,
744 ICmpInst::Predicate Pred) {
745 // If we have X+0, exit early (simplifying logic below) and let it get folded
746 // elsewhere. icmp X+0, X -> icmp X, X
748 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
749 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
752 // (X+4) == X -> false.
753 if (Pred == ICmpInst::ICMP_EQ)
754 return ReplaceInstUsesWith(ICI, Builder->getFalse());
756 // (X+4) != X -> true.
757 if (Pred == ICmpInst::ICMP_NE)
758 return ReplaceInstUsesWith(ICI, Builder->getTrue());
760 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
761 // so the values can never be equal. Similarly for all other "or equals"
764 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
765 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
766 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
767 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
769 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
770 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
773 // (X+1) >u X --> X <u (0-1) --> X != 255
774 // (X+2) >u X --> X <u (0-2) --> X <u 254
775 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
776 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
777 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
779 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
780 ConstantInt *SMax = ConstantInt::get(X->getContext(),
781 APInt::getSignedMaxValue(BitWidth));
783 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
784 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
785 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
786 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
787 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
788 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
789 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
790 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
792 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
793 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
794 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
795 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
796 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
797 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
799 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
800 Constant *C = Builder->getInt(CI->getValue()-1);
801 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
804 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
805 /// and CmpRHS are both known to be integer constants.
806 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
807 ConstantInt *DivRHS) {
808 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
809 const APInt &CmpRHSV = CmpRHS->getValue();
811 // FIXME: If the operand types don't match the type of the divide
812 // then don't attempt this transform. The code below doesn't have the
813 // logic to deal with a signed divide and an unsigned compare (and
814 // vice versa). This is because (x /s C1) <s C2 produces different
815 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
816 // (x /u C1) <u C2. Simply casting the operands and result won't
817 // work. :( The if statement below tests that condition and bails
819 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
820 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
822 if (DivRHS->isZero())
823 return 0; // The ProdOV computation fails on divide by zero.
824 if (DivIsSigned && DivRHS->isAllOnesValue())
825 return 0; // The overflow computation also screws up here
826 if (DivRHS->isOne()) {
827 // This eliminates some funny cases with INT_MIN.
828 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
832 // Compute Prod = CI * DivRHS. We are essentially solving an equation
833 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
834 // C2 (CI). By solving for X we can turn this into a range check
835 // instead of computing a divide.
836 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
838 // Determine if the product overflows by seeing if the product is
839 // not equal to the divide. Make sure we do the same kind of divide
840 // as in the LHS instruction that we're folding.
841 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
842 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
844 // Get the ICmp opcode
845 ICmpInst::Predicate Pred = ICI.getPredicate();
847 /// If the division is known to be exact, then there is no remainder from the
848 /// divide, so the covered range size is unit, otherwise it is the divisor.
849 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
851 // Figure out the interval that is being checked. For example, a comparison
852 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
853 // Compute this interval based on the constants involved and the signedness of
854 // the compare/divide. This computes a half-open interval, keeping track of
855 // whether either value in the interval overflows. After analysis each
856 // overflow variable is set to 0 if it's corresponding bound variable is valid
857 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
858 int LoOverflow = 0, HiOverflow = 0;
859 Constant *LoBound = 0, *HiBound = 0;
861 if (!DivIsSigned) { // udiv
862 // e.g. X/5 op 3 --> [15, 20)
864 HiOverflow = LoOverflow = ProdOV;
866 // If this is not an exact divide, then many values in the range collapse
867 // to the same result value.
868 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
871 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
872 if (CmpRHSV == 0) { // (X / pos) op 0
873 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
874 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
876 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
877 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
878 HiOverflow = LoOverflow = ProdOV;
880 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
881 } else { // (X / pos) op neg
882 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
883 HiBound = AddOne(Prod);
884 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
886 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
887 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
890 } else if (DivRHS->isNegative()) { // Divisor is < 0.
892 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
893 if (CmpRHSV == 0) { // (X / neg) op 0
894 // e.g. X/-5 op 0 --> [-4, 5)
895 LoBound = AddOne(RangeSize);
896 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
897 if (HiBound == DivRHS) { // -INTMIN = INTMIN
898 HiOverflow = 1; // [INTMIN+1, overflow)
899 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
901 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
902 // e.g. X/-5 op 3 --> [-19, -14)
903 HiBound = AddOne(Prod);
904 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
906 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
907 } else { // (X / neg) op neg
908 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
909 LoOverflow = HiOverflow = ProdOV;
911 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
914 // Dividing by a negative swaps the condition. LT <-> GT
915 Pred = ICmpInst::getSwappedPredicate(Pred);
918 Value *X = DivI->getOperand(0);
920 default: llvm_unreachable("Unhandled icmp opcode!");
921 case ICmpInst::ICMP_EQ:
922 if (LoOverflow && HiOverflow)
923 return ReplaceInstUsesWith(ICI, Builder->getFalse());
925 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
926 ICmpInst::ICMP_UGE, X, LoBound);
928 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
929 ICmpInst::ICMP_ULT, X, HiBound);
930 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
932 case ICmpInst::ICMP_NE:
933 if (LoOverflow && HiOverflow)
934 return ReplaceInstUsesWith(ICI, Builder->getTrue());
936 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
937 ICmpInst::ICMP_ULT, X, LoBound);
939 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
940 ICmpInst::ICMP_UGE, X, HiBound);
941 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
942 DivIsSigned, false));
943 case ICmpInst::ICMP_ULT:
944 case ICmpInst::ICMP_SLT:
945 if (LoOverflow == +1) // Low bound is greater than input range.
946 return ReplaceInstUsesWith(ICI, Builder->getTrue());
947 if (LoOverflow == -1) // Low bound is less than input range.
948 return ReplaceInstUsesWith(ICI, Builder->getFalse());
949 return new ICmpInst(Pred, X, LoBound);
950 case ICmpInst::ICMP_UGT:
951 case ICmpInst::ICMP_SGT:
952 if (HiOverflow == +1) // High bound greater than input range.
953 return ReplaceInstUsesWith(ICI, Builder->getFalse());
954 if (HiOverflow == -1) // High bound less than input range.
955 return ReplaceInstUsesWith(ICI, Builder->getTrue());
956 if (Pred == ICmpInst::ICMP_UGT)
957 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
958 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
962 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
963 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
964 ConstantInt *ShAmt) {
965 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
967 // Check that the shift amount is in range. If not, don't perform
968 // undefined shifts. When the shift is visited it will be
970 uint32_t TypeBits = CmpRHSV.getBitWidth();
971 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
972 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
975 if (!ICI.isEquality()) {
976 // If we have an unsigned comparison and an ashr, we can't simplify this.
977 // Similarly for signed comparisons with lshr.
978 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
981 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
982 // by a power of 2. Since we already have logic to simplify these,
983 // transform to div and then simplify the resultant comparison.
984 if (Shr->getOpcode() == Instruction::AShr &&
985 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
988 // Revisit the shift (to delete it).
992 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
995 Shr->getOpcode() == Instruction::AShr ?
996 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
997 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
999 ICI.setOperand(0, Tmp);
1001 // If the builder folded the binop, just return it.
1002 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
1006 // Otherwise, fold this div/compare.
1007 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1008 TheDiv->getOpcode() == Instruction::UDiv);
1010 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1011 assert(Res && "This div/cst should have folded!");
1016 // If we are comparing against bits always shifted out, the
1017 // comparison cannot succeed.
1018 APInt Comp = CmpRHSV << ShAmtVal;
1019 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1020 if (Shr->getOpcode() == Instruction::LShr)
1021 Comp = Comp.lshr(ShAmtVal);
1023 Comp = Comp.ashr(ShAmtVal);
1025 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1026 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1027 Constant *Cst = Builder->getInt1(IsICMP_NE);
1028 return ReplaceInstUsesWith(ICI, Cst);
1031 // Otherwise, check to see if the bits shifted out are known to be zero.
1032 // If so, we can compare against the unshifted value:
1033 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1034 if (Shr->hasOneUse() && Shr->isExact())
1035 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1037 if (Shr->hasOneUse()) {
1038 // Otherwise strength reduce the shift into an and.
1039 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1040 Constant *Mask = Builder->getInt(Val);
1042 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1043 Mask, Shr->getName()+".mask");
1044 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1050 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1052 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1055 const APInt &RHSV = RHS->getValue();
1057 switch (LHSI->getOpcode()) {
1058 case Instruction::Trunc:
1059 if (ICI.isEquality() && LHSI->hasOneUse()) {
1060 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1061 // of the high bits truncated out of x are known.
1062 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1063 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1064 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1065 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1067 // If all the high bits are known, we can do this xform.
1068 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1069 // Pull in the high bits from known-ones set.
