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(ICmpInst &ICI,
743 Value *X, ConstantInt *CI,
744 ICmpInst::Predicate Pred,
746 // If we have X+0, exit early (simplifying logic below) and let it get folded
747 // elsewhere. icmp X+0, X -> icmp X, X
749 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
750 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
753 // (X+4) == X -> false.
754 if (Pred == ICmpInst::ICMP_EQ)
755 return ReplaceInstUsesWith(ICI, Builder->getFalse());
757 // (X+4) != X -> true.
758 if (Pred == ICmpInst::ICMP_NE)
759 return ReplaceInstUsesWith(ICI, Builder->getTrue());
761 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
762 // so the values can never be equal. Similarly for all other "or equals"
765 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
766 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
767 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
768 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
770 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
771 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
774 // (X+1) >u X --> X <u (0-1) --> X != 255
775 // (X+2) >u X --> X <u (0-2) --> X <u 254
776 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
777 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
778 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
780 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
781 ConstantInt *SMax = ConstantInt::get(X->getContext(),
782 APInt::getSignedMaxValue(BitWidth));
784 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
785 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
786 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
787 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
788 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
789 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
790 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
791 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
793 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
794 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
795 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
796 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
797 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
798 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
800 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
801 Constant *C = Builder->getInt(CI->getValue()-1);
802 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
805 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
806 /// and CmpRHS are both known to be integer constants.
807 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
808 ConstantInt *DivRHS) {
809 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
810 const APInt &CmpRHSV = CmpRHS->getValue();
812 // FIXME: If the operand types don't match the type of the divide
813 // then don't attempt this transform. The code below doesn't have the
814 // logic to deal with a signed divide and an unsigned compare (and
815 // vice versa). This is because (x /s C1) <s C2 produces different
816 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
817 // (x /u C1) <u C2. Simply casting the operands and result won't
818 // work. :( The if statement below tests that condition and bails
820 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
821 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
823 if (DivRHS->isZero())
824 return 0; // The ProdOV computation fails on divide by zero.
825 if (DivIsSigned && DivRHS->isAllOnesValue())
826 return 0; // The overflow computation also screws up here
827 if (DivRHS->isOne()) {
828 // This eliminates some funny cases with INT_MIN.
829 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
833 // Compute Prod = CI * DivRHS. We are essentially solving an equation
834 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
835 // C2 (CI). By solving for X we can turn this into a range check
836 // instead of computing a divide.
837 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
839 // Determine if the product overflows by seeing if the product is
840 // not equal to the divide. Make sure we do the same kind of divide
841 // as in the LHS instruction that we're folding.
842 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
843 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
845 // Get the ICmp opcode
846 ICmpInst::Predicate Pred = ICI.getPredicate();
848 /// If the division is known to be exact, then there is no remainder from the
849 /// divide, so the covered range size is unit, otherwise it is the divisor.
850 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
852 // Figure out the interval that is being checked. For example, a comparison
853 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
854 // Compute this interval based on the constants involved and the signedness of
855 // the compare/divide. This computes a half-open interval, keeping track of
856 // whether either value in the interval overflows. After analysis each
857 // overflow variable is set to 0 if it's corresponding bound variable is valid
858 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
859 int LoOverflow = 0, HiOverflow = 0;
860 Constant *LoBound = 0, *HiBound = 0;
862 if (!DivIsSigned) { // udiv
863 // e.g. X/5 op 3 --> [15, 20)
865 HiOverflow = LoOverflow = ProdOV;
867 // If this is not an exact divide, then many values in the range collapse
868 // to the same result value.
869 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
872 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
873 if (CmpRHSV == 0) { // (X / pos) op 0
874 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
875 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
877 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
878 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
879 HiOverflow = LoOverflow = ProdOV;
881 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
882 } else { // (X / pos) op neg
883 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
884 HiBound = AddOne(Prod);
885 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
887 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
888 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
891 } else if (DivRHS->isNegative()) { // Divisor is < 0.
893 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
894 if (CmpRHSV == 0) { // (X / neg) op 0
895 // e.g. X/-5 op 0 --> [-4, 5)
896 LoBound = AddOne(RangeSize);
897 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
898 if (HiBound == DivRHS) { // -INTMIN = INTMIN
899 HiOverflow = 1; // [INTMIN+1, overflow)
900 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
902 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
903 // e.g. X/-5 op 3 --> [-19, -14)
904 HiBound = AddOne(Prod);
905 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
907 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
908 } else { // (X / neg) op neg
909 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
910 LoOverflow = HiOverflow = ProdOV;
912 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
915 // Dividing by a negative swaps the condition. LT <-> GT
916 Pred = ICmpInst::getSwappedPredicate(Pred);
919 Value *X = DivI->getOperand(0);
921 default: llvm_unreachable("Unhandled icmp opcode!");
922 case ICmpInst::ICMP_EQ:
923 if (LoOverflow && HiOverflow)
924 return ReplaceInstUsesWith(ICI, Builder->getFalse());
926 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
927 ICmpInst::ICMP_UGE, X, LoBound);
929 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
930 ICmpInst::ICMP_ULT, X, HiBound);
931 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
933 case ICmpInst::ICMP_NE:
934 if (LoOverflow && HiOverflow)
935 return ReplaceInstUsesWith(ICI, Builder->getTrue());
937 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
938 ICmpInst::ICMP_ULT, X, LoBound);
940 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
941 ICmpInst::ICMP_UGE, X, HiBound);
942 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
943 DivIsSigned, false));
944 case ICmpInst::ICMP_ULT:
945 case ICmpInst::ICMP_SLT:
946 if (LoOverflow == +1) // Low bound is greater than input range.
947 return ReplaceInstUsesWith(ICI, Builder->getTrue());
948 if (LoOverflow == -1) // Low bound is less than input range.
949 return ReplaceInstUsesWith(ICI, Builder->getFalse());
950 return new ICmpInst(Pred, X, LoBound);
951 case ICmpInst::ICMP_UGT:
952 case ICmpInst::ICMP_SGT:
953 if (HiOverflow == +1) // High bound greater than input range.
954 return ReplaceInstUsesWith(ICI, Builder->getFalse());
955 if (HiOverflow == -1) // High bound less than input range.
956 return ReplaceInstUsesWith(ICI, Builder->getTrue());
957 if (Pred == ICmpInst::ICMP_UGT)
958 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
959 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
963 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
964 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
965 ConstantInt *ShAmt) {
966 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
968 // Check that the shift amount is in range. If not, don't perform
969 // undefined shifts. When the shift is visited it will be
971 uint32_t TypeBits = CmpRHSV.getBitWidth();
972 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
973 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
976 if (!ICI.isEquality()) {
977 // If we have an unsigned comparison and an ashr, we can't simplify this.
978 // Similarly for signed comparisons with lshr.
979 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
982 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
983 // by a power of 2. Since we already have logic to simplify these,
984 // transform to div and then simplify the resultant comparison.
985 if (Shr->getOpcode() == Instruction::AShr &&
986 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
989 // Revisit the shift (to delete it).
993 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
996 Shr->getOpcode() == Instruction::AShr ?
997 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
998 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
1000 ICI.setOperand(0, Tmp);
1002 // If the builder folded the binop, just return it.
1003 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
1007 // Otherwise, fold this div/compare.
1008 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1009 TheDiv->getOpcode() == Instruction::UDiv);
1011 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1012 assert(Res && "This div/cst should have folded!");
1017 // If we are comparing against bits always shifted out, the
1018 // comparison cannot succeed.
1019 APInt Comp = CmpRHSV << ShAmtVal;
1020 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1021 if (Shr->getOpcode() == Instruction::LShr)
1022 Comp = Comp.lshr(ShAmtVal);
1024 Comp = Comp.ashr(ShAmtVal);
1026 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1027 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1028 Constant *Cst = Builder->getInt1(IsICMP_NE);
1029 return ReplaceInstUsesWith(ICI, Cst);
1032 // Otherwise, check to see if the bits shifted out are known to be zero.
1033 // If so, we can compare against the unshifted value:
1034 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1035 if (Shr->hasOneUse() && Shr->isExact())
1036 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1038 if (Shr->hasOneUse()) {
1039 // Otherwise strength reduce the shift into an and.
1040 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1041 Constant *Mask = Builder->getInt(Val);
1043 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1044 Mask, Shr->getName()+".mask");
1045 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1051 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1053 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1056 const APInt &RHSV = RHS->getValue();
1058 switch (LHSI->getOpcode()) {
1059 case Instruction::Trunc:
1060 if (ICI.isEquality() && LHSI->hasOneUse()) {
1061 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1062 // of the high bits truncated out of x are known.
1063 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1064 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1065 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1066 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1068 // If all the high bits are known, we can do this xform.
