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/ConstantRange.h"
19 #include "llvm/IR/DataLayout.h"
20 #include "llvm/IR/GetElementPtrTypeIterator.h"
21 #include "llvm/IR/IntrinsicInst.h"
22 #include "llvm/IR/PatternMatch.h"
23 #include "llvm/Target/TargetLibraryInfo.h"
25 using namespace PatternMatch;
27 #define DEBUG_TYPE "instcombine"
29 static ConstantInt *getOne(Constant *C) {
30 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
33 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
34 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
37 static bool HasAddOverflow(ConstantInt *Result,
38 ConstantInt *In1, ConstantInt *In2,
41 return Result->getValue().ult(In1->getValue());
43 if (In2->isNegative())
44 return Result->getValue().sgt(In1->getValue());
45 return Result->getValue().slt(In1->getValue());
48 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
49 /// overflowed for this type.
50 static bool AddWithOverflow(Constant *&Result, Constant *In1,
51 Constant *In2, bool IsSigned = false) {
52 Result = ConstantExpr::getAdd(In1, In2);
54 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
55 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
56 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
57 if (HasAddOverflow(ExtractElement(Result, Idx),
58 ExtractElement(In1, Idx),
59 ExtractElement(In2, Idx),
66 return HasAddOverflow(cast<ConstantInt>(Result),
67 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
71 static bool HasSubOverflow(ConstantInt *Result,
72 ConstantInt *In1, ConstantInt *In2,
75 return Result->getValue().ugt(In1->getValue());
77 if (In2->isNegative())
78 return Result->getValue().slt(In1->getValue());
80 return Result->getValue().sgt(In1->getValue());
83 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
84 /// overflowed for this type.
85 static bool SubWithOverflow(Constant *&Result, Constant *In1,
86 Constant *In2, bool IsSigned = false) {
87 Result = ConstantExpr::getSub(In1, In2);
89 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
90 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
91 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
92 if (HasSubOverflow(ExtractElement(Result, Idx),
93 ExtractElement(In1, Idx),
94 ExtractElement(In2, Idx),
101 return HasSubOverflow(cast<ConstantInt>(Result),
102 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
106 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
107 /// comparison only checks the sign bit. If it only checks the sign bit, set
108 /// TrueIfSigned if the result of the comparison is true when the input value is
110 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
111 bool &TrueIfSigned) {
113 case ICmpInst::ICMP_SLT: // True if LHS s< 0
115 return RHS->isZero();
116 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
118 return RHS->isAllOnesValue();
119 case ICmpInst::ICMP_SGT: // True if LHS s> -1
120 TrueIfSigned = false;
121 return RHS->isAllOnesValue();
122 case ICmpInst::ICMP_UGT:
123 // True if LHS u> RHS and RHS == high-bit-mask - 1
125 return RHS->isMaxValue(true);
126 case ICmpInst::ICMP_UGE:
127 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
129 return RHS->getValue().isSignBit();
135 /// Returns true if the exploded icmp can be expressed as a signed comparison
136 /// to zero and updates the predicate accordingly.
137 /// The signedness of the comparison is preserved.
138 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
139 if (!ICmpInst::isSigned(pred))
143 return ICmpInst::isRelational(pred);
146 if (pred == ICmpInst::ICMP_SLT) {
147 pred = ICmpInst::ICMP_SLE;
150 } else if (RHS->isAllOnesValue()) {
151 if (pred == ICmpInst::ICMP_SGT) {
152 pred = ICmpInst::ICMP_SGE;
160 // isHighOnes - Return true if the constant is of the form 1+0+.
161 // This is the same as lowones(~X).
162 static bool isHighOnes(const ConstantInt *CI) {
163 return (~CI->getValue() + 1).isPowerOf2();
166 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
167 /// set of known zero and one bits, compute the maximum and minimum values that
168 /// could have the specified known zero and known one bits, returning them in
170 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
171 const APInt& KnownOne,
172 APInt& Min, APInt& Max) {
173 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
174 KnownZero.getBitWidth() == Min.getBitWidth() &&
175 KnownZero.getBitWidth() == Max.getBitWidth() &&
176 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
177 APInt UnknownBits = ~(KnownZero|KnownOne);
179 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
180 // bit if it is unknown.
182 Max = KnownOne|UnknownBits;
184 if (UnknownBits.isNegative()) { // Sign bit is unknown
185 Min.setBit(Min.getBitWidth()-1);
186 Max.clearBit(Max.getBitWidth()-1);
190 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
191 // a set of known zero and one bits, compute the maximum and minimum values that
192 // could have the specified known zero and known one bits, returning them in
194 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
195 const APInt &KnownOne,
196 APInt &Min, APInt &Max) {
197 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
198 KnownZero.getBitWidth() == Min.getBitWidth() &&
199 KnownZero.getBitWidth() == Max.getBitWidth() &&
200 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
201 APInt UnknownBits = ~(KnownZero|KnownOne);
203 // The minimum value is when the unknown bits are all zeros.
205 // The maximum value is when the unknown bits are all ones.
206 Max = KnownOne|UnknownBits;
211 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
212 /// cmp pred (load (gep GV, ...)), cmpcst
213 /// where GV is a global variable with a constant initializer. Try to simplify
214 /// this into some simple computation that does not need the load. For example
215 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
217 /// If AndCst is non-null, then the loaded value is masked with that constant
218 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
219 Instruction *InstCombiner::
220 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
221 CmpInst &ICI, ConstantInt *AndCst) {
222 // We need TD information to know the pointer size unless this is inbounds.
223 if (!GEP->isInBounds() && !DL)
226 Constant *Init = GV->getInitializer();
227 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
230 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
231 if (ArrayElementCount > 1024) return nullptr; // Don't blow up on huge arrays.
233 // There are many forms of this optimization we can handle, for now, just do
234 // the simple index into a single-dimensional array.
236 // Require: GEP GV, 0, i {{, constant indices}}
237 if (GEP->getNumOperands() < 3 ||
238 !isa<ConstantInt>(GEP->getOperand(1)) ||
239 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
240 isa<Constant>(GEP->getOperand(2)))
243 // Check that indices after the variable are constants and in-range for the
244 // type they index. Collect the indices. This is typically for arrays of
246 SmallVector<unsigned, 4> LaterIndices;
248 Type *EltTy = Init->getType()->getArrayElementType();
249 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
250 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
251 if (!Idx) return nullptr; // Variable index.
253 uint64_t IdxVal = Idx->getZExtValue();
254 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
256 if (StructType *STy = dyn_cast<StructType>(EltTy))
257 EltTy = STy->getElementType(IdxVal);
258 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
259 if (IdxVal >= ATy->getNumElements()) return nullptr;
260 EltTy = ATy->getElementType();
262 return nullptr; // Unknown type.