1070 APInt NewRHS = RHS->getValue().zext(SrcBits);
1071 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1072 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1073 Builder->getInt(NewRHS));
1078 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1079 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1080 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1082 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1083 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1084 Value *CompareVal = LHSI->getOperand(0);
1086 // If the sign bit of the XorCST is not set, there is no change to
1087 // the operation, just stop using the Xor.
1088 if (!XorCST->isNegative()) {
1089 ICI.setOperand(0, CompareVal);
1094 // Was the old condition true if the operand is positive?
1095 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1097 // If so, the new one isn't.
1098 isTrueIfPositive ^= true;
1100 if (isTrueIfPositive)
1101 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1104 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1108 if (LHSI->hasOneUse()) {
1109 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1110 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1111 const APInt &SignBit = XorCST->getValue();
1112 ICmpInst::Predicate Pred = ICI.isSigned()
1113 ? ICI.getUnsignedPredicate()
1114 : ICI.getSignedPredicate();
1115 return new ICmpInst(Pred, LHSI->getOperand(0),
1116 Builder->getInt(RHSV ^ SignBit));
1119 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1120 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1121 const APInt &NotSignBit = XorCST->getValue();
1122 ICmpInst::Predicate Pred = ICI.isSigned()
1123 ? ICI.getUnsignedPredicate()
1124 : ICI.getSignedPredicate();
1125 Pred = ICI.getSwappedPredicate(Pred);
1126 return new ICmpInst(Pred, LHSI->getOperand(0),
1127 Builder->getInt(RHSV ^ NotSignBit));
1131 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1132 // iff -C is a power of 2
1133 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1134 XorCST->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1135 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCST);
1137 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1138 // iff -C is a power of 2
1139 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1140 XorCST->getValue() == -RHSV && RHSV.isPowerOf2())
1141 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCST);
1144 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1145 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1146 LHSI->getOperand(0)->hasOneUse()) {
1147 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1149 // If the LHS is an AND of a truncating cast, we can widen the
1150 // and/compare to be the input width without changing the value
1151 // produced, eliminating a cast.
1152 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1153 // We can do this transformation if either the AND constant does not
1154 // have its sign bit set or if it is an equality comparison.
1155 // Extending a relational comparison when we're checking the sign
1156 // bit would not work.
1157 if (ICI.isEquality() ||
1158 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1160 Builder->CreateAnd(Cast->getOperand(0),
1161 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1162 NewAnd->takeName(LHSI);
1163 return new ICmpInst(ICI.getPredicate(), NewAnd,
1164 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1168 // If the LHS is an AND of a zext, and we have an equality compare, we can
1169 // shrink the and/compare to the smaller type, eliminating the cast.
1170 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1171 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1172 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1173 // should fold the icmp to true/false in that case.
1174 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1176 Builder->CreateAnd(Cast->getOperand(0),
1177 ConstantExpr::getTrunc(AndCST, Ty));
1178 NewAnd->takeName(LHSI);
1179 return new ICmpInst(ICI.getPredicate(), NewAnd,
1180 ConstantExpr::getTrunc(RHS, Ty));
1184 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1185 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1186 // happens a LOT in code produced by the C front-end, for bitfield
1188 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1189 if (Shift && !Shift->isShift())
1193 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1194 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1195 Type *AndTy = AndCST->getType(); // Type of the and.
1197 // We can fold this as long as we can't shift unknown bits
1198 // into the mask. This can only happen with signed shift
1199 // rights, as they sign-extend.
1201 bool CanFold = Shift->isLogicalShift();
1203 // To test for the bad case of the signed shr, see if any
1204 // of the bits shifted in could be tested after the mask.
1205 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1206 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1208 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1209 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1210 AndCST->getValue()) == 0)
1216 if (Shift->getOpcode() == Instruction::Shl)
1217 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1219 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1221 // Check to see if we are shifting out any of the bits being
1223 if (ConstantExpr::get(Shift->getOpcode(),
1224 NewCst, ShAmt) != RHS) {
1225 // If we shifted bits out, the fold is not going to work out.
1226 // As a special case, check to see if this means that the
1227 // result is always true or false now.
1228 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1229 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1230 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1231 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1233 ICI.setOperand(1, NewCst);
1234 Constant *NewAndCST;
1235 if (Shift->getOpcode() == Instruction::Shl)
1236 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1238 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1239 LHSI->setOperand(1, NewAndCST);
1240 LHSI->setOperand(0, Shift->getOperand(0));
1241 Worklist.Add(Shift); // Shift is dead.
1247 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1248 // preferable because it allows the C<<Y expression to be hoisted out
1249 // of a loop if Y is invariant and X is not.
1250 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1251 ICI.isEquality() && !Shift->isArithmeticShift() &&
1252 !isa<Constant>(Shift->getOperand(0))) {
1255 if (Shift->getOpcode() == Instruction::LShr) {
1256 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1258 // Insert a logical shift.
1259 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1262 // Compute X & (C << Y).
1264 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1266 ICI.setOperand(0, NewAnd);
1270 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
1271 // bit set in (X & AndCST) will produce a result greater than RHSV.
1272 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1273 unsigned NTZ = AndCST->getValue().countTrailingZeros();
1274 if ((NTZ < AndCST->getBitWidth()) &&
1275 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
1276 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1277 Constant::getNullValue(RHS->getType()));
1281 // Try to optimize things like "A[i]&42 == 0" to index computations.
1282 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1283 if (GetElementPtrInst *GEP =
1284 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1285 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1286 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1287 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1288 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1289 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1294 // X & -C == -C -> X > u ~C
1295 // X & -C != -C -> X <= u ~C
1296 // iff C is a power of 2
1297 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1298 return new ICmpInst(
1299 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1300 : ICmpInst::ICMP_ULE,
1301 LHSI->getOperand(0), SubOne(RHS));
1304 case Instruction::Or: {
1305 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1308 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1309 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1310 // -> and (icmp eq P, null), (icmp eq Q, null).
1311 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1312 Constant::getNullValue(P->getType()));
1313 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1314 Constant::getNullValue(Q->getType()));
1316 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1317 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1319 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1325 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1326 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1329 // If this is a signed comparison to 0 and the mul is sign preserving,
1330 // use the mul LHS operand instead.
1331 ICmpInst::Predicate pred = ICI.getPredicate();
1332 if (isSignTest(pred, RHS) && !Val->isZero() &&
1333 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1334 return new ICmpInst(Val->isNegative() ?
1335 ICmpInst::getSwappedPredicate(pred) : pred,
1336 LHSI->getOperand(0),
1337 Constant::getNullValue(RHS->getType()));
1342 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1343 uint32_t TypeBits = RHSV.getBitWidth();
1344 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1347 // (1 << X) pred P2 -> X pred Log2(P2)
1348 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1349 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1350 ICmpInst::Predicate Pred = ICI.getPredicate();
1351 if (ICI.isUnsigned()) {
1352 if (!RHSVIsPowerOf2) {
1353 // (1 << X) < 30 -> X <= 4
1354 // (1 << X) <= 30 -> X <= 4
1355 // (1 << X) >= 30 -> X > 4
1356 // (1 << X) > 30 -> X > 4
1357 if (Pred == ICmpInst::ICMP_ULT)
1358 Pred = ICmpInst::ICMP_ULE;
1359 else if (Pred == ICmpInst::ICMP_UGE)
1360 Pred = ICmpInst::ICMP_UGT;
1362 unsigned RHSLog2 = RHSV.logBase2();
1364 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1365 // (1 << X) > 2147483648 -> X > 31 -> false
1366 // (1 << X) <= 2147483648 -> X <= 31 -> true
1367 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1368 if (RHSLog2 == TypeBits-1) {
1369 if (Pred == ICmpInst::ICMP_UGE)
1370 Pred = ICmpInst::ICMP_EQ;
1371 else if (Pred == ICmpInst::ICMP_UGT)
1372 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1373 else if (Pred == ICmpInst::ICMP_ULE)
1374 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1375 else if (Pred == ICmpInst::ICMP_ULT)
1376 Pred = ICmpInst::ICMP_NE;
1379 return new ICmpInst(Pred, X,
1380 ConstantInt::get(RHS->getType(), RHSLog2));
1381 } else if (ICI.isSigned()) {
1382 if (RHSV.isAllOnesValue()) {
1383 // (1 << X) <= -1 -> X == 31
1384 if (Pred == ICmpInst::ICMP_SLE)
1385 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1386 ConstantInt::get(RHS->getType(), TypeBits-1));
1388 // (1 << X) > -1 -> X != 31
1389 if (Pred == ICmpInst::ICMP_SGT)
1390 return new ICmpInst(ICmpInst::ICMP_NE, X,
1391 ConstantInt::get(RHS->getType(), TypeBits-1));
1393 // (1 << X) < 0 -> X == 31
1394 // (1 << X) <= 0 -> X == 31
1395 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1396 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1397 ConstantInt::get(RHS->getType(), TypeBits-1));
1399 // (1 << X) >= 0 -> X != 31
1400 // (1 << X) > 0 -> X != 31
1401 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1402 return new ICmpInst(ICmpInst::ICMP_NE, X,
1403 ConstantInt::get(RHS->getType(), TypeBits-1));
1405 } else if (ICI.isEquality()) {
1407 return new ICmpInst(
1408 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1410 return ReplaceInstUsesWith(
1411 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse()
1412 : Builder->getTrue());
1418 // Check that the shift amount is in range. If not, don't perform
1419 // undefined shifts. When the shift is visited it will be
1421 if (ShAmt->uge(TypeBits))
1424 if (ICI.isEquality()) {
1425 // If we are comparing against bits always shifted out, the
1426 // comparison cannot succeed.