1069 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1070 // Pull in the high bits from known-ones set.
1071 APInt NewRHS = RHS->getValue().zext(SrcBits);
1072 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1073 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1074 Builder->getInt(NewRHS));
1079 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1080 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1081 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1083 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1084 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1085 Value *CompareVal = LHSI->getOperand(0);
1087 // If the sign bit of the XorCST is not set, there is no change to
1088 // the operation, just stop using the Xor.
1089 if (!XorCST->isNegative()) {
1090 ICI.setOperand(0, CompareVal);
1095 // Was the old condition true if the operand is positive?
1096 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1098 // If so, the new one isn't.
1099 isTrueIfPositive ^= true;
1101 if (isTrueIfPositive)
1102 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1105 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1109 if (LHSI->hasOneUse()) {
1110 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1111 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1112 const APInt &SignBit = XorCST->getValue();
1113 ICmpInst::Predicate Pred = ICI.isSigned()
1114 ? ICI.getUnsignedPredicate()
1115 : ICI.getSignedPredicate();
1116 return new ICmpInst(Pred, LHSI->getOperand(0),
1117 Builder->getInt(RHSV ^ SignBit));
1120 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1121 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1122 const APInt &NotSignBit = XorCST->getValue();
1123 ICmpInst::Predicate Pred = ICI.isSigned()
1124 ? ICI.getUnsignedPredicate()
1125 : ICI.getSignedPredicate();
1126 Pred = ICI.getSwappedPredicate(Pred);
1127 return new ICmpInst(Pred, LHSI->getOperand(0),
1128 Builder->getInt(RHSV ^ NotSignBit));
1132 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1133 // iff -C is a power of 2
1134 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1135 XorCST->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1136 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCST);
1138 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1139 // iff -C is a power of 2
1140 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1141 XorCST->getValue() == -RHSV && RHSV.isPowerOf2())
1142 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCST);
1145 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1146 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1147 LHSI->getOperand(0)->hasOneUse()) {
1148 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1150 // If the LHS is an AND of a truncating cast, we can widen the
1151 // and/compare to be the input width without changing the value
1152 // produced, eliminating a cast.
1153 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1154 // We can do this transformation if either the AND constant does not
1155 // have its sign bit set or if it is an equality comparison.
1156 // Extending a relational comparison when we're checking the sign
1157 // bit would not work.
1158 if (ICI.isEquality() ||
1159 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1161 Builder->CreateAnd(Cast->getOperand(0),
1162 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1163 NewAnd->takeName(LHSI);
1164 return new ICmpInst(ICI.getPredicate(), NewAnd,
1165 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1169 // If the LHS is an AND of a zext, and we have an equality compare, we can
1170 // shrink the and/compare to the smaller type, eliminating the cast.
1171 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1172 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1173 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1174 // should fold the icmp to true/false in that case.
1175 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1177 Builder->CreateAnd(Cast->getOperand(0),
1178 ConstantExpr::getTrunc(AndCST, Ty));
1179 NewAnd->takeName(LHSI);
1180 return new ICmpInst(ICI.getPredicate(), NewAnd,
1181 ConstantExpr::getTrunc(RHS, Ty));
1185 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1186 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1187 // happens a LOT in code produced by the C front-end, for bitfield
1189 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1190 if (Shift && !Shift->isShift())
1194 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1195 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1196 Type *AndTy = AndCST->getType(); // Type of the and.
1198 // We can fold this as long as we can't shift unknown bits
1199 // into the mask. This can only happen with signed shift
1200 // rights, as they sign-extend.
1202 bool CanFold = Shift->isLogicalShift();
1204 // To test for the bad case of the signed shr, see if any
1205 // of the bits shifted in could be tested after the mask.
1206 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1207 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1209 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1210 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1211 AndCST->getValue()) == 0)
1217 if (Shift->getOpcode() == Instruction::Shl)
1218 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1220 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1222 // Check to see if we are shifting out any of the bits being
1224 if (ConstantExpr::get(Shift->getOpcode(),
1225 NewCst, ShAmt) != RHS) {
1226 // If we shifted bits out, the fold is not going to work out.
1227 // As a special case, check to see if this means that the
1228 // result is always true or false now.
1229 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1230 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1231 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1232 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1234 ICI.setOperand(1, NewCst);
1235 Constant *NewAndCST;
1236 if (Shift->getOpcode() == Instruction::Shl)
1237 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1239 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1240 LHSI->setOperand(1, NewAndCST);
1241 LHSI->setOperand(0, Shift->getOperand(0));
1242 Worklist.Add(Shift); // Shift is dead.
1248 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1249 // preferable because it allows the C<<Y expression to be hoisted out
1250 // of a loop if Y is invariant and X is not.
1251 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1252 ICI.isEquality() && !Shift->isArithmeticShift() &&
1253 !isa<Constant>(Shift->getOperand(0))) {
1256 if (Shift->getOpcode() == Instruction::LShr) {
1257 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1259 // Insert a logical shift.
1260 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1263 // Compute X & (C << Y).
1265 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1267 ICI.setOperand(0, NewAnd);
1271 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
1272 // bit set in (X & AndCST) will produce a result greater than RHSV.
1273 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1274 unsigned NTZ = AndCST->getValue().countTrailingZeros();
1275 if ((NTZ < AndCST->getBitWidth()) &&
1276 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
1277 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1278 Constant::getNullValue(RHS->getType()));
1282 // Try to optimize things like "A[i]&42 == 0" to index computations.
1283 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1284 if (GetElementPtrInst *GEP =
1285 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1286 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1287 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1288 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1289 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1290 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1295 // X & -C == -C -> X > u ~C
1296 // X & -C != -C -> X <= u ~C
1297 // iff C is a power of 2
1298 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1299 return new ICmpInst(
1300 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1301 : ICmpInst::ICMP_ULE,
1302 LHSI->getOperand(0), SubOne(RHS));
1305 case Instruction::Or: {
1306 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1309 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1310 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1311 // -> and (icmp eq P, null), (icmp eq Q, null).
1312 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1313 Constant::getNullValue(P->getType()));
1314 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1315 Constant::getNullValue(Q->getType()));
1317 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1318 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1320 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1326 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1327 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1330 // If this is a signed comparison to 0 and the mul is sign preserving,
1331 // use the mul LHS operand instead.
1332 ICmpInst::Predicate pred = ICI.getPredicate();
1333 if (isSignTest(pred, RHS) && !Val->isZero() &&
1334 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1335 return new ICmpInst(Val->isNegative() ?
1336 ICmpInst::getSwappedPredicate(pred) : pred,
1337 LHSI->getOperand(0),
1338 Constant::getNullValue(RHS->getType()));
1343 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1344 uint32_t TypeBits = RHSV.getBitWidth();
1345 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1348 // (1 << X) pred P2 -> X pred Log2(P2)
1349 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1350 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1351 ICmpInst::Predicate Pred = ICI.getPredicate();
1352 if (ICI.isUnsigned()) {
1353 if (!RHSVIsPowerOf2) {
1354 // (1 << X) < 30 -> X <= 4
1355 // (1 << X) <= 30 -> X <= 4
1356 // (1 << X) >= 30 -> X > 4
1357 // (1 << X) > 30 -> X > 4
1358 if (Pred == ICmpInst::ICMP_ULT)
1359 Pred = ICmpInst::ICMP_ULE;
1360 else if (Pred == ICmpInst::ICMP_UGE)
1361 Pred = ICmpInst::ICMP_UGT;
1363 unsigned RHSLog2 = RHSV.logBase2();
1365 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1366 // (1 << X) > 2147483648 -> X > 31 -> false
1367 // (1 << X) <= 2147483648 -> X <= 31 -> true
1368 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1369 if (RHSLog2 == TypeBits-1) {
1370 if (Pred == ICmpInst::ICMP_UGE)
1371 Pred = ICmpInst::ICMP_EQ;
1372 else if (Pred == ICmpInst::ICMP_UGT)
1373 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1374 else if (Pred == ICmpInst::ICMP_ULE)
1375 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1376 else if (Pred == ICmpInst::ICMP_ULT)
1377 Pred = ICmpInst::ICMP_NE;
1380 return new ICmpInst(Pred, X,
1381 ConstantInt::get(RHS->getType(), RHSLog2));
1382 } else if (ICI.isSigned()) {
1383 if (RHSV.isAllOnesValue()) {
1384 // (1 << X) <= -1 -> X == 31
1385 if (Pred == ICmpInst::ICMP_SLE)
1386 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1387 ConstantInt::get(RHS->getType(), TypeBits-1));
1389 // (1 << X) > -1 -> X != 31
1390 if (Pred == ICmpInst::ICMP_SGT)
1391 return new ICmpInst(ICmpInst::ICMP_NE, X,
1392 ConstantInt::get(RHS->getType(), TypeBits-1));
1394 // (1 << X) < 0 -> X == 31
1395 // (1 << X) <= 0 -> X == 31
1396 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1397 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1398 ConstantInt::get(RHS->getType(), TypeBits-1));
1400 // (1 << X) >= 0 -> X != 31
1401 // (1 << X) > 0 -> X != 31
1402 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1403 return new ICmpInst(ICmpInst::ICMP_NE, X,
1404 ConstantInt::get(RHS->getType(), TypeBits-1));
1406 } else if (ICI.isEquality()) {
1408 return new ICmpInst(
1409 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1411 return ReplaceInstUsesWith(
1412 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse()
1413 : Builder->getTrue());
1419 // Check that the shift amount is in range. If not, don't perform
1420 // undefined shifts. When the shift is visited it will be
1422 if (ShAmt->uge(TypeBits))
1425 if (ICI.isEquality()) {
1426 // If we are comparing against bits always shifted out, the
1427 // comparison cannot succeed.