265 LaterIndices.push_back(IdxVal);
268 enum { Overdefined = -3, Undefined = -2 };
270 // Variables for our state machines.
272 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
273 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
274 // and 87 is the second (and last) index. FirstTrueElement is -2 when
275 // undefined, otherwise set to the first true element. SecondTrueElement is
276 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
277 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
279 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
280 // form "i != 47 & i != 87". Same state transitions as for true elements.
281 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
283 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
284 /// define a state machine that triggers for ranges of values that the index
285 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
286 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
287 /// index in the range (inclusive). We use -2 for undefined here because we
288 /// use relative comparisons and don't want 0-1 to match -1.
289 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
291 // MagicBitvector - This is a magic bitvector where we set a bit if the
292 // comparison is true for element 'i'. If there are 64 elements or less in
293 // the array, this will fully represent all the comparison results.
294 uint64_t MagicBitvector = 0;
297 // Scan the array and see if one of our patterns matches.
298 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
299 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
300 Constant *Elt = Init->getAggregateElement(i);
301 if (!Elt) return nullptr;
303 // If this is indexing an array of structures, get the structure element.
304 if (!LaterIndices.empty())
305 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
307 // If the element is masked, handle it.
308 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
310 // Find out if the comparison would be true or false for the i'th element.
311 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
312 CompareRHS, DL, TLI);
313 // If the result is undef for this element, ignore it.
314 if (isa<UndefValue>(C)) {
315 // Extend range state machines to cover this element in case there is an
316 // undef in the middle of the range.
317 if (TrueRangeEnd == (int)i-1)
319 if (FalseRangeEnd == (int)i-1)
324 // If we can't compute the result for any of the elements, we have to give
325 // up evaluating the entire conditional.
326 if (!isa<ConstantInt>(C)) return nullptr;
328 // Otherwise, we know if the comparison is true or false for this element,
329 // update our state machines.
330 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
332 // State machine for single/double/range index comparison.
334 // Update the TrueElement state machine.
335 if (FirstTrueElement == Undefined)
336 FirstTrueElement = TrueRangeEnd = i; // First true element.
338 // Update double-compare state machine.
339 if (SecondTrueElement == Undefined)
340 SecondTrueElement = i;
342 SecondTrueElement = Overdefined;
344 // Update range state machine.
345 if (TrueRangeEnd == (int)i-1)
348 TrueRangeEnd = Overdefined;
351 // Update the FalseElement state machine.
352 if (FirstFalseElement == Undefined)
353 FirstFalseElement = FalseRangeEnd = i; // First false element.
355 // Update double-compare state machine.
356 if (SecondFalseElement == Undefined)
357 SecondFalseElement = i;
359 SecondFalseElement = Overdefined;
361 // Update range state machine.
362 if (FalseRangeEnd == (int)i-1)
365 FalseRangeEnd = Overdefined;
370 // If this element is in range, update our magic bitvector.
371 if (i < 64 && IsTrueForElt)
372 MagicBitvector |= 1ULL << i;
374 // If all of our states become overdefined, bail out early. Since the
375 // predicate is expensive, only check it every 8 elements. This is only
376 // really useful for really huge arrays.
377 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
378 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
379 FalseRangeEnd == Overdefined)
383 // Now that we've scanned the entire array, emit our new comparison(s). We
384 // order the state machines in complexity of the generated code.
385 Value *Idx = GEP->getOperand(2);
387 // If the index is larger than the pointer size of the target, truncate the
388 // index down like the GEP would do implicitly. We don't have to do this for
389 // an inbounds GEP because the index can't be out of range.
390 if (!GEP->isInBounds()) {
391 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
392 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
393 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
394 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
397 // If the comparison is only true for one or two elements, emit direct
399 if (SecondTrueElement != Overdefined) {
400 // None true -> false.
401 if (FirstTrueElement == Undefined)
402 return ReplaceInstUsesWith(ICI, Builder->getFalse());
404 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
406 // True for one element -> 'i == 47'.
407 if (SecondTrueElement == Undefined)
408 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
410 // True for two elements -> 'i == 47 | i == 72'.
411 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
412 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
413 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
414 return BinaryOperator::CreateOr(C1, C2);
417 // If the comparison is only false for one or two elements, emit direct
419 if (SecondFalseElement != Overdefined) {
420 // None false -> true.
421 if (FirstFalseElement == Undefined)
422 return ReplaceInstUsesWith(ICI, Builder->getTrue());
424 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
426 // False for one element -> 'i != 47'.
427 if (SecondFalseElement == Undefined)
428 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
430 // False for two elements -> 'i != 47 & i != 72'.
431 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
432 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
433 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
434 return BinaryOperator::CreateAnd(C1, C2);
437 // If the comparison can be replaced with a range comparison for the elements
438 // where it is true, emit the range check.
439 if (TrueRangeEnd != Overdefined) {
440 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
442 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
443 if (FirstTrueElement) {
444 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
445 Idx = Builder->CreateAdd(Idx, Offs);
448 Value *End = ConstantInt::get(Idx->getType(),
449 TrueRangeEnd-FirstTrueElement+1);
450 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
453 // False range check.
454 if (FalseRangeEnd != Overdefined) {
455 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
456 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
457 if (FirstFalseElement) {
458 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
459 Idx = Builder->CreateAdd(Idx, Offs);
462 Value *End = ConstantInt::get(Idx->getType(),
463 FalseRangeEnd-FirstFalseElement);
464 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
468 // If a magic bitvector captures the entire comparison state
469 // of this load, replace it with computation that does:
470 // ((magic_cst >> i) & 1) != 0
474 // Look for an appropriate type:
475 // - The type of Idx if the magic fits
476 // - The smallest fitting legal type if we have a DataLayout
478 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
481 Ty = DL->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
482 else if (ArrayElementCount <= 32)
483 Ty = Type::getInt32Ty(Init->getContext());
486 Value *V = Builder->CreateIntCast(Idx, Ty, false);
487 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
488 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
489 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
497 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
498 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
499 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
500 /// be complex, and scales are involved. The above expression would also be
501 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
502 /// This later form is less amenable to optimization though, and we are allowed
503 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
505 /// If we can't emit an optimized form for this expression, this returns null.
507 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
508 const DataLayout &DL = *IC.getDataLayout();
509 gep_type_iterator GTI = gep_type_begin(GEP);
511 // Check to see if this gep only has a single variable index. If so, and if
512 // any constant indices are a multiple of its scale, then we can compute this
513 // in terms of the scale of the variable index. For example, if the GEP
514 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
515 // because the expression will cross zero at the same point.
516 unsigned i, e = GEP->getNumOperands();
518 for (i = 1; i != e; ++i, ++GTI) {
519 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
520 // Compute the aggregate offset of constant indices.