1428 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1430 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1431 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1432 Constant *Cst = Builder->getInt1(IsICMP_NE);
1433 return ReplaceInstUsesWith(ICI, Cst);
1436 // If the shift is NUW, then it is just shifting out zeros, no need for an
1438 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1439 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1440 ConstantExpr::getLShr(RHS, ShAmt));
1442 // If the shift is NSW and we compare to 0, then it is just shifting out
1443 // sign bits, no need for an AND either.
1444 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1445 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1446 ConstantExpr::getLShr(RHS, ShAmt));
1448 if (LHSI->hasOneUse()) {
1449 // Otherwise strength reduce the shift into an and.
1450 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1451 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1452 TypeBits - ShAmtVal));
1455 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1456 return new ICmpInst(ICI.getPredicate(), And,
1457 ConstantExpr::getLShr(RHS, ShAmt));
1461 // If this is a signed comparison to 0 and the shift is sign preserving,
1462 // use the shift LHS operand instead.
1463 ICmpInst::Predicate pred = ICI.getPredicate();
1464 if (isSignTest(pred, RHS) &&
1465 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1466 return new ICmpInst(pred,
1467 LHSI->getOperand(0),
1468 Constant::getNullValue(RHS->getType()));
1470 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1471 bool TrueIfSigned = false;
1472 if (LHSI->hasOneUse() &&
1473 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1474 // (X << 31) <s 0 --> (X&1) != 0
1475 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1476 APInt::getOneBitSet(TypeBits,
1477 TypeBits-ShAmt->getZExtValue()-1));
1479 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1480 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1481 And, Constant::getNullValue(And->getType()));
1484 // Transform (icmp pred iM (shl iM %v, N), CI)
1485 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1486 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1487 // This enables to get rid of the shift in favor of a trunc which can be
1488 // free on the target. It has the additional benefit of comparing to a
1489 // smaller constant, which will be target friendly.
1490 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1491 if (LHSI->hasOneUse() &&
1492 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1493 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1494 Constant *NCI = ConstantExpr::getTrunc(
1495 ConstantExpr::getAShr(RHS,
1496 ConstantInt::get(RHS->getType(), Amt)),
1498 return new ICmpInst(ICI.getPredicate(),
1499 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1506 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1507 case Instruction::AShr: {
1508 // Handle equality comparisons of shift-by-constant.
1509 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1510 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1511 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1515 // Handle exact shr's.
1516 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1517 if (RHSV.isMinValue())
1518 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1523 case Instruction::SDiv:
1524 case Instruction::UDiv:
1525 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1526 // Fold this div into the comparison, producing a range check.
1527 // Determine, based on the divide type, what the range is being
1528 // checked. If there is an overflow on the low or high side, remember
1529 // it, otherwise compute the range [low, hi) bounding the new value.
1530 // See: InsertRangeTest above for the kinds of replacements possible.
1531 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1532 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1537 case Instruction::Sub: {
1538 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1540 const APInt &LHSV = LHSC->getValue();
1542 // C1-X <u C2 -> (X|(C2-1)) == C1
1543 // iff C1 & (C2-1) == C2-1
1544 // C2 is a power of 2
1545 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1546 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1547 return new ICmpInst(ICmpInst::ICMP_EQ,
1548 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1551 // C1-X >u C2 -> (X|C2) != C1
1552 // iff C1 & C2 == C2
1553 // C2+1 is a power of 2
1554 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1555 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1556 return new ICmpInst(ICmpInst::ICMP_NE,
1557 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1561 case Instruction::Add:
1562 // Fold: icmp pred (add X, C1), C2
1563 if (!ICI.isEquality()) {
1564 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1566 const APInt &LHSV = LHSC->getValue();
1568 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1571 if (ICI.isSigned()) {
1572 if (CR.getLower().isSignBit()) {
1573 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1574 Builder->getInt(CR.getUpper()));
1575 } else if (CR.getUpper().isSignBit()) {
1576 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1577 Builder->getInt(CR.getLower()));
1580 if (CR.getLower().isMinValue()) {
1581 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1582 Builder->getInt(CR.getUpper()));
1583 } else if (CR.getUpper().isMinValue()) {
1584 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1585 Builder->getInt(CR.getLower()));
1589 // X-C1 <u C2 -> (X & -C2) == C1
1590 // iff C1 & (C2-1) == 0
1591 // C2 is a power of 2
1592 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1593 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1594 return new ICmpInst(ICmpInst::ICMP_EQ,
1595 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1596 ConstantExpr::getNeg(LHSC));
1598 // X-C1 >u C2 -> (X & ~C2) != C1
1600 // C2+1 is a power of 2
1601 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1602 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1603 return new ICmpInst(ICmpInst::ICMP_NE,
1604 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1605 ConstantExpr::getNeg(LHSC));
1610 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1611 if (ICI.isEquality()) {
1612 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1614 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1615 // the second operand is a constant, simplify a bit.
1616 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1617 switch (BO->getOpcode()) {
1618 case Instruction::SRem:
1619 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1620 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1621 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1622 if (V.sgt(1) && V.isPowerOf2()) {
1624 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1626 return new ICmpInst(ICI.getPredicate(), NewRem,
1627 Constant::getNullValue(BO->getType()));
1631 case Instruction::Add:
1632 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1633 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1634 if (BO->hasOneUse())
1635 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1636 ConstantExpr::getSub(RHS, BOp1C));
1637 } else if (RHSV == 0) {
1638 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1639 // efficiently invertible, or if the add has just this one use.
1640 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1642 if (Value *NegVal = dyn_castNegVal(BOp1))
1643 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1644 if (Value *NegVal = dyn_castNegVal(BOp0))
1645 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1646 if (BO->hasOneUse()) {
1647 Value *Neg = Builder->CreateNeg(BOp1);
1649 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1653 case Instruction::Xor:
1654 // For the xor case, we can xor two constants together, eliminating
1655 // the explicit xor.
1656 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1657 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1658 ConstantExpr::getXor(RHS, BOC));
1659 } else if (RHSV == 0) {
1660 // Replace ((xor A, B) != 0) with (A != B)
1661 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1665 case Instruction::Sub:
1666 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1667 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1668 if (BO->hasOneUse())
1669 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1670 ConstantExpr::getSub(BOp0C, RHS));
1671 } else if (RHSV == 0) {
1672 // Replace ((sub A, B) != 0) with (A != B)
1673 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1677 case Instruction::Or:
1678 // If bits are being or'd in that are not present in the constant we
1679 // are comparing against, then the comparison could never succeed!
1680 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1681 Constant *NotCI = ConstantExpr::getNot(RHS);
1682 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1683 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1687 case Instruction::And:
1688 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1689 // If bits are being compared against that are and'd out, then the
1690 // comparison can never succeed!
1691 if ((RHSV & ~BOC->getValue()) != 0)
1692 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1694 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1695 if (RHS == BOC && RHSV.isPowerOf2())
1696 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1697 ICmpInst::ICMP_NE, LHSI,
1698 Constant::getNullValue(RHS->getType()));
1700 // Don't perform the following transforms if the AND has multiple uses
1701 if (!BO->hasOneUse())
1704 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1705 if (BOC->getValue().isSignBit()) {
1706 Value *X = BO->getOperand(0);
1707 Constant *Zero = Constant::getNullValue(X->getType());
1708 ICmpInst::Predicate pred = isICMP_NE ?
1709 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1710 return new ICmpInst(pred, X, Zero);
1713 // ((X & ~7) == 0) --> X < 8
1714 if (RHSV == 0 && isHighOnes(BOC)) {
1715 Value *X = BO->getOperand(0);
1716 Constant *NegX = ConstantExpr::getNeg(BOC);
1717 ICmpInst::Predicate pred = isICMP_NE ?
1718 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1719 return new ICmpInst(pred, X, NegX);
1723 case Instruction::Mul:
1724 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1725 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1726 // The trivial case (mul X, 0) is handled by InstSimplify
1727 // General case : (mul X, C) != 0 iff X != 0
1728 // (mul X, C) == 0 iff X == 0
1730 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1731 Constant::getNullValue(RHS->getType()));
1737 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1738 // Handle icmp {eq|ne} <intrinsic>, intcst.