1429 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1431 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1432 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1433 Constant *Cst = Builder->getInt1(IsICMP_NE);
1434 return ReplaceInstUsesWith(ICI, Cst);
1437 // If the shift is NUW, then it is just shifting out zeros, no need for an
1439 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1440 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1441 ConstantExpr::getLShr(RHS, ShAmt));
1443 // If the shift is NSW and we compare to 0, then it is just shifting out
1444 // sign bits, no need for an AND either.
1445 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1446 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1447 ConstantExpr::getLShr(RHS, ShAmt));
1449 if (LHSI->hasOneUse()) {
1450 // Otherwise strength reduce the shift into an and.
1451 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1452 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1453 TypeBits - ShAmtVal));
1456 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1457 return new ICmpInst(ICI.getPredicate(), And,
1458 ConstantExpr::getLShr(RHS, ShAmt));
1462 // If this is a signed comparison to 0 and the shift is sign preserving,
1463 // use the shift LHS operand instead.
1464 ICmpInst::Predicate pred = ICI.getPredicate();
1465 if (isSignTest(pred, RHS) &&
1466 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1467 return new ICmpInst(pred,
1468 LHSI->getOperand(0),
1469 Constant::getNullValue(RHS->getType()));
1471 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1472 bool TrueIfSigned = false;
1473 if (LHSI->hasOneUse() &&
1474 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1475 // (X << 31) <s 0 --> (X&1) != 0
1476 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1477 APInt::getOneBitSet(TypeBits,
1478 TypeBits-ShAmt->getZExtValue()-1));
1480 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1481 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1482 And, Constant::getNullValue(And->getType()));
1485 // Transform (icmp pred iM (shl iM %v, N), CI)
1486 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1487 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1488 // This enables to get rid of the shift in favor of a trunc which can be
1489 // free on the target. It has the additional benefit of comparing to a
1490 // smaller constant, which will be target friendly.
1491 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1492 if (LHSI->hasOneUse() &&
1493 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1494 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1495 Constant *NCI = ConstantExpr::getTrunc(
1496 ConstantExpr::getAShr(RHS,
1497 ConstantInt::get(RHS->getType(), Amt)),
1499 return new ICmpInst(ICI.getPredicate(),
1500 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1507 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1508 case Instruction::AShr: {
1509 // Handle equality comparisons of shift-by-constant.
1510 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1511 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1512 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1516 // Handle exact shr's.
1517 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1518 if (RHSV.isMinValue())
1519 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1524 case Instruction::SDiv:
1525 case Instruction::UDiv:
1526 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1527 // Fold this div into the comparison, producing a range check.
1528 // Determine, based on the divide type, what the range is being
1529 // checked. If there is an overflow on the low or high side, remember
1530 // it, otherwise compute the range [low, hi) bounding the new value.
1531 // See: InsertRangeTest above for the kinds of replacements possible.
1532 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1533 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1538 case Instruction::Sub: {
1539 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1541 const APInt &LHSV = LHSC->getValue();
1543 // C1-X <u C2 -> (X|(C2-1)) == C1
1544 // iff C1 & (C2-1) == C2-1
1545 // C2 is a power of 2
1546 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1547 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1548 return new ICmpInst(ICmpInst::ICMP_EQ,
1549 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1552 // C1-X >u C2 -> (X|C2) != C1
1553 // iff C1 & C2 == C2
1554 // C2+1 is a power of 2
1555 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1556 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1557 return new ICmpInst(ICmpInst::ICMP_NE,
1558 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1562 case Instruction::Add:
1563 // Fold: icmp pred (add X, C1), C2
1564 if (!ICI.isEquality()) {
1565 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1567 const APInt &LHSV = LHSC->getValue();
1569 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1572 if (ICI.isSigned()) {
1573 if (CR.getLower().isSignBit()) {
1574 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1575 Builder->getInt(CR.getUpper()));
1576 } else if (CR.getUpper().isSignBit()) {
1577 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1578 Builder->getInt(CR.getLower()));
1581 if (CR.getLower().isMinValue()) {
1582 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1583 Builder->getInt(CR.getUpper()));
1584 } else if (CR.getUpper().isMinValue()) {
1585 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1586 Builder->getInt(CR.getLower()));
1590 // X-C1 <u C2 -> (X & -C2) == C1
1591 // iff C1 & (C2-1) == 0
1592 // C2 is a power of 2
1593 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1594 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1595 return new ICmpInst(ICmpInst::ICMP_EQ,
1596 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1597 ConstantExpr::getNeg(LHSC));
1599 // X-C1 >u C2 -> (X & ~C2) != C1
1601 // C2+1 is a power of 2
1602 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1603 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1604 return new ICmpInst(ICmpInst::ICMP_NE,
1605 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1606 ConstantExpr::getNeg(LHSC));
1611 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1612 if (ICI.isEquality()) {
1613 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1615 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1616 // the second operand is a constant, simplify a bit.
1617 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1618 switch (BO->getOpcode()) {
1619 case Instruction::SRem:
1620 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1621 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1622 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1623 if (V.sgt(1) && V.isPowerOf2()) {
1625 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1627 return new ICmpInst(ICI.getPredicate(), NewRem,
1628 Constant::getNullValue(BO->getType()));
1632 case Instruction::Add:
1633 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1634 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1635 if (BO->hasOneUse())
1636 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1637 ConstantExpr::getSub(RHS, BOp1C));
1638 } else if (RHSV == 0) {
1639 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1640 // efficiently invertible, or if the add has just this one use.
1641 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1643 if (Value *NegVal = dyn_castNegVal(BOp1))
1644 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1645 if (Value *NegVal = dyn_castNegVal(BOp0))
1646 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1647 if (BO->hasOneUse()) {
1648 Value *Neg = Builder->CreateNeg(BOp1);
1650 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1654 case Instruction::Xor:
1655 // For the xor case, we can xor two constants together, eliminating
1656 // the explicit xor.
1657 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1658 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1659 ConstantExpr::getXor(RHS, BOC));
1660 } else if (RHSV == 0) {
1661 // Replace ((xor A, B) != 0) with (A != B)
1662 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1666 case Instruction::Sub:
1667 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1668 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1669 if (BO->hasOneUse())
1670 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1671 ConstantExpr::getSub(BOp0C, RHS));
1672 } else if (RHSV == 0) {
1673 // Replace ((sub A, B) != 0) with (A != B)
1674 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1678 case Instruction::Or:
1679 // If bits are being or'd in that are not present in the constant we
1680 // are comparing against, then the comparison could never succeed!
1681 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1682 Constant *NotCI = ConstantExpr::getNot(RHS);
1683 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1684 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1688 case Instruction::And:
1689 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1690 // If bits are being compared against that are and'd out, then the
1691 // comparison can never succeed!
1692 if ((RHSV & ~BOC->getValue()) != 0)
1693 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1695 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1696 if (RHS == BOC && RHSV.isPowerOf2())
1697 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1698 ICmpInst::ICMP_NE, LHSI,
1699 Constant::getNullValue(RHS->getType()));
1701 // Don't perform the following transforms if the AND has multiple uses
1702 if (!BO->hasOneUse())
1705 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1706 if (BOC->getValue().isSignBit()) {
1707 Value *X = BO->getOperand(0);
1708 Constant *Zero = Constant::getNullValue(X->getType());
1709 ICmpInst::Predicate pred = isICMP_NE ?
1710 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1711 return new ICmpInst(pred, X, Zero);
1714 // ((X & ~7) == 0) --> X < 8
1715 if (RHSV == 0 && isHighOnes(BOC)) {
1716 Value *X = BO->getOperand(0);
1717 Constant *NegX = ConstantExpr::getNeg(BOC);
1718 ICmpInst::Predicate pred = isICMP_NE ?
1719 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1720 return new ICmpInst(pred, X, NegX);
1724 case Instruction::Mul:
1725 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1726 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1727 // The trivial case (mul X, 0) is handled by InstSimplify
1728 // General case : (mul X, C) != 0 iff X != 0
1729 // (mul X, C) == 0 iff X == 0
1731 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1732 Constant::getNullValue(RHS->getType()));
1738 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1739 // Handle icmp {eq|ne} <intrinsic>, intcst.