521 if (CI->isZero()) continue;
523 // Handle a struct index, which adds its field offset to the pointer.
524 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
525 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
527 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
528 Offset += Size*CI->getSExtValue();
531 // Found our variable index.
536 // If there are no variable indices, we must have a constant offset, just
537 // evaluate it the general way.
538 if (i == e) return nullptr;
540 Value *VariableIdx = GEP->getOperand(i);
541 // Determine the scale factor of the variable element. For example, this is
542 // 4 if the variable index is into an array of i32.
543 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
545 // Verify that there are no other variable indices. If so, emit the hard way.
546 for (++i, ++GTI; i != e; ++i, ++GTI) {
547 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
548 if (!CI) return nullptr;
550 // Compute the aggregate offset of constant indices.
551 if (CI->isZero()) continue;
553 // Handle a struct index, which adds its field offset to the pointer.
554 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
555 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
557 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
558 Offset += Size*CI->getSExtValue();
564 // Okay, we know we have a single variable index, which must be a
565 // pointer/array/vector index. If there is no offset, life is simple, return
567 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
568 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
570 // Cast to intptrty in case a truncation occurs. If an extension is needed,
571 // we don't need to bother extending: the extension won't affect where the
572 // computation crosses zero.
573 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
574 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
579 // Otherwise, there is an index. The computation we will do will be modulo
580 // the pointer size, so get it.
581 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
583 Offset &= PtrSizeMask;
584 VariableScale &= PtrSizeMask;
586 // To do this transformation, any constant index must be a multiple of the
587 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
588 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
589 // multiple of the variable scale.
590 int64_t NewOffs = Offset / (int64_t)VariableScale;
591 if (Offset != NewOffs*(int64_t)VariableScale)
594 // Okay, we can do this evaluation. Start by converting the index to intptr.
595 if (VariableIdx->getType() != IntPtrTy)
596 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
598 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
599 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
602 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
603 /// else. At this point we know that the GEP is on the LHS of the comparison.
604 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
605 ICmpInst::Predicate Cond,
607 // Don't transform signed compares of GEPs into index compares. Even if the
608 // GEP is inbounds, the final add of the base pointer can have signed overflow
609 // and would change the result of the icmp.
610 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
611 // the maximum signed value for the pointer type.
612 if (ICmpInst::isSigned(Cond))
615 // Look through bitcasts and addrspacecasts. We do not however want to remove
617 if (!isa<GetElementPtrInst>(RHS))
618 RHS = RHS->stripPointerCasts();
620 Value *PtrBase = GEPLHS->getOperand(0);
621 if (DL && PtrBase == RHS && GEPLHS->isInBounds()) {
622 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
623 // This transformation (ignoring the base and scales) is valid because we
624 // know pointers can't overflow since the gep is inbounds. See if we can
625 // output an optimized form.
626 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
628 // If not, synthesize the offset the hard way.
630 Offset = EmitGEPOffset(GEPLHS);
631 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
632 Constant::getNullValue(Offset->getType()));
633 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
634 // If the base pointers are different, but the indices are the same, just
635 // compare the base pointer.
636 if (PtrBase != GEPRHS->getOperand(0)) {
637 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
638 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
639 GEPRHS->getOperand(0)->getType();
641 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
642 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
643 IndicesTheSame = false;
647 // If all indices are the same, just compare the base pointers.
649 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
651 // If we're comparing GEPs with two base pointers that only differ in type
652 // and both GEPs have only constant indices or just one use, then fold
653 // the compare with the adjusted indices.
654 if (DL && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
655 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
656 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
657 PtrBase->stripPointerCasts() ==
658 GEPRHS->getOperand(0)->stripPointerCasts()) {
659 Value *LOffset = EmitGEPOffset(GEPLHS);
660 Value *ROffset = EmitGEPOffset(GEPRHS);
662 // If we looked through an addrspacecast between different sized address
663 // spaces, the LHS and RHS pointers are different sized
664 // integers. Truncate to the smaller one.
665 Type *LHSIndexTy = LOffset->getType();
666 Type *RHSIndexTy = ROffset->getType();
667 if (LHSIndexTy != RHSIndexTy) {
668 if (LHSIndexTy->getPrimitiveSizeInBits() <
669 RHSIndexTy->getPrimitiveSizeInBits()) {
670 ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy);
672 LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy);
675 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
677 return ReplaceInstUsesWith(I, Cmp);
680 // Otherwise, the base pointers are different and the indices are
681 // different, bail out.
685 // If one of the GEPs has all zero indices, recurse.
686 bool AllZeros = true;
687 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
688 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
689 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
694 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
695 ICmpInst::getSwappedPredicate(Cond), I);
697 // If the other GEP has all zero indices, recurse.
699 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
700 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
701 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
706 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
708 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
709 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
710 // If the GEPs only differ by one index, compare it.
711 unsigned NumDifferences = 0; // Keep track of # differences.
712 unsigned DiffOperand = 0; // The operand that differs.
713 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
714 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
715 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
716 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
717 // Irreconcilable differences.
721 if (NumDifferences++) break;
726 if (NumDifferences == 0) // SAME GEP?
727 return ReplaceInstUsesWith(I, // No comparison is needed here.
728 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
730 else if (NumDifferences == 1 && GEPsInBounds) {
731 Value *LHSV = GEPLHS->getOperand(DiffOperand);
732 Value *RHSV = GEPRHS->getOperand(DiffOperand);
733 // Make sure we do a signed comparison here.
734 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
738 // Only lower this if the icmp is the only user of the GEP or if we expect
739 // the result to fold to a constant!
742 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
743 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
744 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
745 Value *L = EmitGEPOffset(GEPLHS);
746 Value *R = EmitGEPOffset(GEPRHS);
747 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
753 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
754 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
755 Value *X, ConstantInt *CI,
756 ICmpInst::Predicate Pred) {
757 // If we have X+0, exit early (simplifying logic below) and let it get folded
758 // elsewhere. icmp X+0, X -> icmp X, X
760 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
761 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
764 // (X+4) == X -> false.
765 if (Pred == ICmpInst::ICMP_EQ)
766 return ReplaceInstUsesWith(ICI, Builder->getFalse());
768 // (X+4) != X -> true.
769 if (Pred == ICmpInst::ICMP_NE)
770 return ReplaceInstUsesWith(ICI, Builder->getTrue());
772 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
773 // so the values can never be equal. Similarly for all other "or equals"
776 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
777 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
778 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
779 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
781 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
782 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
785 // (X+1) >u X --> X <u (0-1) --> X != 255
786 // (X+2) >u X --> X <u (0-2) --> X <u 254
787 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
788 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
789 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
791 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
792 ConstantInt *SMax = ConstantInt::get(X->getContext(),
793 APInt::getSignedMaxValue(BitWidth));
795 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
796 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
797 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
798 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
799 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
800 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
801 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
802 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
804 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
805 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
806 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
807 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
808 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
809 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
811 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
812 Constant *C = Builder->getInt(CI->getValue()-1);
813 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
816 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
817 /// and CmpRHS are both known to be integer constants.