1739 switch (II->getIntrinsicID()) {
1740 case Intrinsic::bswap:
1742 ICI.setOperand(0, II->getArgOperand(0));
1743 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1745 case Intrinsic::ctlz:
1746 case Intrinsic::cttz:
1747 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1748 if (RHSV == RHS->getType()->getBitWidth()) {
1750 ICI.setOperand(0, II->getArgOperand(0));
1751 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1755 case Intrinsic::ctpop:
1756 // popcount(A) == 0 -> A == 0 and likewise for !=
1757 if (RHS->isZero()) {
1759 ICI.setOperand(0, II->getArgOperand(0));
1760 ICI.setOperand(1, RHS);
1772 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1773 /// We only handle extending casts so far.
1775 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1776 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1777 Value *LHSCIOp = LHSCI->getOperand(0);
1778 Type *SrcTy = LHSCIOp->getType();
1779 Type *DestTy = LHSCI->getType();
1782 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1783 // integer type is the same size as the pointer type.
1784 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1785 TD->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1787 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1788 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1789 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1790 RHSOp = RHSC->getOperand(0);
1791 // If the pointer types don't match, insert a bitcast.
1792 if (LHSCIOp->getType() != RHSOp->getType())
1793 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1797 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1800 // The code below only handles extension cast instructions, so far.
1802 if (LHSCI->getOpcode() != Instruction::ZExt &&
1803 LHSCI->getOpcode() != Instruction::SExt)
1806 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1807 bool isSignedCmp = ICI.isSigned();
1809 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1810 // Not an extension from the same type?
1811 RHSCIOp = CI->getOperand(0);
1812 if (RHSCIOp->getType() != LHSCIOp->getType())
1815 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1816 // and the other is a zext), then we can't handle this.
1817 if (CI->getOpcode() != LHSCI->getOpcode())
1820 // Deal with equality cases early.
1821 if (ICI.isEquality())
1822 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1824 // A signed comparison of sign extended values simplifies into a
1825 // signed comparison.
1826 if (isSignedCmp && isSignedExt)
1827 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1829 // The other three cases all fold into an unsigned comparison.
1830 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1833 // If we aren't dealing with a constant on the RHS, exit early
1834 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1838 // Compute the constant that would happen if we truncated to SrcTy then
1839 // reextended to DestTy.
1840 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1841 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1844 // If the re-extended constant didn't change...
1846 // Deal with equality cases early.
1847 if (ICI.isEquality())
1848 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1850 // A signed comparison of sign extended values simplifies into a
1851 // signed comparison.
1852 if (isSignedExt && isSignedCmp)
1853 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1855 // The other three cases all fold into an unsigned comparison.
1856 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1859 // The re-extended constant changed so the constant cannot be represented
1860 // in the shorter type. Consequently, we cannot emit a simple comparison.
1861 // All the cases that fold to true or false will have already been handled
1862 // by SimplifyICmpInst, so only deal with the tricky case.
1864 if (isSignedCmp || !isSignedExt)
1867 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1868 // should have been folded away previously and not enter in here.
1870 // We're performing an unsigned comp with a sign extended value.
1871 // This is true if the input is >= 0. [aka >s -1]
1872 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1873 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1875 // Finally, return the value computed.
1876 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1877 return ReplaceInstUsesWith(ICI, Result);
1879 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1880 return BinaryOperator::CreateNot(Result);
1883 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1884 /// I = icmp ugt (add (add A, B), CI2), CI1
1885 /// If this is of the form:
1887 /// if (sum+128 >u 255)
1888 /// Then replace it with llvm.sadd.with.overflow.i8.
1890 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1891 ConstantInt *CI2, ConstantInt *CI1,
1893 // The transformation we're trying to do here is to transform this into an
1894 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1895 // with a narrower add, and discard the add-with-constant that is part of the
1896 // range check (if we can't eliminate it, this isn't profitable).
1898 // In order to eliminate the add-with-constant, the compare can be its only
1900 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1901 if (!AddWithCst->hasOneUse()) return 0;
1903 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1904 if (!CI2->getValue().isPowerOf2()) return 0;
1905 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1906 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1908 // The width of the new add formed is 1 more than the bias.
1911 // Check to see that CI1 is an all-ones value with NewWidth bits.
1912 if (CI1->getBitWidth() == NewWidth ||
1913 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1916 // This is only really a signed overflow check if the inputs have been
1917 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1918 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1919 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1920 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1921 IC.ComputeNumSignBits(B) < NeededSignBits)
1924 // In order to replace the original add with a narrower
1925 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1926 // and truncates that discard the high bits of the add. Verify that this is
1928 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1929 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1931 if (*UI == AddWithCst) continue;
1933 // Only accept truncates for now. We would really like a nice recursive
1934 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1935 // chain to see which bits of a value are actually demanded. If the
1936 // original add had another add which was then immediately truncated, we
1937 // could still do the transformation.
1938 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1940 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1943 // If the pattern matches, truncate the inputs to the narrower type and
1944 // use the sadd_with_overflow intrinsic to efficiently compute both the
1945 // result and the overflow bit.
1946 Module *M = I.getParent()->getParent()->getParent();
1948 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1949 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1952 InstCombiner::BuilderTy *Builder = IC.Builder;
1954 // Put the new code above the original add, in case there are any uses of the
1955 // add between the add and the compare.
1956 Builder->SetInsertPoint(OrigAdd);
1958 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1959 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1960 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1961 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1962 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1964 // The inner add was the result of the narrow add, zero extended to the
1965 // wider type. Replace it with the result computed by the intrinsic.
1966 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1968 // The original icmp gets replaced with the overflow value.
1969 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1972 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1974 // Don't bother doing this transformation for pointers, don't do it for
1976 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1978 // If the add is a constant expr, then we don't bother transforming it.
1979 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1980 if (OrigAdd == 0) return 0;
1982 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1984 // Put the new code above the original add, in case there are any uses of the
1985 // add between the add and the compare.
1986 InstCombiner::BuilderTy *Builder = IC.Builder;
1987 Builder->SetInsertPoint(OrigAdd);
1989 Module *M = I.getParent()->getParent()->getParent();
1990 Type *Ty = LHS->getType();
1991 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1992 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1993 Value *Add = Builder->CreateExtractValue(Call, 0);
1995 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1997 // The original icmp gets replaced with the overflow value.
1998 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2001 // DemandedBitsLHSMask - When performing a comparison against a constant,
2002 // it is possible that not all the bits in the LHS are demanded. This helper
2003 // method computes the mask that IS demanded.
2004 static APInt DemandedBitsLHSMask(ICmpInst &I,
2005 unsigned BitWidth, bool isSignCheck) {
2007 return APInt::getSignBit(BitWidth);
2009 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2010 if (!CI) return APInt::getAllOnesValue(BitWidth);
2011 const APInt &RHS = CI->getValue();
2013 switch (I.getPredicate()) {
2014 // For a UGT comparison, we don't care about any bits that
2015 // correspond to the trailing ones of the comparand. The value of these
2016 // bits doesn't impact the outcome of the comparison, because any value
2017 // greater than the RHS must differ in a bit higher than these due to carry.
2018 case ICmpInst::ICMP_UGT: {
2019 unsigned trailingOnes = RHS.countTrailingOnes();
2020 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2024 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2025 // Any value less than the RHS must differ in a higher bit because of carries.
2026 case ICmpInst::ICMP_ULT: {
2027 unsigned trailingZeros = RHS.countTrailingZeros();
2028 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2033 return APInt::getAllOnesValue(BitWidth);
2038 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2039 /// should be swapped.
2040 /// The descision is based on how many times these two operands are reused
2041 /// as subtract operands and their positions in those instructions.
2042 /// The rational is that several architectures use the same instruction for
2043 /// both subtract and cmp, thus it is better if the order of those operands
2045 /// \return true if Op0 and Op1 should be swapped.
2046 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2047 const Value * Op1) {
2048 // Filter out pointer value as those cannot appears directly in subtract.
2049 // FIXME: we may want to go through inttoptrs or bitcasts.
2050 if (Op0->getType()->isPointerTy())
2052 // Count every uses of both Op0 and Op1 in a subtract.