1740 switch (II->getIntrinsicID()) {
1741 case Intrinsic::bswap:
1743 ICI.setOperand(0, II->getArgOperand(0));
1744 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1746 case Intrinsic::ctlz:
1747 case Intrinsic::cttz:
1748 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1749 if (RHSV == RHS->getType()->getBitWidth()) {
1751 ICI.setOperand(0, II->getArgOperand(0));
1752 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1756 case Intrinsic::ctpop:
1757 // popcount(A) == 0 -> A == 0 and likewise for !=
1758 if (RHS->isZero()) {
1760 ICI.setOperand(0, II->getArgOperand(0));
1761 ICI.setOperand(1, RHS);
1773 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1774 /// We only handle extending casts so far.
1776 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1777 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1778 Value *LHSCIOp = LHSCI->getOperand(0);
1779 Type *SrcTy = LHSCIOp->getType();
1780 Type *DestTy = LHSCI->getType();
1783 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1784 // integer type is the same size as the pointer type.
1785 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1786 TD->getPointerSizeInBits() ==
1787 cast<IntegerType>(DestTy)->getBitWidth()) {
1789 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1790 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1791 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1792 RHSOp = RHSC->getOperand(0);
1793 // If the pointer types don't match, insert a bitcast.
1794 if (LHSCIOp->getType() != RHSOp->getType())
1795 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1799 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1802 // The code below only handles extension cast instructions, so far.
1804 if (LHSCI->getOpcode() != Instruction::ZExt &&
1805 LHSCI->getOpcode() != Instruction::SExt)
1808 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1809 bool isSignedCmp = ICI.isSigned();
1811 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1812 // Not an extension from the same type?
1813 RHSCIOp = CI->getOperand(0);
1814 if (RHSCIOp->getType() != LHSCIOp->getType())
1817 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1818 // and the other is a zext), then we can't handle this.
1819 if (CI->getOpcode() != LHSCI->getOpcode())
1822 // Deal with equality cases early.
1823 if (ICI.isEquality())
1824 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1826 // A signed comparison of sign extended values simplifies into a
1827 // signed comparison.
1828 if (isSignedCmp && isSignedExt)
1829 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1831 // The other three cases all fold into an unsigned comparison.
1832 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1835 // If we aren't dealing with a constant on the RHS, exit early
1836 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1840 // Compute the constant that would happen if we truncated to SrcTy then
1841 // reextended to DestTy.
1842 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1843 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1846 // If the re-extended constant didn't change...
1848 // Deal with equality cases early.
1849 if (ICI.isEquality())
1850 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1852 // A signed comparison of sign extended values simplifies into a
1853 // signed comparison.
1854 if (isSignedExt && isSignedCmp)
1855 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1857 // The other three cases all fold into an unsigned comparison.
1858 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1861 // The re-extended constant changed so the constant cannot be represented
1862 // in the shorter type. Consequently, we cannot emit a simple comparison.
1863 // All the cases that fold to true or false will have already been handled
1864 // by SimplifyICmpInst, so only deal with the tricky case.
1866 if (isSignedCmp || !isSignedExt)
1869 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1870 // should have been folded away previously and not enter in here.
1872 // We're performing an unsigned comp with a sign extended value.
1873 // This is true if the input is >= 0. [aka >s -1]
1874 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1875 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1877 // Finally, return the value computed.
1878 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1879 return ReplaceInstUsesWith(ICI, Result);
1881 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1882 return BinaryOperator::CreateNot(Result);
1885 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1886 /// I = icmp ugt (add (add A, B), CI2), CI1
1887 /// If this is of the form:
1889 /// if (sum+128 >u 255)
1890 /// Then replace it with llvm.sadd.with.overflow.i8.
1892 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1893 ConstantInt *CI2, ConstantInt *CI1,
1895 // The transformation we're trying to do here is to transform this into an
1896 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1897 // with a narrower add, and discard the add-with-constant that is part of the
1898 // range check (if we can't eliminate it, this isn't profitable).
1900 // In order to eliminate the add-with-constant, the compare can be its only
1902 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1903 if (!AddWithCst->hasOneUse()) return 0;
1905 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1906 if (!CI2->getValue().isPowerOf2()) return 0;
1907 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1908 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1910 // The width of the new add formed is 1 more than the bias.
1913 // Check to see that CI1 is an all-ones value with NewWidth bits.
1914 if (CI1->getBitWidth() == NewWidth ||
1915 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1918 // This is only really a signed overflow check if the inputs have been
1919 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1920 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1921 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1922 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1923 IC.ComputeNumSignBits(B) < NeededSignBits)
1926 // In order to replace the original add with a narrower
1927 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1928 // and truncates that discard the high bits of the add. Verify that this is
1930 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1931 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1933 if (*UI == AddWithCst) continue;
1935 // Only accept truncates for now. We would really like a nice recursive
1936 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1937 // chain to see which bits of a value are actually demanded. If the
1938 // original add had another add which was then immediately truncated, we
1939 // could still do the transformation.
1940 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1942 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1945 // If the pattern matches, truncate the inputs to the narrower type and
1946 // use the sadd_with_overflow intrinsic to efficiently compute both the
1947 // result and the overflow bit.
1948 Module *M = I.getParent()->getParent()->getParent();
1950 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1951 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1954 InstCombiner::BuilderTy *Builder = IC.Builder;
1956 // Put the new code above the original add, in case there are any uses of the
1957 // add between the add and the compare.
1958 Builder->SetInsertPoint(OrigAdd);
1960 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1961 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1962 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1963 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1964 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1966 // The inner add was the result of the narrow add, zero extended to the
1967 // wider type. Replace it with the result computed by the intrinsic.
1968 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1970 // The original icmp gets replaced with the overflow value.
1971 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1974 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1976 // Don't bother doing this transformation for pointers, don't do it for
1978 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1980 // If the add is a constant expr, then we don't bother transforming it.
1981 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1982 if (OrigAdd == 0) return 0;
1984 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1986 // Put the new code above the original add, in case there are any uses of the
1987 // add between the add and the compare.
1988 InstCombiner::BuilderTy *Builder = IC.Builder;
1989 Builder->SetInsertPoint(OrigAdd);
1991 Module *M = I.getParent()->getParent()->getParent();
1992 Type *Ty = LHS->getType();
1993 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1994 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1995 Value *Add = Builder->CreateExtractValue(Call, 0);
1997 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1999 // The original icmp gets replaced with the overflow value.
2000 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2003 // DemandedBitsLHSMask - When performing a comparison against a constant,
2004 // it is possible that not all the bits in the LHS are demanded. This helper
2005 // method computes the mask that IS demanded.
2006 static APInt DemandedBitsLHSMask(ICmpInst &I,
2007 unsigned BitWidth, bool isSignCheck) {
2009 return APInt::getSignBit(BitWidth);
2011 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2012 if (!CI) return APInt::getAllOnesValue(BitWidth);
2013 const APInt &RHS = CI->getValue();
2015 switch (I.getPredicate()) {
2016 // For a UGT comparison, we don't care about any bits that
2017 // correspond to the trailing ones of the comparand. The value of these
2018 // bits doesn't impact the outcome of the comparison, because any value
2019 // greater than the RHS must differ in a bit higher than these due to carry.
2020 case ICmpInst::ICMP_UGT: {
2021 unsigned trailingOnes = RHS.countTrailingOnes();
2022 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2026 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2027 // Any value less than the RHS must differ in a higher bit because of carries.
2028 case ICmpInst::ICMP_ULT: {
2029 unsigned trailingZeros = RHS.countTrailingZeros();
2030 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2035 return APInt::getAllOnesValue(BitWidth);
2040 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2041 /// should be swapped.
2042 /// The descision is based on how many times these two operands are reused
2043 /// as subtract operands and their positions in those instructions.
2044 /// The rational is that several architectures use the same instruction for
2045 /// both subtract and cmp, thus it is better if the order of those operands
2047 /// \return true if Op0 and Op1 should be swapped.
2048 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2049 const Value * Op1) {
2050 // Filter out pointer value as those cannot appears directly in subtract.
2051 // FIXME: we may want to go through inttoptrs or bitcasts.
2052 if (Op0->getType()->isPointerTy())
2054 // Count every uses of both Op0 and Op1 in a subtract.