818 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
819 ConstantInt *DivRHS) {
820 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
821 const APInt &CmpRHSV = CmpRHS->getValue();
823 // FIXME: If the operand types don't match the type of the divide
824 // then don't attempt this transform. The code below doesn't have the
825 // logic to deal with a signed divide and an unsigned compare (and
826 // vice versa). This is because (x /s C1) <s C2 produces different
827 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
828 // (x /u C1) <u C2. Simply casting the operands and result won't
829 // work. :( The if statement below tests that condition and bails
831 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
832 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
834 if (DivRHS->isZero())
835 return nullptr; // The ProdOV computation fails on divide by zero.
836 if (DivIsSigned && DivRHS->isAllOnesValue())
837 return nullptr; // The overflow computation also screws up here
838 if (DivRHS->isOne()) {
839 // This eliminates some funny cases with INT_MIN.
840 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
844 // Compute Prod = CI * DivRHS. We are essentially solving an equation
845 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
846 // C2 (CI). By solving for X we can turn this into a range check
847 // instead of computing a divide.
848 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
850 // Determine if the product overflows by seeing if the product is
851 // not equal to the divide. Make sure we do the same kind of divide
852 // as in the LHS instruction that we're folding.
853 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
854 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
856 // Get the ICmp opcode
857 ICmpInst::Predicate Pred = ICI.getPredicate();
859 /// If the division is known to be exact, then there is no remainder from the
860 /// divide, so the covered range size is unit, otherwise it is the divisor.
861 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
863 // Figure out the interval that is being checked. For example, a comparison
864 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
865 // Compute this interval based on the constants involved and the signedness of
866 // the compare/divide. This computes a half-open interval, keeping track of
867 // whether either value in the interval overflows. After analysis each
868 // overflow variable is set to 0 if it's corresponding bound variable is valid
869 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
870 int LoOverflow = 0, HiOverflow = 0;
871 Constant *LoBound = nullptr, *HiBound = nullptr;
873 if (!DivIsSigned) { // udiv
874 // e.g. X/5 op 3 --> [15, 20)
876 HiOverflow = LoOverflow = ProdOV;
878 // If this is not an exact divide, then many values in the range collapse
879 // to the same result value.
880 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
883 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
884 if (CmpRHSV == 0) { // (X / pos) op 0
885 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
886 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
888 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
889 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
890 HiOverflow = LoOverflow = ProdOV;
892 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
893 } else { // (X / pos) op neg
894 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
895 HiBound = AddOne(Prod);
896 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
898 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
899 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
902 } else if (DivRHS->isNegative()) { // Divisor is < 0.
904 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
905 if (CmpRHSV == 0) { // (X / neg) op 0
906 // e.g. X/-5 op 0 --> [-4, 5)
907 LoBound = AddOne(RangeSize);
908 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
909 if (HiBound == DivRHS) { // -INTMIN = INTMIN
910 HiOverflow = 1; // [INTMIN+1, overflow)
911 HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
913 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
914 // e.g. X/-5 op 3 --> [-19, -14)
915 HiBound = AddOne(Prod);
916 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
918 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
919 } else { // (X / neg) op neg
920 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
921 LoOverflow = HiOverflow = ProdOV;
923 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
926 // Dividing by a negative swaps the condition. LT <-> GT
927 Pred = ICmpInst::getSwappedPredicate(Pred);
930 Value *X = DivI->getOperand(0);
932 default: llvm_unreachable("Unhandled icmp opcode!");
933 case ICmpInst::ICMP_EQ:
934 if (LoOverflow && HiOverflow)
935 return ReplaceInstUsesWith(ICI, Builder->getFalse());
937 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
938 ICmpInst::ICMP_UGE, X, LoBound);
940 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
941 ICmpInst::ICMP_ULT, X, HiBound);
942 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
944 case ICmpInst::ICMP_NE:
945 if (LoOverflow && HiOverflow)
946 return ReplaceInstUsesWith(ICI, Builder->getTrue());
948 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
949 ICmpInst::ICMP_ULT, X, LoBound);
951 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
952 ICmpInst::ICMP_UGE, X, HiBound);
953 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
954 DivIsSigned, false));
955 case ICmpInst::ICMP_ULT:
956 case ICmpInst::ICMP_SLT:
957 if (LoOverflow == +1) // Low bound is greater than input range.
958 return ReplaceInstUsesWith(ICI, Builder->getTrue());
959 if (LoOverflow == -1) // Low bound is less than input range.
960 return ReplaceInstUsesWith(ICI, Builder->getFalse());
961 return new ICmpInst(Pred, X, LoBound);
962 case ICmpInst::ICMP_UGT:
963 case ICmpInst::ICMP_SGT:
964 if (HiOverflow == +1) // High bound greater than input range.
965 return ReplaceInstUsesWith(ICI, Builder->getFalse());
966 if (HiOverflow == -1) // High bound less than input range.
967 return ReplaceInstUsesWith(ICI, Builder->getTrue());
968 if (Pred == ICmpInst::ICMP_UGT)
969 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
970 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
974 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
975 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
976 ConstantInt *ShAmt) {
977 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
979 // Check that the shift amount is in range. If not, don't perform
980 // undefined shifts. When the shift is visited it will be
982 uint32_t TypeBits = CmpRHSV.getBitWidth();
983 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
984 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
987 if (!ICI.isEquality()) {
988 // If we have an unsigned comparison and an ashr, we can't simplify this.
989 // Similarly for signed comparisons with lshr.
990 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
993 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
994 // by a power of 2. Since we already have logic to simplify these,
995 // transform to div and then simplify the resultant comparison.
996 if (Shr->getOpcode() == Instruction::AShr &&
997 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
1000 // Revisit the shift (to delete it).
1004 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
1007 Shr->getOpcode() == Instruction::AShr ?
1008 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
1009 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
1011 ICI.setOperand(0, Tmp);
1013 // If the builder folded the binop, just return it.
1014 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
1018 // Otherwise, fold this div/compare.
1019 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1020 TheDiv->getOpcode() == Instruction::UDiv);
1022 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1023 assert(Res && "This div/cst should have folded!");
1028 // If we are comparing against bits always shifted out, the
1029 // comparison cannot succeed.
1030 APInt Comp = CmpRHSV << ShAmtVal;
1031 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1032 if (Shr->getOpcode() == Instruction::LShr)
1033 Comp = Comp.lshr(ShAmtVal);
1035 Comp = Comp.ashr(ShAmtVal);
1037 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1038 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1039 Constant *Cst = Builder->getInt1(IsICMP_NE);
1040 return ReplaceInstUsesWith(ICI, Cst);
1043 // Otherwise, check to see if the bits shifted out are known to be zero.