2053 // Each time Op0 is the first operand, count -1: swapping is bad, the
2054 // subtract has already the same layout as the compare.
2055 // Each time Op0 is the second operand, count +1: swapping is good, the
2056 // subtract has a diffrent layout as the compare.
2057 // At the end, if the benefit is greater than 0, Op0 should come second to
2058 // expose more CSE opportunities.
2059 int GlobalSwapBenefits = 0;
2060 for (Value::const_use_iterator UI = Op0->use_begin(), UIEnd = Op0->use_end(); UI != UIEnd; ++UI) {
2061 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(*UI);
2062 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2064 // If Op0 is the first argument, this is not beneficial to swap the
2066 int LocalSwapBenefits = -1;
2067 unsigned Op1Idx = 1;
2068 if (BinOp->getOperand(Op1Idx) == Op0) {
2070 LocalSwapBenefits = 1;
2072 if (BinOp->getOperand(Op1Idx) != Op1)
2074 GlobalSwapBenefits += LocalSwapBenefits;
2076 return GlobalSwapBenefits > 0;
2079 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2080 bool Changed = false;
2081 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2082 unsigned Op0Cplxity = getComplexity(Op0);
2083 unsigned Op1Cplxity = getComplexity(Op1);
2085 /// Orders the operands of the compare so that they are listed from most
2086 /// complex to least complex. This puts constants before unary operators,
2087 /// before binary operators.
2088 if (Op0Cplxity < Op1Cplxity ||
2089 (Op0Cplxity == Op1Cplxity &&
2090 swapMayExposeCSEOpportunities(Op0, Op1))) {
2092 std::swap(Op0, Op1);
2096 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
2097 return ReplaceInstUsesWith(I, V);
2099 // comparing -val or val with non-zero is the same as just comparing val
2100 // ie, abs(val) != 0 -> val != 0
2101 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2103 Value *Cond, *SelectTrue, *SelectFalse;
2104 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2105 m_Value(SelectFalse)))) {
2106 if (Value *V = dyn_castNegVal(SelectTrue)) {
2107 if (V == SelectFalse)
2108 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2110 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2111 if (V == SelectTrue)
2112 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2117 Type *Ty = Op0->getType();
2119 // icmp's with boolean values can always be turned into bitwise operations
2120 if (Ty->isIntegerTy(1)) {
2121 switch (I.getPredicate()) {
2122 default: llvm_unreachable("Invalid icmp instruction!");
2123 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2124 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2125 return BinaryOperator::CreateNot(Xor);
2127 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2128 return BinaryOperator::CreateXor(Op0, Op1);
2130 case ICmpInst::ICMP_UGT:
2131 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2133 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2134 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2135 return BinaryOperator::CreateAnd(Not, Op1);
2137 case ICmpInst::ICMP_SGT:
2138 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2140 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2141 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2142 return BinaryOperator::CreateAnd(Not, Op0);
2144 case ICmpInst::ICMP_UGE:
2145 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2147 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2148 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2149 return BinaryOperator::CreateOr(Not, Op1);
2151 case ICmpInst::ICMP_SGE:
2152 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2154 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2155 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2156 return BinaryOperator::CreateOr(Not, Op0);
2161 unsigned BitWidth = 0;
2162 if (Ty->isIntOrIntVectorTy())
2163 BitWidth = Ty->getScalarSizeInBits();
2164 else if (TD) // Pointers require TD info to get their size.
2165 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
2167 bool isSignBit = false;
2169 // See if we are doing a comparison with a constant.
2170 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2171 Value *A = 0, *B = 0;
2173 // Match the following pattern, which is a common idiom when writing
2174 // overflow-safe integer arithmetic function. The source performs an
2175 // addition in wider type, and explicitly checks for overflow using
2176 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2177 // sadd_with_overflow intrinsic.
2179 // TODO: This could probably be generalized to handle other overflow-safe
2180 // operations if we worked out the formulas to compute the appropriate
2184 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2186 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2187 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2188 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2189 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2193 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2194 if (I.isEquality() && CI->isZero() &&
2195 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2196 // (icmp cond A B) if cond is equality
2197 return new ICmpInst(I.getPredicate(), A, B);
2200 // If we have an icmp le or icmp ge instruction, turn it into the
2201 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2202 // them being folded in the code below. The SimplifyICmpInst code has
2203 // already handled the edge cases for us, so we just assert on them.
2204 switch (I.getPredicate()) {
2206 case ICmpInst::ICMP_ULE:
2207 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2208 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2209 Builder->getInt(CI->getValue()+1));
2210 case ICmpInst::ICMP_SLE:
2211 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2212 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2213 Builder->getInt(CI->getValue()+1));
2214 case ICmpInst::ICMP_UGE:
2215 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2216 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2217 Builder->getInt(CI->getValue()-1));
2218 case ICmpInst::ICMP_SGE:
2219 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2220 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2221 Builder->getInt(CI->getValue()-1));
2224 // If this comparison is a normal comparison, it demands all
2225 // bits, if it is a sign bit comparison, it only demands the sign bit.
2227 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2230 // See if we can fold the comparison based on range information we can get
2231 // by checking whether bits are known to be zero or one in the input.
2232 if (BitWidth != 0) {
2233 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2234 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2236 if (SimplifyDemandedBits(I.getOperandUse(0),
2237 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2238 Op0KnownZero, Op0KnownOne, 0))
2240 if (SimplifyDemandedBits(I.getOperandUse(1),
2241 APInt::getAllOnesValue(BitWidth),
2242 Op1KnownZero, Op1KnownOne, 0))
2245 // Given the known and unknown bits, compute a range that the LHS could be
2246 // in. Compute the Min, Max and RHS values based on the known bits. For the
2247 // EQ and NE we use unsigned values.
2248 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2249 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2251 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2253 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2256 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2258 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2262 // If Min and Max are known to be the same, then SimplifyDemandedBits
2263 // figured out that the LHS is a constant. Just constant fold this now so
2264 // that code below can assume that Min != Max.
2265 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2266 return new ICmpInst(I.getPredicate(),
2267 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2268 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2269 return new ICmpInst(I.getPredicate(), Op0,
2270 ConstantInt::get(Op1->getType(), Op1Min));
2272 // Based on the range information we know about the LHS, see if we can
2273 // simplify this comparison. For example, (x&4) < 8 is always true.
2274 switch (I.getPredicate()) {
2275 default: llvm_unreachable("Unknown icmp opcode!");
2276 case ICmpInst::ICMP_EQ: {
2277 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2278 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2280 // If all bits are known zero except for one, then we know at most one
2281 // bit is set. If the comparison is against zero, then this is a check
2282 // to see if *that* bit is set.
2283 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2284 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2285 // If the LHS is an AND with the same constant, look through it.
2287 ConstantInt *LHSC = 0;
2288 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2289 LHSC->getValue() != Op0KnownZeroInverted)
2292 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2293 // then turn "((1 << x)&8) == 0" into "x != 3".
2295 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2296 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2297 return new ICmpInst(ICmpInst::ICMP_NE, X,
2298 ConstantInt::get(X->getType(), CmpVal));
2301 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2302 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2304 if (Op0KnownZeroInverted == 1 &&
2305 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2306 return new ICmpInst(ICmpInst::ICMP_NE, X,
2307 ConstantInt::get(X->getType(),
2308 CI->countTrailingZeros()));
2313 case ICmpInst::ICMP_NE: {
2314 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2315 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2317 // If all bits are known zero except for one, then we know at most one
2318 // bit is set. If the comparison is against zero, then this is a check
2319 // to see if *that* bit is set.
2320 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2321 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2322 // If the LHS is an AND with the same constant, look through it.
2324 ConstantInt *LHSC = 0;
2325 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2326 LHSC->getValue() != Op0KnownZeroInverted)
2329 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2330 // then turn "((1 << x)&8) != 0" into "x == 3".