2055 // Each time Op0 is the first operand, count -1: swapping is bad, the
2056 // subtract has already the same layout as the compare.
2057 // Each time Op0 is the second operand, count +1: swapping is good, the
2058 // subtract has a diffrent layout as the compare.
2059 // At the end, if the benefit is greater than 0, Op0 should come second to
2060 // expose more CSE opportunities.
2061 int GlobalSwapBenefits = 0;
2062 for (Value::const_use_iterator UI = Op0->use_begin(), UIEnd = Op0->use_end(); UI != UIEnd; ++UI) {
2063 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(*UI);
2064 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2066 // If Op0 is the first argument, this is not beneficial to swap the
2068 int LocalSwapBenefits = -1;
2069 unsigned Op1Idx = 1;
2070 if (BinOp->getOperand(Op1Idx) == Op0) {
2072 LocalSwapBenefits = 1;
2074 if (BinOp->getOperand(Op1Idx) != Op1)
2076 GlobalSwapBenefits += LocalSwapBenefits;
2078 return GlobalSwapBenefits > 0;
2081 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2082 bool Changed = false;
2083 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2084 unsigned Op0Cplxity = getComplexity(Op0);
2085 unsigned Op1Cplxity = getComplexity(Op1);
2087 /// Orders the operands of the compare so that they are listed from most
2088 /// complex to least complex. This puts constants before unary operators,
2089 /// before binary operators.
2090 if (Op0Cplxity < Op1Cplxity ||
2091 (Op0Cplxity == Op1Cplxity &&
2092 swapMayExposeCSEOpportunities(Op0, Op1))) {
2094 std::swap(Op0, Op1);
2098 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
2099 return ReplaceInstUsesWith(I, V);
2101 // comparing -val or val with non-zero is the same as just comparing val
2102 // ie, abs(val) != 0 -> val != 0
2103 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2105 Value *Cond, *SelectTrue, *SelectFalse;
2106 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2107 m_Value(SelectFalse)))) {
2108 if (Value *V = dyn_castNegVal(SelectTrue)) {
2109 if (V == SelectFalse)
2110 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2112 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2113 if (V == SelectTrue)
2114 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2119 Type *Ty = Op0->getType();
2121 // icmp's with boolean values can always be turned into bitwise operations
2122 if (Ty->isIntegerTy(1)) {
2123 switch (I.getPredicate()) {
2124 default: llvm_unreachable("Invalid icmp instruction!");
2125 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2126 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2127 return BinaryOperator::CreateNot(Xor);
2129 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2130 return BinaryOperator::CreateXor(Op0, Op1);
2132 case ICmpInst::ICMP_UGT:
2133 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2135 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2136 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2137 return BinaryOperator::CreateAnd(Not, Op1);
2139 case ICmpInst::ICMP_SGT:
2140 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2142 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2143 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2144 return BinaryOperator::CreateAnd(Not, Op0);
2146 case ICmpInst::ICMP_UGE:
2147 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2149 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2150 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2151 return BinaryOperator::CreateOr(Not, Op1);
2153 case ICmpInst::ICMP_SGE:
2154 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2156 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2157 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2158 return BinaryOperator::CreateOr(Not, Op0);
2163 unsigned BitWidth = 0;
2164 if (Ty->isIntOrIntVectorTy())
2165 BitWidth = Ty->getScalarSizeInBits();
2166 else if (TD) // Pointers require TD info to get their size.
2167 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
2169 bool isSignBit = false;
2171 // See if we are doing a comparison with a constant.
2172 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2173 Value *A = 0, *B = 0;
2175 // Match the following pattern, which is a common idiom when writing
2176 // overflow-safe integer arithmetic function. The source performs an
2177 // addition in wider type, and explicitly checks for overflow using
2178 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2179 // sadd_with_overflow intrinsic.
2181 // TODO: This could probably be generalized to handle other overflow-safe
2182 // operations if we worked out the formulas to compute the appropriate
2186 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2188 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2189 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2190 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2191 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2195 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2196 if (I.isEquality() && CI->isZero() &&
2197 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2198 // (icmp cond A B) if cond is equality
2199 return new ICmpInst(I.getPredicate(), A, B);
2202 // If we have an icmp le or icmp ge instruction, turn it into the
2203 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2204 // them being folded in the code below. The SimplifyICmpInst code has
2205 // already handled the edge cases for us, so we just assert on them.
2206 switch (I.getPredicate()) {
2208 case ICmpInst::ICMP_ULE:
2209 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2210 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2211 Builder->getInt(CI->getValue()+1));
2212 case ICmpInst::ICMP_SLE:
2213 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2214 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2215 Builder->getInt(CI->getValue()+1));
2216 case ICmpInst::ICMP_UGE:
2217 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2218 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2219 Builder->getInt(CI->getValue()-1));
2220 case ICmpInst::ICMP_SGE:
2221 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2222 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2223 Builder->getInt(CI->getValue()-1));
2226 // If this comparison is a normal comparison, it demands all
2227 // bits, if it is a sign bit comparison, it only demands the sign bit.
2229 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2232 // See if we can fold the comparison based on range information we can get
2233 // by checking whether bits are known to be zero or one in the input.
2234 if (BitWidth != 0) {
2235 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2236 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2238 if (SimplifyDemandedBits(I.getOperandUse(0),
2239 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2240 Op0KnownZero, Op0KnownOne, 0))
2242 if (SimplifyDemandedBits(I.getOperandUse(1),
2243 APInt::getAllOnesValue(BitWidth),
2244 Op1KnownZero, Op1KnownOne, 0))
2247 // Given the known and unknown bits, compute a range that the LHS could be
2248 // in. Compute the Min, Max and RHS values based on the known bits. For the
2249 // EQ and NE we use unsigned values.
2250 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2251 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2253 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2255 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2258 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2260 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2264 // If Min and Max are known to be the same, then SimplifyDemandedBits
2265 // figured out that the LHS is a constant. Just constant fold this now so
2266 // that code below can assume that Min != Max.
2267 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2268 return new ICmpInst(I.getPredicate(),
2269 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2270 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2271 return new ICmpInst(I.getPredicate(), Op0,
2272 ConstantInt::get(Op1->getType(), Op1Min));
2274 // Based on the range information we know about the LHS, see if we can
2275 // simplify this comparison. For example, (x&4) < 8 is always true.
2276 switch (I.getPredicate()) {
2277 default: llvm_unreachable("Unknown icmp opcode!");
2278 case ICmpInst::ICMP_EQ: {
2279 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2280 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2282 // If all bits are known zero except for one, then we know at most one
2283 // bit is set. If the comparison is against zero, then this is a check
2284 // to see if *that* bit is set.
2285 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2286 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2287 // If the LHS is an AND with the same constant, look through it.
2289 ConstantInt *LHSC = 0;
2290 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2291 LHSC->getValue() != Op0KnownZeroInverted)
2294 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2295 // then turn "((1 << x)&8) == 0" into "x != 3".
2297 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2298 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2299 return new ICmpInst(ICmpInst::ICMP_NE, X,
2300 ConstantInt::get(X->getType(), CmpVal));
2303 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2304 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2306 if (Op0KnownZeroInverted == 1 &&
2307 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2308 return new ICmpInst(ICmpInst::ICMP_NE, X,
2309 ConstantInt::get(X->getType(),
2310 CI->countTrailingZeros()));
2315 case ICmpInst::ICMP_NE: {
2316 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2317 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2319 // If all bits are known zero except for one, then we know at most one
2320 // bit is set. If the comparison is against zero, then this is a check
2321 // to see if *that* bit is set.
2322 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2323 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2324 // If the LHS is an AND with the same constant, look through it.
2326 ConstantInt *LHSC = 0;
2327 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2328 LHSC->getValue() != Op0KnownZeroInverted)
2331 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2332 // then turn "((1 << x)&8) != 0" into "x == 3".