1044 // If so, we can compare against the unshifted value:
1045 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1046 if (Shr->hasOneUse() && Shr->isExact())
1047 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1049 if (Shr->hasOneUse()) {
1050 // Otherwise strength reduce the shift into an and.
1051 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1052 Constant *Mask = Builder->getInt(Val);
1054 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1055 Mask, Shr->getName()+".mask");
1056 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1062 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1064 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1067 const APInt &RHSV = RHS->getValue();
1069 switch (LHSI->getOpcode()) {
1070 case Instruction::Trunc:
1071 if (ICI.isEquality() && LHSI->hasOneUse()) {
1072 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1073 // of the high bits truncated out of x are known.
1074 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1075 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1076 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1077 computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne);
1079 // If all the high bits are known, we can do this xform.
1080 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1081 // Pull in the high bits from known-ones set.
1082 APInt NewRHS = RHS->getValue().zext(SrcBits);
1083 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1084 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1085 Builder->getInt(NewRHS));
1090 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1091 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1092 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1094 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1095 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1096 Value *CompareVal = LHSI->getOperand(0);
1098 // If the sign bit of the XorCst is not set, there is no change to
1099 // the operation, just stop using the Xor.
1100 if (!XorCst->isNegative()) {
1101 ICI.setOperand(0, CompareVal);
1106 // Was the old condition true if the operand is positive?
1107 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1109 // If so, the new one isn't.
1110 isTrueIfPositive ^= true;
1112 if (isTrueIfPositive)
1113 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1116 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1120 if (LHSI->hasOneUse()) {
1121 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1122 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1123 const APInt &SignBit = XorCst->getValue();
1124 ICmpInst::Predicate Pred = ICI.isSigned()
1125 ? ICI.getUnsignedPredicate()
1126 : ICI.getSignedPredicate();
1127 return new ICmpInst(Pred, LHSI->getOperand(0),
1128 Builder->getInt(RHSV ^ SignBit));
1131 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1132 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1133 const APInt &NotSignBit = XorCst->getValue();
1134 ICmpInst::Predicate Pred = ICI.isSigned()
1135 ? ICI.getUnsignedPredicate()
1136 : ICI.getSignedPredicate();
1137 Pred = ICI.getSwappedPredicate(Pred);
1138 return new ICmpInst(Pred, LHSI->getOperand(0),
1139 Builder->getInt(RHSV ^ NotSignBit));
1143 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1144 // iff -C is a power of 2
1145 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1146 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1147 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1149 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1150 // iff -C is a power of 2
1151 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1152 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1153 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1156 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1157 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1158 LHSI->getOperand(0)->hasOneUse()) {
1159 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1161 // If the LHS is an AND of a truncating cast, we can widen the
1162 // and/compare to be the input width without changing the value
1163 // produced, eliminating a cast.
1164 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1165 // We can do this transformation if either the AND constant does not
1166 // have its sign bit set or if it is an equality comparison.
1167 // Extending a relational comparison when we're checking the sign
1168 // bit would not work.
1169 if (ICI.isEquality() ||
1170 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1172 Builder->CreateAnd(Cast->getOperand(0),
1173 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1174 NewAnd->takeName(LHSI);
1175 return new ICmpInst(ICI.getPredicate(), NewAnd,
1176 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1180 // If the LHS is an AND of a zext, and we have an equality compare, we can
1181 // shrink the and/compare to the smaller type, eliminating the cast.
1182 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1183 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1184 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1185 // should fold the icmp to true/false in that case.
1186 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1188 Builder->CreateAnd(Cast->getOperand(0),
1189 ConstantExpr::getTrunc(AndCst, Ty));
1190 NewAnd->takeName(LHSI);
1191 return new ICmpInst(ICI.getPredicate(), NewAnd,
1192 ConstantExpr::getTrunc(RHS, Ty));
1196 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1197 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1198 // happens a LOT in code produced by the C front-end, for bitfield
1200 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1201 if (Shift && !Shift->isShift())
1205 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
1207 // This seemingly simple opportunity to fold away a shift turns out to
1208 // be rather complicated. See PR17827
1209 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1211 bool CanFold = false;
1212 unsigned ShiftOpcode = Shift->getOpcode();
1213 if (ShiftOpcode == Instruction::AShr) {
1214 // There may be some constraints that make this possible,
1215 // but nothing simple has been discovered yet.
1217 } else if (ShiftOpcode == Instruction::Shl) {
1218 // For a left shift, we can fold if the comparison is not signed.
1219 // We can also fold a signed comparison if the mask value and
1220 // comparison value are not negative. These constraints may not be
1221 // obvious, but we can prove that they are correct using an SMT
1223 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1225 } else if (ShiftOpcode == Instruction::LShr) {
1226 // For a logical right shift, we can fold if the comparison is not
1227 // signed. We can also fold a signed comparison if the shifted mask
1228 // value and the shifted comparison value are not negative.
1229 // These constraints may not be obvious, but we can prove that they
1230 // are correct using an SMT solver.
1231 if (!ICI.isSigned())
1234 ConstantInt *ShiftedAndCst =
1235 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1236 ConstantInt *ShiftedRHSCst =
1237 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1239 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1246 if (ShiftOpcode == Instruction::Shl)
1247 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1249 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1251 // Check to see if we are shifting out any of the bits being
1253 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1254 // If we shifted bits out, the fold is not going to work out.
1255 // As a special case, check to see if this means that the
1256 // result is always true or false now.
1257 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1258 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1259 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1260 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1262 ICI.setOperand(1, NewCst);
1263 Constant *NewAndCst;
1264 if (ShiftOpcode == Instruction::Shl)
1265 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1267 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1268 LHSI->setOperand(1, NewAndCst);
1269 LHSI->setOperand(0, Shift->getOperand(0));
1270 Worklist.Add(Shift); // Shift is dead.
1276 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1277 // preferable because it allows the C<<Y expression to be hoisted out
1278 // of a loop if Y is invariant and X is not.
1279 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1280 ICI.isEquality() && !Shift->isArithmeticShift() &&
1281 !isa<Constant>(Shift->getOperand(0))) {
1284 if (Shift->getOpcode() == Instruction::LShr) {
1285 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1287 // Insert a logical shift.
1288 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1291 // Compute X & (C << Y).
1293 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1295 ICI.setOperand(0, NewAnd);
1299 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1300 // bit set in (X & AndCst) will produce a result greater than RHSV.
1301 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1302 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1303 if ((NTZ < AndCst->getBitWidth()) &&
1304 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1305 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1306 Constant::getNullValue(RHS->getType()));
1310 // Try to optimize things like "A[i]&42 == 0" to index computations.