2332 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2333 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2334 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2335 ConstantInt::get(X->getType(), CmpVal));
2338 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2339 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2341 if (Op0KnownZeroInverted == 1 &&
2342 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2343 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2344 ConstantInt::get(X->getType(),
2345 CI->countTrailingZeros()));
2350 case ICmpInst::ICMP_ULT:
2351 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2352 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2353 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2354 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2355 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2356 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2357 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2358 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2359 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2360 Builder->getInt(CI->getValue()-1));
2362 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2363 if (CI->isMinValue(true))
2364 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2365 Constant::getAllOnesValue(Op0->getType()));
2368 case ICmpInst::ICMP_UGT:
2369 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2370 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2371 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2372 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2374 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2375 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2376 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2377 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2378 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2379 Builder->getInt(CI->getValue()+1));
2381 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2382 if (CI->isMaxValue(true))
2383 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2384 Constant::getNullValue(Op0->getType()));
2387 case ICmpInst::ICMP_SLT:
2388 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2389 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2390 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2391 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2392 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2393 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2394 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2395 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2396 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2397 Builder->getInt(CI->getValue()-1));
2400 case ICmpInst::ICMP_SGT:
2401 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2402 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2403 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2404 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2406 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2407 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2408 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2409 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2410 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2411 Builder->getInt(CI->getValue()+1));
2414 case ICmpInst::ICMP_SGE:
2415 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2416 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2417 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2418 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2419 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2421 case ICmpInst::ICMP_SLE:
2422 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2423 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2424 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2425 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2426 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2428 case ICmpInst::ICMP_UGE:
2429 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2430 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2431 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2432 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2433 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2435 case ICmpInst::ICMP_ULE:
2436 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2437 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2438 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2439 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2440 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2444 // Turn a signed comparison into an unsigned one if both operands
2445 // are known to have the same sign.
2447 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2448 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2449 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2452 // Test if the ICmpInst instruction is used exclusively by a select as
2453 // part of a minimum or maximum operation. If so, refrain from doing
2454 // any other folding. This helps out other analyses which understand
2455 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2456 // and CodeGen. And in this case, at least one of the comparison
2457 // operands has at least one user besides the compare (the select),
2458 // which would often largely negate the benefit of folding anyway.
2460 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2461 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2462 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2465 // See if we are doing a comparison between a constant and an instruction that
2466 // can be folded into the comparison.
2467 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2468 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2469 // instruction, see if that instruction also has constants so that the
2470 // instruction can be folded into the icmp
2471 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2472 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2476 // Handle icmp with constant (but not simple integer constant) RHS
2477 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2478 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2479 switch (LHSI->getOpcode()) {
2480 case Instruction::GetElementPtr:
2481 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2482 if (RHSC->isNullValue() &&
2483 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2484 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2485 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2487 case Instruction::PHI:
2488 // Only fold icmp into the PHI if the phi and icmp are in the same
2489 // block. If in the same block, we're encouraging jump threading. If
2490 // not, we are just pessimizing the code by making an i1 phi.
2491 if (LHSI->getParent() == I.getParent())
2492 if (Instruction *NV = FoldOpIntoPhi(I))
2495 case Instruction::Select: {
2496 // If either operand of the select is a constant, we can fold the
2497 // comparison into the select arms, which will cause one to be
2498 // constant folded and the select turned into a bitwise or.
2499 Value *Op1 = 0, *Op2 = 0;
2500 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2501 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2502 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2503 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2505 // We only want to perform this transformation if it will not lead to
2506 // additional code. This is true if either both sides of the select
2507 // fold to a constant (in which case the icmp is replaced with a select
2508 // which will usually simplify) or this is the only user of the
2509 // select (in which case we are trading a select+icmp for a simpler
2511 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2513 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2516 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2518 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2522 case Instruction::IntToPtr:
2523 // icmp pred inttoptr(X), null -> icmp pred X, 0
2524 if (RHSC->isNullValue() && TD &&
2525 TD->getIntPtrType(RHSC->getType()) ==
2526 LHSI->getOperand(0)->getType())
2527 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2528 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2531 case Instruction::Load:
2532 // Try to optimize things like "A[i] > 4" to index computations.
2533 if (GetElementPtrInst *GEP =
2534 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2535 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2536 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2537 !cast<LoadInst>(LHSI)->isVolatile())
2538 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2545 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2546 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2547 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2549 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2550 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2551 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2554 // Test to see if the operands of the icmp are casted versions of other
2555 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2557 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2558 if (Op0->getType()->isPointerTy() &&
2559 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2560 // We keep moving the cast from the left operand over to the right
2561 // operand, where it can often be eliminated completely.
2562 Op0 = CI->getOperand(0);
2564 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2565 // so eliminate it as well.
2566 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2567 Op1 = CI2->getOperand(0);
2569 // If Op1 is a constant, we can fold the cast into the constant.
2570 if (Op0->getType() != Op1->getType()) {
2571 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2572 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2574 // Otherwise, cast the RHS right before the icmp
2575 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2578 return new ICmpInst(I.getPredicate(), Op0, Op1);
2582 if (isa<CastInst>(Op0)) {
2583 // Handle the special case of: icmp (cast bool to X), <cst>
2584 // This comes up when you have code like
2587 // For generality, we handle any zero-extension of any operand comparison
2588 // with a constant or another cast from the same type.
2589 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2590 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2594 // Special logic for binary operators.
2595 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2596 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2598 CmpInst::Predicate Pred = I.getPredicate();
2599 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2600 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2601 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2602 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2603 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2604 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2605 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2606 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2607 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2609 // Analyze the case when either Op0 or Op1 is an add instruction.
2610 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2611 Value *A = 0, *B = 0, *C = 0, *D = 0;
2612 if (BO0 && BO0->getOpcode() == Instruction::Add)
2613 A = BO0->getOperand(0), B = BO0->getOperand(1);
2614 if (BO1 && BO1->getOpcode() == Instruction::Add)
2615 C = BO1->getOperand(0), D = BO1->getOperand(1);
2617 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2618 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2619 return new ICmpInst(Pred, A == Op1 ? B : A,
2620 Constant::getNullValue(Op1->getType()));
2622 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2623 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2624 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2627 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2628 if (A && C && (A == C || A == D || B == C || B == D) &&
2629 NoOp0WrapProblem && NoOp1WrapProblem &&
2630 // Try not to increase register pressure.
2631 BO0->hasOneUse() && BO1->hasOneUse()) {
2632 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2635 // C + B == C + D -> B == D
2638 } else if (A == D) {
2639 // D + B == C + D -> B == C
2642 } else if (B == C) {
2643 // A + C == C + D -> A == D
2648 // A + D == C + D -> A == C
2652 return new ICmpInst(Pred, Y, Z);
2655 // icmp slt (X + -1), Y -> icmp sle X, Y
2656 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2657 match(B, m_AllOnes()))
2658 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2660 // icmp sge (X + -1), Y -> icmp sgt X, Y
2661 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2662 match(B, m_AllOnes()))
2663 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2665 // icmp sle (X + 1), Y -> icmp slt X, Y
2666 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
2668 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2670 // icmp sgt (X + 1), Y -> icmp sge X, Y
2671 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
2673 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2675 // if C1 has greater magnitude than C2:
2676 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2677 // s.t. C3 = C1 - C2
2679 // if C2 has greater magnitude than C1:
2680 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2681 // s.t. C3 = C2 - C1
2682 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2683 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2684 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2685 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2686 const APInt &AP1 = C1->getValue();
2687 const APInt &AP2 = C2->getValue();
2688 if (AP1.isNegative() == AP2.isNegative()) {
2689 APInt AP1Abs = C1->getValue().abs();
2690 APInt AP2Abs = C2->getValue().abs();
2691 if (AP1Abs.uge(AP2Abs)) {
2692 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2693 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2694 return new ICmpInst(Pred, NewAdd, C);
2696 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2697 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2698 return new ICmpInst(Pred, A, NewAdd);
2704 // Analyze the case when either Op0 or Op1 is a sub instruction.
2705 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2706 A = 0; B = 0; C = 0; D = 0;
2707 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2708 A = BO0->getOperand(0), B = BO0->getOperand(1);
2709 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2710 C = BO1->getOperand(0), D = BO1->getOperand(1);
2712 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2713 if (A == Op1 && NoOp0WrapProblem)
2714 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2716 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2717 if (C == Op0 && NoOp1WrapProblem)
2718 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2720 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2721 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2722 // Try not to increase register pressure.
2723 BO0->hasOneUse() && BO1->hasOneUse())
2724 return new ICmpInst(Pred, A, C);
2726 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2727 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2728 // Try not to increase register pressure.
2729 BO0->hasOneUse() && BO1->hasOneUse())
2730 return new ICmpInst(Pred, D, B);
2732 BinaryOperator *SRem = NULL;
2733 // icmp (srem X, Y), Y
2734 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2735 Op1 == BO0->getOperand(1))
2737 // icmp Y, (srem X, Y)
2738 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2739 Op0 == BO1->getOperand(1))
2742 // We don't check hasOneUse to avoid increasing register pressure because
2743 // the value we use is the same value this instruction was already using.