2334 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2335 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2336 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2337 ConstantInt::get(X->getType(), CmpVal));
2340 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2341 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2343 if (Op0KnownZeroInverted == 1 &&
2344 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2345 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2346 ConstantInt::get(X->getType(),
2347 CI->countTrailingZeros()));
2352 case ICmpInst::ICMP_ULT:
2353 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2354 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2355 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2356 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2357 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2358 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2359 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2360 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2361 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2362 Builder->getInt(CI->getValue()-1));
2364 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2365 if (CI->isMinValue(true))
2366 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2367 Constant::getAllOnesValue(Op0->getType()));
2370 case ICmpInst::ICMP_UGT:
2371 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2372 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2373 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2374 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2376 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2377 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2378 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2379 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2380 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2381 Builder->getInt(CI->getValue()+1));
2383 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2384 if (CI->isMaxValue(true))
2385 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2386 Constant::getNullValue(Op0->getType()));
2389 case ICmpInst::ICMP_SLT:
2390 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2391 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2392 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2393 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2394 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2395 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2396 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2397 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2398 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2399 Builder->getInt(CI->getValue()-1));
2402 case ICmpInst::ICMP_SGT:
2403 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2404 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2405 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2406 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2408 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2409 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2410 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2411 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2412 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2413 Builder->getInt(CI->getValue()+1));
2416 case ICmpInst::ICMP_SGE:
2417 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2418 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2419 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2420 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2421 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2423 case ICmpInst::ICMP_SLE:
2424 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2425 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2426 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2427 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2428 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2430 case ICmpInst::ICMP_UGE:
2431 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2432 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2433 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2434 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2435 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2437 case ICmpInst::ICMP_ULE:
2438 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2439 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2440 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2441 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2442 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2446 // Turn a signed comparison into an unsigned one if both operands
2447 // are known to have the same sign.
2449 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2450 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2451 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2454 // Test if the ICmpInst instruction is used exclusively by a select as
2455 // part of a minimum or maximum operation. If so, refrain from doing
2456 // any other folding. This helps out other analyses which understand
2457 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2458 // and CodeGen. And in this case, at least one of the comparison
2459 // operands has at least one user besides the compare (the select),
2460 // which would often largely negate the benefit of folding anyway.
2462 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2463 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2464 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2467 // See if we are doing a comparison between a constant and an instruction that
2468 // can be folded into the comparison.
2469 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2470 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2471 // instruction, see if that instruction also has constants so that the
2472 // instruction can be folded into the icmp
2473 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2474 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2478 // Handle icmp with constant (but not simple integer constant) RHS
2479 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2480 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2481 switch (LHSI->getOpcode()) {
2482 case Instruction::GetElementPtr:
2483 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2484 if (RHSC->isNullValue() &&
2485 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2486 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2487 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2489 case Instruction::PHI:
2490 // Only fold icmp into the PHI if the phi and icmp are in the same
2491 // block. If in the same block, we're encouraging jump threading. If
2492 // not, we are just pessimizing the code by making an i1 phi.
2493 if (LHSI->getParent() == I.getParent())
2494 if (Instruction *NV = FoldOpIntoPhi(I))
2497 case Instruction::Select: {
2498 // If either operand of the select is a constant, we can fold the
2499 // comparison into the select arms, which will cause one to be
2500 // constant folded and the select turned into a bitwise or.
2501 Value *Op1 = 0, *Op2 = 0;
2502 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2503 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2504 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2505 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2507 // We only want to perform this transformation if it will not lead to
2508 // additional code. This is true if either both sides of the select
2509 // fold to a constant (in which case the icmp is replaced with a select
2510 // which will usually simplify) or this is the only user of the
2511 // select (in which case we are trading a select+icmp for a simpler
2513 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2515 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2518 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2520 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2524 case Instruction::IntToPtr:
2525 // icmp pred inttoptr(X), null -> icmp pred X, 0
2526 if (RHSC->isNullValue() && TD &&
2527 TD->getIntPtrType(RHSC->getType()) ==
2528 LHSI->getOperand(0)->getType())
2529 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2530 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2533 case Instruction::Load:
2534 // Try to optimize things like "A[i] > 4" to index computations.
2535 if (GetElementPtrInst *GEP =
2536 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2537 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2538 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2539 !cast<LoadInst>(LHSI)->isVolatile())
2540 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2547 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2548 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2549 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2551 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2552 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2553 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2556 // Test to see if the operands of the icmp are casted versions of other
2557 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2559 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2560 if (Op0->getType()->isPointerTy() &&
2561 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2562 // We keep moving the cast from the left operand over to the right
2563 // operand, where it can often be eliminated completely.
2564 Op0 = CI->getOperand(0);
2566 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2567 // so eliminate it as well.
2568 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2569 Op1 = CI2->getOperand(0);
2571 // If Op1 is a constant, we can fold the cast into the constant.
2572 if (Op0->getType() != Op1->getType()) {
2573 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2574 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2576 // Otherwise, cast the RHS right before the icmp
2577 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2580 return new ICmpInst(I.getPredicate(), Op0, Op1);
2584 if (isa<CastInst>(Op0)) {
2585 // Handle the special case of: icmp (cast bool to X), <cst>
2586 // This comes up when you have code like
2589 // For generality, we handle any zero-extension of any operand comparison
2590 // with a constant or another cast from the same type.
2591 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2592 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2596 // Special logic for binary operators.
2597 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2598 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2600 CmpInst::Predicate Pred = I.getPredicate();
2601 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2602 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2603 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2604 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2605 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2606 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2607 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2608 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2609 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2611 // Analyze the case when either Op0 or Op1 is an add instruction.
2612 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2613 Value *A = 0, *B = 0, *C = 0, *D = 0;
2614 if (BO0 && BO0->getOpcode() == Instruction::Add)
2615 A = BO0->getOperand(0), B = BO0->getOperand(1);
2616 if (BO1 && BO1->getOpcode() == Instruction::Add)
2617 C = BO1->getOperand(0), D = BO1->getOperand(1);
2619 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2620 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2621 return new ICmpInst(Pred, A == Op1 ? B : A,
2622 Constant::getNullValue(Op1->getType()));
2624 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2625 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2626 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2629 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2630 if (A && C && (A == C || A == D || B == C || B == D) &&
2631 NoOp0WrapProblem && NoOp1WrapProblem &&
2632 // Try not to increase register pressure.
2633 BO0->hasOneUse() && BO1->hasOneUse()) {
2634 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2637 // C + B == C + D -> B == D
2640 } else if (A == D) {
2641 // D + B == C + D -> B == C
2644 } else if (B == C) {
2645 // A + C == C + D -> A == D
2650 // A + D == C + D -> A == C
2654 return new ICmpInst(Pred, Y, Z);
2657 // icmp slt (X + -1), Y -> icmp sle X, Y
2658 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2659 match(B, m_AllOnes()))
2660 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2662 // icmp sge (X + -1), Y -> icmp sgt X, Y
2663 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2664 match(B, m_AllOnes()))
2665 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2667 // icmp sle (X + 1), Y -> icmp slt X, Y
2668 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
2670 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2672 // icmp sgt (X + 1), Y -> icmp sge X, Y
2673 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
2675 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2677 // if C1 has greater magnitude than C2:
2678 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2679 // s.t. C3 = C1 - C2
2681 // if C2 has greater magnitude than C1:
2682 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2683 // s.t. C3 = C2 - C1
2684 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2685 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2686 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2687 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2688 const APInt &AP1 = C1->getValue();
2689 const APInt &AP2 = C2->getValue();
2690 if (AP1.isNegative() == AP2.isNegative()) {
2691 APInt AP1Abs = C1->getValue().abs();
2692 APInt AP2Abs = C2->getValue().abs();
2693 if (AP1Abs.uge(AP2Abs)) {
2694 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2695 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2696 return new ICmpInst(Pred, NewAdd, C);
2698 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2699 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2700 return new ICmpInst(Pred, A, NewAdd);
2706 // Analyze the case when either Op0 or Op1 is a sub instruction.
2707 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2708 A = 0; B = 0; C = 0; D = 0;
2709 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2710 A = BO0->getOperand(0), B = BO0->getOperand(1);
2711 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2712 C = BO1->getOperand(0), D = BO1->getOperand(1);
2714 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2715 if (A == Op1 && NoOp0WrapProblem)
2716 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2718 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2719 if (C == Op0 && NoOp1WrapProblem)
2720 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2722 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2723 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2724 // Try not to increase register pressure.
2725 BO0->hasOneUse() && BO1->hasOneUse())
2726 return new ICmpInst(Pred, A, C);
2728 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2729 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2730 // Try not to increase register pressure.
2731 BO0->hasOneUse() && BO1->hasOneUse())
2732 return new ICmpInst(Pred, D, B);
2734 BinaryOperator *SRem = NULL;
2735 // icmp (srem X, Y), Y
2736 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2737 Op1 == BO0->getOperand(1))
2739 // icmp Y, (srem X, Y)
2740 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2741 Op0 == BO1->getOperand(1))
2744 // We don't check hasOneUse to avoid increasing register pressure because
2745 // the value we use is the same value this instruction was already using.