1311 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1312 if (GetElementPtrInst *GEP =
1313 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1314 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1315 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1316 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1317 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1318 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1323 // X & -C == -C -> X > u ~C
1324 // X & -C != -C -> X <= u ~C
1325 // iff C is a power of 2
1326 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1327 return new ICmpInst(
1328 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1329 : ICmpInst::ICMP_ULE,
1330 LHSI->getOperand(0), SubOne(RHS));
1333 case Instruction::Or: {
1334 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1337 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1338 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1339 // -> and (icmp eq P, null), (icmp eq Q, null).
1340 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1341 Constant::getNullValue(P->getType()));
1342 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1343 Constant::getNullValue(Q->getType()));
1345 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1346 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1348 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1354 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1355 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1358 // If this is a signed comparison to 0 and the mul is sign preserving,
1359 // use the mul LHS operand instead.
1360 ICmpInst::Predicate pred = ICI.getPredicate();
1361 if (isSignTest(pred, RHS) && !Val->isZero() &&
1362 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1363 return new ICmpInst(Val->isNegative() ?
1364 ICmpInst::getSwappedPredicate(pred) : pred,
1365 LHSI->getOperand(0),
1366 Constant::getNullValue(RHS->getType()));
1371 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1372 uint32_t TypeBits = RHSV.getBitWidth();
1373 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1376 // (1 << X) pred P2 -> X pred Log2(P2)
1377 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1378 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1379 ICmpInst::Predicate Pred = ICI.getPredicate();
1380 if (ICI.isUnsigned()) {
1381 if (!RHSVIsPowerOf2) {
1382 // (1 << X) < 30 -> X <= 4
1383 // (1 << X) <= 30 -> X <= 4
1384 // (1 << X) >= 30 -> X > 4
1385 // (1 << X) > 30 -> X > 4
1386 if (Pred == ICmpInst::ICMP_ULT)
1387 Pred = ICmpInst::ICMP_ULE;
1388 else if (Pred == ICmpInst::ICMP_UGE)
1389 Pred = ICmpInst::ICMP_UGT;
1391 unsigned RHSLog2 = RHSV.logBase2();
1393 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1394 // (1 << X) > 2147483648 -> X > 31 -> false
1395 // (1 << X) <= 2147483648 -> X <= 31 -> true
1396 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1397 if (RHSLog2 == TypeBits-1) {
1398 if (Pred == ICmpInst::ICMP_UGE)
1399 Pred = ICmpInst::ICMP_EQ;
1400 else if (Pred == ICmpInst::ICMP_UGT)
1401 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1402 else if (Pred == ICmpInst::ICMP_ULE)
1403 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1404 else if (Pred == ICmpInst::ICMP_ULT)
1405 Pred = ICmpInst::ICMP_NE;
1408 return new ICmpInst(Pred, X,
1409 ConstantInt::get(RHS->getType(), RHSLog2));
1410 } else if (ICI.isSigned()) {
1411 if (RHSV.isAllOnesValue()) {
1412 // (1 << X) <= -1 -> X == 31
1413 if (Pred == ICmpInst::ICMP_SLE)
1414 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1415 ConstantInt::get(RHS->getType(), TypeBits-1));
1417 // (1 << X) > -1 -> X != 31
1418 if (Pred == ICmpInst::ICMP_SGT)
1419 return new ICmpInst(ICmpInst::ICMP_NE, X,
1420 ConstantInt::get(RHS->getType(), TypeBits-1));
1422 // (1 << X) < 0 -> X == 31
1423 // (1 << X) <= 0 -> X == 31
1424 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1425 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1426 ConstantInt::get(RHS->getType(), TypeBits-1));
1428 // (1 << X) >= 0 -> X != 31
1429 // (1 << X) > 0 -> X != 31
1430 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1431 return new ICmpInst(ICmpInst::ICMP_NE, X,
1432 ConstantInt::get(RHS->getType(), TypeBits-1));
1434 } else if (ICI.isEquality()) {
1436 return new ICmpInst(
1437 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1439 return ReplaceInstUsesWith(
1440 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse()
1441 : Builder->getTrue());
1447 // Check that the shift amount is in range. If not, don't perform
1448 // undefined shifts. When the shift is visited it will be
1450 if (ShAmt->uge(TypeBits))
1453 if (ICI.isEquality()) {
1454 // If we are comparing against bits always shifted out, the
1455 // comparison cannot succeed.
1457 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1459 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1460 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1461 Constant *Cst = Builder->getInt1(IsICMP_NE);
1462 return ReplaceInstUsesWith(ICI, Cst);
1465 // If the shift is NUW, then it is just shifting out zeros, no need for an
1467 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1468 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1469 ConstantExpr::getLShr(RHS, ShAmt));
1471 // If the shift is NSW and we compare to 0, then it is just shifting out
1472 // sign bits, no need for an AND either.
1473 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1474 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1475 ConstantExpr::getLShr(RHS, ShAmt));
1477 if (LHSI->hasOneUse()) {
1478 // Otherwise strength reduce the shift into an and.
1479 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1480 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1481 TypeBits - ShAmtVal));
1484 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1485 return new ICmpInst(ICI.getPredicate(), And,
1486 ConstantExpr::getLShr(RHS, ShAmt));
1490 // If this is a signed comparison to 0 and the shift is sign preserving,
1491 // use the shift LHS operand instead.
1492 ICmpInst::Predicate pred = ICI.getPredicate();
1493 if (isSignTest(pred, RHS) &&
1494 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1495 return new ICmpInst(pred,
1496 LHSI->getOperand(0),
1497 Constant::getNullValue(RHS->getType()));
1499 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1500 bool TrueIfSigned = false;
1501 if (LHSI->hasOneUse() &&
1502 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1503 // (X << 31) <s 0 --> (X&1) != 0
1504 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1505 APInt::getOneBitSet(TypeBits,
1506 TypeBits-ShAmt->getZExtValue()-1));
1508 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1509 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1510 And, Constant::getNullValue(And->getType()));
1513 // Transform (icmp pred iM (shl iM %v, N), CI)
1514 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1515 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1516 // This enables to get rid of the shift in favor of a trunc which can be
1517 // free on the target. It has the additional benefit of comparing to a
1518 // smaller constant, which will be target friendly.
1519 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1520 if (LHSI->hasOneUse() &&
1521 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1522 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1523 Constant *NCI = ConstantExpr::getTrunc(
1524 ConstantExpr::getAShr(RHS,
1525 ConstantInt::get(RHS->getType(), Amt)),
1527 return new ICmpInst(ICI.getPredicate(),
1528 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1535 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1536 case Instruction::AShr: {
1537 // Handle equality comparisons of shift-by-constant.
1538 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1539 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1540 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1544 // Handle exact shr's.