2744 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2746 case ICmpInst::ICMP_EQ:
2747 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2748 case ICmpInst::ICMP_NE:
2749 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2750 case ICmpInst::ICMP_SGT:
2751 case ICmpInst::ICMP_SGE:
2752 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2753 Constant::getAllOnesValue(SRem->getType()));
2754 case ICmpInst::ICMP_SLT:
2755 case ICmpInst::ICMP_SLE:
2756 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2757 Constant::getNullValue(SRem->getType()));
2761 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2762 BO0->hasOneUse() && BO1->hasOneUse() &&
2763 BO0->getOperand(1) == BO1->getOperand(1)) {
2764 switch (BO0->getOpcode()) {
2766 case Instruction::Add:
2767 case Instruction::Sub:
2768 case Instruction::Xor:
2769 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2770 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2771 BO1->getOperand(0));
2772 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2773 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2774 if (CI->getValue().isSignBit()) {
2775 ICmpInst::Predicate Pred = I.isSigned()
2776 ? I.getUnsignedPredicate()
2777 : I.getSignedPredicate();
2778 return new ICmpInst(Pred, BO0->getOperand(0),
2779 BO1->getOperand(0));
2782 if (CI->isMaxValue(true)) {
2783 ICmpInst::Predicate Pred = I.isSigned()
2784 ? I.getUnsignedPredicate()
2785 : I.getSignedPredicate();
2786 Pred = I.getSwappedPredicate(Pred);
2787 return new ICmpInst(Pred, BO0->getOperand(0),
2788 BO1->getOperand(0));
2792 case Instruction::Mul:
2793 if (!I.isEquality())
2796 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2797 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2798 // Mask = -1 >> count-trailing-zeros(Cst).
2799 if (!CI->isZero() && !CI->isOne()) {
2800 const APInt &AP = CI->getValue();
2801 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2802 APInt::getLowBitsSet(AP.getBitWidth(),
2804 AP.countTrailingZeros()));
2805 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2806 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2807 return new ICmpInst(I.getPredicate(), And1, And2);
2811 case Instruction::UDiv:
2812 case Instruction::LShr:
2816 case Instruction::SDiv:
2817 case Instruction::AShr:
2818 if (!BO0->isExact() || !BO1->isExact())
2820 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2821 BO1->getOperand(0));
2822 case Instruction::Shl: {
2823 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2824 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2827 if (!NSW && I.isSigned())
2829 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2830 BO1->getOperand(0));
2837 // Transform (A & ~B) == 0 --> (A & B) != 0
2838 // and (A & ~B) != 0 --> (A & B) == 0
2839 // if A is a power of 2.
2840 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2841 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
2842 return new ICmpInst(I.getInversePredicate(),
2843 Builder->CreateAnd(A, B),
2846 // ~x < ~y --> y < x
2847 // ~x < cst --> ~cst < x
2848 if (match(Op0, m_Not(m_Value(A)))) {
2849 if (match(Op1, m_Not(m_Value(B))))
2850 return new ICmpInst(I.getPredicate(), B, A);
2851 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2852 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2855 // (a+b) <u a --> llvm.uadd.with.overflow.
2856 // (a+b) <u b --> llvm.uadd.with.overflow.
2857 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2858 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2859 (Op1 == A || Op1 == B))
2860 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2863 // a >u (a+b) --> llvm.uadd.with.overflow.
2864 // b >u (a+b) --> llvm.uadd.with.overflow.
2865 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2866 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2867 (Op0 == A || Op0 == B))
2868 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2872 if (I.isEquality()) {
2873 Value *A, *B, *C, *D;
2875 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2876 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2877 Value *OtherVal = A == Op1 ? B : A;
2878 return new ICmpInst(I.getPredicate(), OtherVal,
2879 Constant::getNullValue(A->getType()));
2882 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2883 // A^c1 == C^c2 --> A == C^(c1^c2)
2884 ConstantInt *C1, *C2;
2885 if (match(B, m_ConstantInt(C1)) &&
2886 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2887 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
2888 Value *Xor = Builder->CreateXor(C, NC);
2889 return new ICmpInst(I.getPredicate(), A, Xor);
2892 // A^B == A^D -> B == D
2893 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2894 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2895 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2896 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2900 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2901 (A == Op0 || B == Op0)) {
2902 // A == (A^B) -> B == 0
2903 Value *OtherVal = A == Op0 ? B : A;
2904 return new ICmpInst(I.getPredicate(), OtherVal,
2905 Constant::getNullValue(A->getType()));
2908 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2909 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2910 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2911 Value *X = 0, *Y = 0, *Z = 0;
2914 X = B; Y = D; Z = A;
2915 } else if (A == D) {
2916 X = B; Y = C; Z = A;
2917 } else if (B == C) {
2918 X = A; Y = D; Z = B;
2919 } else if (B == D) {
2920 X = A; Y = C; Z = B;
2923 if (X) { // Build (X^Y) & Z
2924 Op1 = Builder->CreateXor(X, Y);
2925 Op1 = Builder->CreateAnd(Op1, Z);
2926 I.setOperand(0, Op1);
2927 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2932 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2933 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2935 if ((Op0->hasOneUse() &&
2936 match(Op0, m_ZExt(m_Value(A))) &&
2937 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2938 (Op1->hasOneUse() &&
2939 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2940 match(Op1, m_ZExt(m_Value(A))))) {
2941 APInt Pow2 = Cst1->getValue() + 1;
2942 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2943 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2944 return new ICmpInst(I.getPredicate(), A,
2945 Builder->CreateTrunc(B, A->getType()));
2948 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2949 // "icmp (and X, mask), cst"
2951 if (Op0->hasOneUse() &&
2952 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2953 m_ConstantInt(ShAmt))))) &&
2954 match(Op1, m_ConstantInt(Cst1)) &&
2955 // Only do this when A has multiple uses. This is most important to do
2956 // when it exposes other optimizations.
2958 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2960 if (ShAmt < ASize) {
2962 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2965 APInt CmpV = Cst1->getValue().zext(ASize);
2968 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2969 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2975 Value *X; ConstantInt *Cst;
2977 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2978 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
2981 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2982 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
2984 return Changed ? &I : 0;
2987 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2989 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2992 if (!isa<ConstantFP>(RHSC)) return 0;
2993 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2995 // Get the width of the mantissa. We don't want to hack on conversions that
2996 // might lose information from the integer, e.g. "i64 -> float"
2997 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2998 if (MantissaWidth == -1) return 0; // Unknown.
3000 // Check to see that the input is converted from an integer type that is small
3001 // enough that preserves all bits. TODO: check here for "known" sign bits.
3002 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3003 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
3005 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3006 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3010 // If the conversion would lose info, don't hack on this.
3011 if ((int)InputSize > MantissaWidth)
3014 // Otherwise, we can potentially simplify the comparison. We know that it
3015 // will always come through as an integer value and we know the constant is
3016 // not a NAN (it would have been previously simplified).
3017 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3019 ICmpInst::Predicate Pred;
3020 switch (I.getPredicate()) {
3021 default: llvm_unreachable("Unexpected predicate!");
3022 case FCmpInst::FCMP_UEQ:
3023 case FCmpInst::FCMP_OEQ:
3024 Pred = ICmpInst::ICMP_EQ;
3026 case FCmpInst::FCMP_UGT:
3027 case FCmpInst::FCMP_OGT:
3028 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3030 case FCmpInst::FCMP_UGE:
3031 case FCmpInst::FCMP_OGE:
3032 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3034 case FCmpInst::FCMP_ULT:
3035 case FCmpInst::FCMP_OLT:
3036 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3038 case FCmpInst::FCMP_ULE:
3039 case FCmpInst::FCMP_OLE:
3040 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3042 case FCmpInst::FCMP_UNE:
3043 case FCmpInst::FCMP_ONE:
3044 Pred = ICmpInst::ICMP_NE;
3046 case FCmpInst::FCMP_ORD:
3047 return ReplaceInstUsesWith(I, Builder->getTrue());
3048 case FCmpInst::FCMP_UNO:
3049 return ReplaceInstUsesWith(I, Builder->getFalse());
3052 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3054 // Now we know that the APFloat is a normal number, zero or inf.
3056 // See if the FP constant is too large for the integer. For example,
3057 // comparing an i8 to 300.0.
3058 unsigned IntWidth = IntTy->getScalarSizeInBits();
3061 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3062 // and large values.
3063 APFloat SMax(RHS.getSemantics());
3064 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3065 APFloat::rmNearestTiesToEven);
3066 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3067 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3068 Pred == ICmpInst::ICMP_SLE)
3069 return ReplaceInstUsesWith(I, Builder->getTrue());
3070 return ReplaceInstUsesWith(I, Builder->getFalse());
3073 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3074 // +INF and large values.
3075 APFloat UMax(RHS.getSemantics());
3076 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3077 APFloat::rmNearestTiesToEven);
3078 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3079 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3080 Pred == ICmpInst::ICMP_ULE)
3081 return ReplaceInstUsesWith(I, Builder->getTrue());
3082 return ReplaceInstUsesWith(I, Builder->getFalse());
3087 // See if the RHS value is < SignedMin.