2746 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2748 case ICmpInst::ICMP_EQ:
2749 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2750 case ICmpInst::ICMP_NE:
2751 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2752 case ICmpInst::ICMP_SGT:
2753 case ICmpInst::ICMP_SGE:
2754 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2755 Constant::getAllOnesValue(SRem->getType()));
2756 case ICmpInst::ICMP_SLT:
2757 case ICmpInst::ICMP_SLE:
2758 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2759 Constant::getNullValue(SRem->getType()));
2763 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2764 BO0->hasOneUse() && BO1->hasOneUse() &&
2765 BO0->getOperand(1) == BO1->getOperand(1)) {
2766 switch (BO0->getOpcode()) {
2768 case Instruction::Add:
2769 case Instruction::Sub:
2770 case Instruction::Xor:
2771 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2772 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2773 BO1->getOperand(0));
2774 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2775 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2776 if (CI->getValue().isSignBit()) {
2777 ICmpInst::Predicate Pred = I.isSigned()
2778 ? I.getUnsignedPredicate()
2779 : I.getSignedPredicate();
2780 return new ICmpInst(Pred, BO0->getOperand(0),
2781 BO1->getOperand(0));
2784 if (CI->isMaxValue(true)) {
2785 ICmpInst::Predicate Pred = I.isSigned()
2786 ? I.getUnsignedPredicate()
2787 : I.getSignedPredicate();
2788 Pred = I.getSwappedPredicate(Pred);
2789 return new ICmpInst(Pred, BO0->getOperand(0),
2790 BO1->getOperand(0));
2794 case Instruction::Mul:
2795 if (!I.isEquality())
2798 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2799 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2800 // Mask = -1 >> count-trailing-zeros(Cst).
2801 if (!CI->isZero() && !CI->isOne()) {
2802 const APInt &AP = CI->getValue();
2803 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2804 APInt::getLowBitsSet(AP.getBitWidth(),
2806 AP.countTrailingZeros()));
2807 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2808 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2809 return new ICmpInst(I.getPredicate(), And1, And2);
2813 case Instruction::UDiv:
2814 case Instruction::LShr:
2818 case Instruction::SDiv:
2819 case Instruction::AShr:
2820 if (!BO0->isExact() || !BO1->isExact())
2822 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2823 BO1->getOperand(0));
2824 case Instruction::Shl: {
2825 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2826 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2829 if (!NSW && I.isSigned())
2831 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2832 BO1->getOperand(0));
2839 // Transform (A & ~B) == 0 --> (A & B) != 0
2840 // and (A & ~B) != 0 --> (A & B) == 0
2841 // if A is a power of 2.
2842 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2843 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
2844 return new ICmpInst(I.getInversePredicate(),
2845 Builder->CreateAnd(A, B),
2848 // ~x < ~y --> y < x
2849 // ~x < cst --> ~cst < x
2850 if (match(Op0, m_Not(m_Value(A)))) {
2851 if (match(Op1, m_Not(m_Value(B))))
2852 return new ICmpInst(I.getPredicate(), B, A);
2853 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2854 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2857 // (a+b) <u a --> llvm.uadd.with.overflow.
2858 // (a+b) <u b --> llvm.uadd.with.overflow.
2859 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2860 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2861 (Op1 == A || Op1 == B))
2862 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2865 // a >u (a+b) --> llvm.uadd.with.overflow.
2866 // b >u (a+b) --> llvm.uadd.with.overflow.
2867 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2868 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2869 (Op0 == A || Op0 == B))
2870 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2874 if (I.isEquality()) {
2875 Value *A, *B, *C, *D;
2877 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2878 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2879 Value *OtherVal = A == Op1 ? B : A;
2880 return new ICmpInst(I.getPredicate(), OtherVal,
2881 Constant::getNullValue(A->getType()));
2884 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2885 // A^c1 == C^c2 --> A == C^(c1^c2)
2886 ConstantInt *C1, *C2;
2887 if (match(B, m_ConstantInt(C1)) &&
2888 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2889 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
2890 Value *Xor = Builder->CreateXor(C, NC);
2891 return new ICmpInst(I.getPredicate(), A, Xor);
2894 // A^B == A^D -> B == D
2895 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2896 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2897 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2898 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2902 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2903 (A == Op0 || B == Op0)) {
2904 // A == (A^B) -> B == 0
2905 Value *OtherVal = A == Op0 ? B : A;
2906 return new ICmpInst(I.getPredicate(), OtherVal,
2907 Constant::getNullValue(A->getType()));
2910 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2911 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2912 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2913 Value *X = 0, *Y = 0, *Z = 0;
2916 X = B; Y = D; Z = A;
2917 } else if (A == D) {
2918 X = B; Y = C; Z = A;
2919 } else if (B == C) {
2920 X = A; Y = D; Z = B;
2921 } else if (B == D) {
2922 X = A; Y = C; Z = B;
2925 if (X) { // Build (X^Y) & Z
2926 Op1 = Builder->CreateXor(X, Y);
2927 Op1 = Builder->CreateAnd(Op1, Z);
2928 I.setOperand(0, Op1);
2929 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2934 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
2935 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
2937 if ((Op0->hasOneUse() &&
2938 match(Op0, m_ZExt(m_Value(A))) &&
2939 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
2940 (Op1->hasOneUse() &&
2941 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
2942 match(Op1, m_ZExt(m_Value(A))))) {
2943 APInt Pow2 = Cst1->getValue() + 1;
2944 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
2945 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
2946 return new ICmpInst(I.getPredicate(), A,
2947 Builder->CreateTrunc(B, A->getType()));
2950 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2951 // "icmp (and X, mask), cst"
2953 if (Op0->hasOneUse() &&
2954 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2955 m_ConstantInt(ShAmt))))) &&
2956 match(Op1, m_ConstantInt(Cst1)) &&
2957 // Only do this when A has multiple uses. This is most important to do
2958 // when it exposes other optimizations.
2960 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2962 if (ShAmt < ASize) {
2964 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2967 APInt CmpV = Cst1->getValue().zext(ASize);
2970 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2971 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2977 Value *X; ConstantInt *Cst;
2979 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2980 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2983 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2984 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2986 return Changed ? &I : 0;
2994 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2996 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2999 if (!isa<ConstantFP>(RHSC)) return 0;
3000 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3002 // Get the width of the mantissa. We don't want to hack on conversions that
3003 // might lose information from the integer, e.g. "i64 -> float"
3004 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3005 if (MantissaWidth == -1) return 0; // Unknown.
3007 // Check to see that the input is converted from an integer type that is small
3008 // enough that preserves all bits. TODO: check here for "known" sign bits.
3009 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3010 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
3012 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3013 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3017 // If the conversion would lose info, don't hack on this.
3018 if ((int)InputSize > MantissaWidth)
3021 // Otherwise, we can potentially simplify the comparison. We know that it
3022 // will always come through as an integer value and we know the constant is
3023 // not a NAN (it would have been previously simplified).
3024 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3026 ICmpInst::Predicate Pred;
3027 switch (I.getPredicate()) {
3028 default: llvm_unreachable("Unexpected predicate!");
3029 case FCmpInst::FCMP_UEQ:
3030 case FCmpInst::FCMP_OEQ:
3031 Pred = ICmpInst::ICMP_EQ;
3033 case FCmpInst::FCMP_UGT:
3034 case FCmpInst::FCMP_OGT:
3035 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3037 case FCmpInst::FCMP_UGE:
3038 case FCmpInst::FCMP_OGE:
3039 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3041 case FCmpInst::FCMP_ULT:
3042 case FCmpInst::FCMP_OLT:
3043 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3045 case FCmpInst::FCMP_ULE:
3046 case FCmpInst::FCMP_OLE:
3047 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3049 case FCmpInst::FCMP_UNE:
3050 case FCmpInst::FCMP_ONE:
3051 Pred = ICmpInst::ICMP_NE;
3053 case FCmpInst::FCMP_ORD:
3054 return ReplaceInstUsesWith(I, Builder->getTrue());
3055 case FCmpInst::FCMP_UNO:
3056 return ReplaceInstUsesWith(I, Builder->getFalse());
3059 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3061 // Now we know that the APFloat is a normal number, zero or inf.
3063 // See if the FP constant is too large for the integer. For example,
3064 // comparing an i8 to 300.0.
3065 unsigned IntWidth = IntTy->getScalarSizeInBits();
3068 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3069 // and large values.
3070 APFloat SMax(RHS.getSemantics());
3071 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3072 APFloat::rmNearestTiesToEven);
3073 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3074 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3075 Pred == ICmpInst::ICMP_SLE)
3076 return ReplaceInstUsesWith(I, Builder->getTrue());
3077 return ReplaceInstUsesWith(I, Builder->getFalse());
3080 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3081 // +INF and large values.
3082 APFloat UMax(RHS.getSemantics());
3083 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3084 APFloat::rmNearestTiesToEven);
3085 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3086 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3087 Pred == ICmpInst::ICMP_ULE)
3088 return ReplaceInstUsesWith(I, Builder->getTrue());
3089 return ReplaceInstUsesWith(I, Builder->getFalse());
3094 // See if the RHS value is < SignedMin.