1545 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1546 if (RHSV.isMinValue())
1547 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1552 case Instruction::SDiv:
1553 case Instruction::UDiv:
1554 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1555 // Fold this div into the comparison, producing a range check.
1556 // Determine, based on the divide type, what the range is being
1557 // checked. If there is an overflow on the low or high side, remember
1558 // it, otherwise compute the range [low, hi) bounding the new value.
1559 // See: InsertRangeTest above for the kinds of replacements possible.
1560 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1561 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1566 case Instruction::Sub: {
1567 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1569 const APInt &LHSV = LHSC->getValue();
1571 // C1-X <u C2 -> (X|(C2-1)) == C1
1572 // iff C1 & (C2-1) == C2-1
1573 // C2 is a power of 2
1574 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1575 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1576 return new ICmpInst(ICmpInst::ICMP_EQ,
1577 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1580 // C1-X >u C2 -> (X|C2) != C1
1581 // iff C1 & C2 == C2
1582 // C2+1 is a power of 2
1583 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1584 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1585 return new ICmpInst(ICmpInst::ICMP_NE,
1586 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1590 case Instruction::Add:
1591 // Fold: icmp pred (add X, C1), C2
1592 if (!ICI.isEquality()) {
1593 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1595 const APInt &LHSV = LHSC->getValue();
1597 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1600 if (ICI.isSigned()) {
1601 if (CR.getLower().isSignBit()) {
1602 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1603 Builder->getInt(CR.getUpper()));
1604 } else if (CR.getUpper().isSignBit()) {
1605 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1606 Builder->getInt(CR.getLower()));
1609 if (CR.getLower().isMinValue()) {
1610 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1611 Builder->getInt(CR.getUpper()));
1612 } else if (CR.getUpper().isMinValue()) {
1613 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1614 Builder->getInt(CR.getLower()));
1618 // X-C1 <u C2 -> (X & -C2) == C1
1619 // iff C1 & (C2-1) == 0
1620 // C2 is a power of 2
1621 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1622 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1623 return new ICmpInst(ICmpInst::ICMP_EQ,
1624 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1625 ConstantExpr::getNeg(LHSC));
1627 // X-C1 >u C2 -> (X & ~C2) != C1
1629 // C2+1 is a power of 2
1630 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1631 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1632 return new ICmpInst(ICmpInst::ICMP_NE,
1633 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1634 ConstantExpr::getNeg(LHSC));
1639 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1640 if (ICI.isEquality()) {
1641 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1643 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1644 // the second operand is a constant, simplify a bit.
1645 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1646 switch (BO->getOpcode()) {
1647 case Instruction::SRem:
1648 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1649 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1650 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1651 if (V.sgt(1) && V.isPowerOf2()) {
1653 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1655 return new ICmpInst(ICI.getPredicate(), NewRem,
1656 Constant::getNullValue(BO->getType()));
1660 case Instruction::Add:
1661 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1662 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1663 if (BO->hasOneUse())
1664 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1665 ConstantExpr::getSub(RHS, BOp1C));
1666 } else if (RHSV == 0) {
1667 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1668 // efficiently invertible, or if the add has just this one use.
1669 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1671 if (Value *NegVal = dyn_castNegVal(BOp1))
1672 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1673 if (Value *NegVal = dyn_castNegVal(BOp0))
1674 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1675 if (BO->hasOneUse()) {
1676 Value *Neg = Builder->CreateNeg(BOp1);
1678 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1682 case Instruction::Xor:
1683 // For the xor case, we can xor two constants together, eliminating
1684 // the explicit xor.
1685 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1686 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1687 ConstantExpr::getXor(RHS, BOC));
1688 } else if (RHSV == 0) {
1689 // Replace ((xor A, B) != 0) with (A != B)
1690 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1694 case Instruction::Sub:
1695 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1696 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1697 if (BO->hasOneUse())
1698 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1699 ConstantExpr::getSub(BOp0C, RHS));
1700 } else if (RHSV == 0) {
1701 // Replace ((sub A, B) != 0) with (A != B)
1702 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1706 case Instruction::Or:
1707 // If bits are being or'd in that are not present in the constant we
1708 // are comparing against, then the comparison could never succeed!
1709 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1710 Constant *NotCI = ConstantExpr::getNot(RHS);
1711 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1712 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1716 case Instruction::And:
1717 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1718 // If bits are being compared against that are and'd out, then the
1719 // comparison can never succeed!
1720 if ((RHSV & ~BOC->getValue()) != 0)
1721 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1723 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1724 if (RHS == BOC && RHSV.isPowerOf2())
1725 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1726 ICmpInst::ICMP_NE, LHSI,
1727 Constant::getNullValue(RHS->getType()));
1729 // Don't perform the following transforms if the AND has multiple uses
1730 if (!BO->hasOneUse())
1733 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1734 if (BOC->getValue().isSignBit()) {
1735 Value *X = BO->getOperand(0);
1736 Constant *Zero = Constant::getNullValue(X->getType());
1737 ICmpInst::Predicate pred = isICMP_NE ?
1738 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1739 return new ICmpInst(pred, X, Zero);
1742 // ((X & ~7) == 0) --> X < 8
1743 if (RHSV == 0 && isHighOnes(BOC)) {
1744 Value *X = BO->getOperand(0);
1745 Constant *NegX = ConstantExpr::getNeg(BOC);
1746 ICmpInst::Predicate pred = isICMP_NE ?
1747 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1748 return new ICmpInst(pred, X, NegX);
1752 case Instruction::Mul:
1753 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1754 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1755 // The trivial case (mul X, 0) is handled by InstSimplify
1756 // General case : (mul X, C) != 0 iff X != 0
1757 // (mul X, C) == 0 iff X == 0
1759 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1760 Constant::getNullValue(RHS->getType()));
1766 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1767 // Handle icmp {eq|ne} <intrinsic>, intcst.
1768 switch (II->getIntrinsicID()) {
1769 case Intrinsic::bswap:
1771 ICI.setOperand(0, II->getArgOperand(0));
1772 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1774 case Intrinsic::ctlz:
1775 case Intrinsic::cttz:
1776 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1777 if (RHSV == RHS->getType()->getBitWidth()) {
1779 ICI.setOperand(0, II->getArgOperand(0));
1780 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1784 case Intrinsic::ctpop:
1785 // popcount(A) == 0 -> A == 0 and likewise for !=
1786 if (RHS->isZero()) {
1788 ICI.setOperand(0, II->getArgOperand(0));
1789 ICI.setOperand(1, RHS);
1801 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1802 /// We only handle extending casts so far.
1804 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1805 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1806 Value *LHSCIOp = LHSCI->getOperand(0);
1807 Type *SrcTy = LHSCIOp->getType();
1808 Type *DestTy = LHSCI->getType();
1811 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1812 // integer type is the same size as the pointer type.