3088 APFloat SMin(RHS.getSemantics());
3089 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3090 APFloat::rmNearestTiesToEven);
3091 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3092 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3093 Pred == ICmpInst::ICMP_SGE)
3094 return ReplaceInstUsesWith(I, Builder->getTrue());
3095 return ReplaceInstUsesWith(I, Builder->getFalse());
3098 // See if the RHS value is < UnsignedMin.
3099 APFloat SMin(RHS.getSemantics());
3100 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3101 APFloat::rmNearestTiesToEven);
3102 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3103 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3104 Pred == ICmpInst::ICMP_UGE)
3105 return ReplaceInstUsesWith(I, Builder->getTrue());
3106 return ReplaceInstUsesWith(I, Builder->getFalse());
3110 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3111 // [0, UMAX], but it may still be fractional. See if it is fractional by
3112 // casting the FP value to the integer value and back, checking for equality.
3113 // Don't do this for zero, because -0.0 is not fractional.
3114 Constant *RHSInt = LHSUnsigned
3115 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3116 : ConstantExpr::getFPToSI(RHSC, IntTy);
3117 if (!RHS.isZero()) {
3118 bool Equal = LHSUnsigned
3119 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3120 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3122 // If we had a comparison against a fractional value, we have to adjust
3123 // the compare predicate and sometimes the value. RHSC is rounded towards
3124 // zero at this point.
3126 default: llvm_unreachable("Unexpected integer comparison!");
3127 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3128 return ReplaceInstUsesWith(I, Builder->getTrue());
3129 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3130 return ReplaceInstUsesWith(I, Builder->getFalse());
3131 case ICmpInst::ICMP_ULE:
3132 // (float)int <= 4.4 --> int <= 4
3133 // (float)int <= -4.4 --> false
3134 if (RHS.isNegative())
3135 return ReplaceInstUsesWith(I, Builder->getFalse());
3137 case ICmpInst::ICMP_SLE:
3138 // (float)int <= 4.4 --> int <= 4
3139 // (float)int <= -4.4 --> int < -4
3140 if (RHS.isNegative())
3141 Pred = ICmpInst::ICMP_SLT;
3143 case ICmpInst::ICMP_ULT:
3144 // (float)int < -4.4 --> false
3145 // (float)int < 4.4 --> int <= 4
3146 if (RHS.isNegative())
3147 return ReplaceInstUsesWith(I, Builder->getFalse());
3148 Pred = ICmpInst::ICMP_ULE;
3150 case ICmpInst::ICMP_SLT:
3151 // (float)int < -4.4 --> int < -4
3152 // (float)int < 4.4 --> int <= 4
3153 if (!RHS.isNegative())
3154 Pred = ICmpInst::ICMP_SLE;
3156 case ICmpInst::ICMP_UGT:
3157 // (float)int > 4.4 --> int > 4
3158 // (float)int > -4.4 --> true
3159 if (RHS.isNegative())
3160 return ReplaceInstUsesWith(I, Builder->getTrue());
3162 case ICmpInst::ICMP_SGT:
3163 // (float)int > 4.4 --> int > 4
3164 // (float)int > -4.4 --> int >= -4
3165 if (RHS.isNegative())
3166 Pred = ICmpInst::ICMP_SGE;
3168 case ICmpInst::ICMP_UGE:
3169 // (float)int >= -4.4 --> true
3170 // (float)int >= 4.4 --> int > 4
3171 if (RHS.isNegative())
3172 return ReplaceInstUsesWith(I, Builder->getTrue());
3173 Pred = ICmpInst::ICMP_UGT;
3175 case ICmpInst::ICMP_SGE:
3176 // (float)int >= -4.4 --> int >= -4
3177 // (float)int >= 4.4 --> int > 4
3178 if (!RHS.isNegative())
3179 Pred = ICmpInst::ICMP_SGT;
3185 // Lower this FP comparison into an appropriate integer version of the
3187 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3190 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3191 bool Changed = false;
3193 /// Orders the operands of the compare so that they are listed from most
3194 /// complex to least complex. This puts constants before unary operators,
3195 /// before binary operators.
3196 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3201 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3203 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
3204 return ReplaceInstUsesWith(I, V);
3206 // Simplify 'fcmp pred X, X'
3208 switch (I.getPredicate()) {
3209 default: llvm_unreachable("Unknown predicate!");
3210 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3211 case FCmpInst::FCMP_ULT: // True if unordered or less than
3212 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3213 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3214 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3215 I.setPredicate(FCmpInst::FCMP_UNO);
3216 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3219 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3220 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3221 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3222 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3223 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3224 I.setPredicate(FCmpInst::FCMP_ORD);
3225 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3230 // Handle fcmp with constant RHS
3231 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3232 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3233 switch (LHSI->getOpcode()) {
3234 case Instruction::FPExt: {
3235 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3236 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3237 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3241 const fltSemantics *Sem;
3242 // FIXME: This shouldn't be here.
3243 if (LHSExt->getSrcTy()->isHalfTy())
3244 Sem = &APFloat::IEEEhalf;
3245 else if (LHSExt->getSrcTy()->isFloatTy())
3246 Sem = &APFloat::IEEEsingle;
3247 else if (LHSExt->getSrcTy()->isDoubleTy())
3248 Sem = &APFloat::IEEEdouble;
3249 else if (LHSExt->getSrcTy()->isFP128Ty())
3250 Sem = &APFloat::IEEEquad;
3251 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3252 Sem = &APFloat::x87DoubleExtended;
3253 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3254 Sem = &APFloat::PPCDoubleDouble;
3259 APFloat F = RHSF->getValueAPF();
3260 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3262 // Avoid lossy conversions and denormals. Zero is a special case
3263 // that's OK to convert.
3267 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3268 APFloat::cmpLessThan) || Fabs.isZero()))
3270 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3271 ConstantFP::get(RHSC->getContext(), F));
3274 case Instruction::PHI:
3275 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3276 // block. If in the same block, we're encouraging jump threading. If
3277 // not, we are just pessimizing the code by making an i1 phi.
3278 if (LHSI->getParent() == I.getParent())
3279 if (Instruction *NV = FoldOpIntoPhi(I))
3282 case Instruction::SIToFP:
3283 case Instruction::UIToFP:
3284 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3287 case Instruction::Select: {
3288 // If either operand of the select is a constant, we can fold the
3289 // comparison into the select arms, which will cause one to be
3290 // constant folded and the select turned into a bitwise or.
3291 Value *Op1 = 0, *Op2 = 0;
3292 if (LHSI->hasOneUse()) {
3293 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3294 // Fold the known value into the constant operand.
3295 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3296 // Insert a new FCmp of the other select operand.
3297 Op2 = Builder->CreateFCmp(I.getPredicate(),
3298 LHSI->getOperand(2), RHSC, I.getName());
3299 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3300 // Fold the known value into the constant operand.
3301 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3302 // Insert a new FCmp of the other select operand.
3303 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
3309 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3312 case Instruction::FSub: {
3313 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3315 if (match(LHSI, m_FNeg(m_Value(Op))))
3316 return new FCmpInst(I.getSwappedPredicate(), Op,
3317 ConstantExpr::getFNeg(RHSC));
3320 case Instruction::Load:
3321 if (GetElementPtrInst *GEP =
3322 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3323 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3324 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3325 !cast<LoadInst>(LHSI)->isVolatile())
3326 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3330 case Instruction::Call: {
3331 CallInst *CI = cast<CallInst>(LHSI);
3333 // Various optimization for fabs compared with zero.
3334 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3335 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3337 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3338 Func == LibFunc::fabsl) {
3339 switch (I.getPredicate()) {
3341 // fabs(x) < 0 --> false
3342 case FCmpInst::FCMP_OLT:
3343 return ReplaceInstUsesWith(I, Builder->getFalse());
3344 // fabs(x) > 0 --> x != 0
3345 case FCmpInst::FCMP_OGT:
3346 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3348 // fabs(x) <= 0 --> x == 0
3349 case FCmpInst::FCMP_OLE:
3350 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3352 // fabs(x) >= 0 --> !isnan(x)
3353 case FCmpInst::FCMP_OGE:
3354 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3356 // fabs(x) == 0 --> x == 0
3357 // fabs(x) != 0 --> x != 0
3358 case FCmpInst::FCMP_OEQ:
3359 case FCmpInst::FCMP_UEQ:
3360 case FCmpInst::FCMP_ONE:
3361 case FCmpInst::FCMP_UNE:
3362 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3371 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3373 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3374 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3376 // fcmp (fpext x), (fpext y) -> fcmp x, y
3377 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3378 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3379 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3380 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3381 RHSExt->getOperand(0));
3383 return Changed ? &I : 0;