3095 APFloat SMin(RHS.getSemantics());
3096 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3097 APFloat::rmNearestTiesToEven);
3098 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3099 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3100 Pred == ICmpInst::ICMP_SGE)
3101 return ReplaceInstUsesWith(I, Builder->getTrue());
3102 return ReplaceInstUsesWith(I, Builder->getFalse());
3105 // See if the RHS value is < UnsignedMin.
3106 APFloat SMin(RHS.getSemantics());
3107 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3108 APFloat::rmNearestTiesToEven);
3109 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3110 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3111 Pred == ICmpInst::ICMP_UGE)
3112 return ReplaceInstUsesWith(I, Builder->getTrue());
3113 return ReplaceInstUsesWith(I, Builder->getFalse());
3117 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3118 // [0, UMAX], but it may still be fractional. See if it is fractional by
3119 // casting the FP value to the integer value and back, checking for equality.
3120 // Don't do this for zero, because -0.0 is not fractional.
3121 Constant *RHSInt = LHSUnsigned
3122 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3123 : ConstantExpr::getFPToSI(RHSC, IntTy);
3124 if (!RHS.isZero()) {
3125 bool Equal = LHSUnsigned
3126 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3127 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3129 // If we had a comparison against a fractional value, we have to adjust
3130 // the compare predicate and sometimes the value. RHSC is rounded towards
3131 // zero at this point.
3133 default: llvm_unreachable("Unexpected integer comparison!");
3134 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3135 return ReplaceInstUsesWith(I, Builder->getTrue());
3136 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3137 return ReplaceInstUsesWith(I, Builder->getFalse());
3138 case ICmpInst::ICMP_ULE:
3139 // (float)int <= 4.4 --> int <= 4
3140 // (float)int <= -4.4 --> false
3141 if (RHS.isNegative())
3142 return ReplaceInstUsesWith(I, Builder->getFalse());
3144 case ICmpInst::ICMP_SLE:
3145 // (float)int <= 4.4 --> int <= 4
3146 // (float)int <= -4.4 --> int < -4
3147 if (RHS.isNegative())
3148 Pred = ICmpInst::ICMP_SLT;
3150 case ICmpInst::ICMP_ULT:
3151 // (float)int < -4.4 --> false
3152 // (float)int < 4.4 --> int <= 4
3153 if (RHS.isNegative())
3154 return ReplaceInstUsesWith(I, Builder->getFalse());
3155 Pred = ICmpInst::ICMP_ULE;
3157 case ICmpInst::ICMP_SLT:
3158 // (float)int < -4.4 --> int < -4
3159 // (float)int < 4.4 --> int <= 4
3160 if (!RHS.isNegative())
3161 Pred = ICmpInst::ICMP_SLE;
3163 case ICmpInst::ICMP_UGT:
3164 // (float)int > 4.4 --> int > 4
3165 // (float)int > -4.4 --> true
3166 if (RHS.isNegative())
3167 return ReplaceInstUsesWith(I, Builder->getTrue());
3169 case ICmpInst::ICMP_SGT:
3170 // (float)int > 4.4 --> int > 4
3171 // (float)int > -4.4 --> int >= -4
3172 if (RHS.isNegative())
3173 Pred = ICmpInst::ICMP_SGE;
3175 case ICmpInst::ICMP_UGE:
3176 // (float)int >= -4.4 --> true
3177 // (float)int >= 4.4 --> int > 4
3178 if (RHS.isNegative())
3179 return ReplaceInstUsesWith(I, Builder->getTrue());
3180 Pred = ICmpInst::ICMP_UGT;
3182 case ICmpInst::ICMP_SGE:
3183 // (float)int >= -4.4 --> int >= -4
3184 // (float)int >= 4.4 --> int > 4
3185 if (!RHS.isNegative())
3186 Pred = ICmpInst::ICMP_SGT;
3192 // Lower this FP comparison into an appropriate integer version of the
3194 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3197 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3198 bool Changed = false;
3200 /// Orders the operands of the compare so that they are listed from most
3201 /// complex to least complex. This puts constants before unary operators,
3202 /// before binary operators.
3203 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3208 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3210 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
3211 return ReplaceInstUsesWith(I, V);
3213 // Simplify 'fcmp pred X, X'
3215 switch (I.getPredicate()) {
3216 default: llvm_unreachable("Unknown predicate!");
3217 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3218 case FCmpInst::FCMP_ULT: // True if unordered or less than
3219 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3220 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3221 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3222 I.setPredicate(FCmpInst::FCMP_UNO);
3223 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3226 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3227 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3228 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3229 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3230 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3231 I.setPredicate(FCmpInst::FCMP_ORD);
3232 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3237 // Handle fcmp with constant RHS
3238 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3239 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3240 switch (LHSI->getOpcode()) {
3241 case Instruction::FPExt: {
3242 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3243 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3244 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3248 const fltSemantics *Sem;
3249 // FIXME: This shouldn't be here.
3250 if (LHSExt->getSrcTy()->isHalfTy())
3251 Sem = &APFloat::IEEEhalf;
3252 else if (LHSExt->getSrcTy()->isFloatTy())
3253 Sem = &APFloat::IEEEsingle;
3254 else if (LHSExt->getSrcTy()->isDoubleTy())
3255 Sem = &APFloat::IEEEdouble;
3256 else if (LHSExt->getSrcTy()->isFP128Ty())
3257 Sem = &APFloat::IEEEquad;
3258 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3259 Sem = &APFloat::x87DoubleExtended;
3260 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3261 Sem = &APFloat::PPCDoubleDouble;
3266 APFloat F = RHSF->getValueAPF();
3267 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3269 // Avoid lossy conversions and denormals. Zero is a special case
3270 // that's OK to convert.
3274 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3275 APFloat::cmpLessThan) || Fabs.isZero()))
3277 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3278 ConstantFP::get(RHSC->getContext(), F));
3281 case Instruction::PHI:
3282 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3283 // block. If in the same block, we're encouraging jump threading. If
3284 // not, we are just pessimizing the code by making an i1 phi.
3285 if (LHSI->getParent() == I.getParent())
3286 if (Instruction *NV = FoldOpIntoPhi(I))
3289 case Instruction::SIToFP:
3290 case Instruction::UIToFP:
3291 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3294 case Instruction::Select: {
3295 // If either operand of the select is a constant, we can fold the
3296 // comparison into the select arms, which will cause one to be
3297 // constant folded and the select turned into a bitwise or.
3298 Value *Op1 = 0, *Op2 = 0;
3299 if (LHSI->hasOneUse()) {
3300 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3301 // Fold the known value into the constant operand.
3302 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3303 // Insert a new FCmp of the other select operand.
3304 Op2 = Builder->CreateFCmp(I.getPredicate(),
3305 LHSI->getOperand(2), RHSC, I.getName());
3306 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3307 // Fold the known value into the constant operand.
3308 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
3309 // Insert a new FCmp of the other select operand.
3310 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
3316 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3319 case Instruction::FSub: {
3320 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3322 if (match(LHSI, m_FNeg(m_Value(Op))))
3323 return new FCmpInst(I.getSwappedPredicate(), Op,
3324 ConstantExpr::getFNeg(RHSC));
3327 case Instruction::Load:
3328 if (GetElementPtrInst *GEP =
3329 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3330 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3331 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3332 !cast<LoadInst>(LHSI)->isVolatile())
3333 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3337 case Instruction::Call: {
3338 CallInst *CI = cast<CallInst>(LHSI);
3340 // Various optimization for fabs compared with zero.
3341 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3342 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3344 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3345 Func == LibFunc::fabsl) {
3346 switch (I.getPredicate()) {
3348 // fabs(x) < 0 --> false
3349 case FCmpInst::FCMP_OLT:
3350 return ReplaceInstUsesWith(I, Builder->getFalse());
3351 // fabs(x) > 0 --> x != 0
3352 case FCmpInst::FCMP_OGT:
3353 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3355 // fabs(x) <= 0 --> x == 0
3356 case FCmpInst::FCMP_OLE:
3357 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3359 // fabs(x) >= 0 --> !isnan(x)
3360 case FCmpInst::FCMP_OGE:
3361 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3363 // fabs(x) == 0 --> x == 0
3364 // fabs(x) != 0 --> x != 0
3365 case FCmpInst::FCMP_OEQ:
3366 case FCmpInst::FCMP_UEQ:
3367 case FCmpInst::FCMP_ONE:
3368 case FCmpInst::FCMP_UNE:
3369 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3378 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3380 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3381 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3383 // fcmp (fpext x), (fpext y) -> fcmp x, y
3384 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3385 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3386 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3387 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3388 RHSExt->getOperand(0));
3390 return Changed ? &I : 0;