1813 if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
1814 DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1815 Value *RHSOp = nullptr;
1816 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1817 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1818 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1819 RHSOp = RHSC->getOperand(0);
1820 // If the pointer types don't match, insert a bitcast.
1821 if (LHSCIOp->getType() != RHSOp->getType())
1822 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1826 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1829 // The code below only handles extension cast instructions, so far.
1831 if (LHSCI->getOpcode() != Instruction::ZExt &&
1832 LHSCI->getOpcode() != Instruction::SExt)
1835 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1836 bool isSignedCmp = ICI.isSigned();
1838 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1839 // Not an extension from the same type?
1840 RHSCIOp = CI->getOperand(0);
1841 if (RHSCIOp->getType() != LHSCIOp->getType())
1844 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1845 // and the other is a zext), then we can't handle this.
1846 if (CI->getOpcode() != LHSCI->getOpcode())
1849 // Deal with equality cases early.
1850 if (ICI.isEquality())
1851 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1853 // A signed comparison of sign extended values simplifies into a
1854 // signed comparison.
1855 if (isSignedCmp && isSignedExt)
1856 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1858 // The other three cases all fold into an unsigned comparison.
1859 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1862 // If we aren't dealing with a constant on the RHS, exit early
1863 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1867 // Compute the constant that would happen if we truncated to SrcTy then
1868 // reextended to DestTy.
1869 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1870 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1873 // If the re-extended constant didn't change...
1875 // Deal with equality cases early.
1876 if (ICI.isEquality())
1877 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1879 // A signed comparison of sign extended values simplifies into a
1880 // signed comparison.
1881 if (isSignedExt && isSignedCmp)
1882 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1884 // The other three cases all fold into an unsigned comparison.
1885 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1888 // The re-extended constant changed so the constant cannot be represented
1889 // in the shorter type. Consequently, we cannot emit a simple comparison.
1890 // All the cases that fold to true or false will have already been handled
1891 // by SimplifyICmpInst, so only deal with the tricky case.
1893 if (isSignedCmp || !isSignedExt)
1896 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1897 // should have been folded away previously and not enter in here.
1899 // We're performing an unsigned comp with a sign extended value.
1900 // This is true if the input is >= 0. [aka >s -1]
1901 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1902 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1904 // Finally, return the value computed.
1905 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1906 return ReplaceInstUsesWith(ICI, Result);
1908 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1909 return BinaryOperator::CreateNot(Result);
1912 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1913 /// I = icmp ugt (add (add A, B), CI2), CI1
1914 /// If this is of the form:
1916 /// if (sum+128 >u 255)
1917 /// Then replace it with llvm.sadd.with.overflow.i8.
1919 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1920 ConstantInt *CI2, ConstantInt *CI1,
1922 // The transformation we're trying to do here is to transform this into an
1923 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1924 // with a narrower add, and discard the add-with-constant that is part of the
1925 // range check (if we can't eliminate it, this isn't profitable).
1927 // In order to eliminate the add-with-constant, the compare can be its only
1929 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1930 if (!AddWithCst->hasOneUse()) return nullptr;
1932 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1933 if (!CI2->getValue().isPowerOf2()) return nullptr;
1934 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1935 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
1937 // The width of the new add formed is 1 more than the bias.
1940 // Check to see that CI1 is an all-ones value with NewWidth bits.
1941 if (CI1->getBitWidth() == NewWidth ||
1942 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1945 // This is only really a signed overflow check if the inputs have been
1946 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1947 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1948 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1949 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1950 IC.ComputeNumSignBits(B) < NeededSignBits)
1953 // In order to replace the original add with a narrower
1954 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1955 // and truncates that discard the high bits of the add. Verify that this is
1957 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1958 for (User *U : OrigAdd->users()) {
1959 if (U == AddWithCst) continue;
1961 // Only accept truncates for now. We would really like a nice recursive
1962 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1963 // chain to see which bits of a value are actually demanded. If the
1964 // original add had another add which was then immediately truncated, we
1965 // could still do the transformation.
1966 TruncInst *TI = dyn_cast<TruncInst>(U);
1967 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1971 // If the pattern matches, truncate the inputs to the narrower type and
1972 // use the sadd_with_overflow intrinsic to efficiently compute both the
1973 // result and the overflow bit.
1974 Module *M = I.getParent()->getParent()->getParent();
1976 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1977 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1980 InstCombiner::BuilderTy *Builder = IC.Builder;
1982 // Put the new code above the original add, in case there are any uses of the
1983 // add between the add and the compare.
1984 Builder->SetInsertPoint(OrigAdd);
1986 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1987 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1988 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1989 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1990 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1992 // The inner add was the result of the narrow add, zero extended to the
1993 // wider type. Replace it with the result computed by the intrinsic.
1994 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1996 // The original icmp gets replaced with the overflow value.
1997 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
2000 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
2002 // Don't bother doing this transformation for pointers, don't do it for
2004 if (!isa<IntegerType>(OrigAddV->getType())) return nullptr;
2006 // If the add is a constant expr, then we don't bother transforming it.
2007 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
2008 if (!OrigAdd) return nullptr;
2010 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
2012 // Put the new code above the original add, in case there are any uses of the
2013 // add between the add and the compare.
2014 InstCombiner::BuilderTy *Builder = IC.Builder;
2015 Builder->SetInsertPoint(OrigAdd);
2017 Module *M = I.getParent()->getParent()->getParent();
2018 Type *Ty = LHS->getType();
2019 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
2020 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
2021 Value *Add = Builder->CreateExtractValue(Call, 0);
2023 IC.ReplaceInstUsesWith(*OrigAdd, Add);
2025 // The original icmp gets replaced with the overflow value.
2026 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2029 /// \brief Recognize and process idiom involving test for multiplication
2032 /// The caller has matched a pattern of the form:
2033 /// I = cmp u (mul(zext A, zext B), V
2034 /// The function checks if this is a test for overflow and if so replaces
2035 /// multiplication with call to 'mul.with.overflow' intrinsic.
2037 /// \param I Compare instruction.
2038 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
2039 /// the compare instruction. Must be of integer type.
2040 /// \param OtherVal The other argument of compare instruction.
2041 /// \returns Instruction which must replace the compare instruction, NULL if no
2042 /// replacement required.
2043 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
2044 Value *OtherVal, InstCombiner &IC) {
2045 // Don't bother doing this transformation for pointers, don't do it for
2047 if (!isa<IntegerType>(MulVal->getType()))
2050 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
2051 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
2052 Instruction *MulInstr = cast<Instruction>(MulVal);
2053 assert(MulInstr->getOpcode() == Instruction::Mul);
2055 Instruction *LHS = cast<Instruction>(MulInstr->getOperand(0)),
2056 *RHS = cast<Instruction>(MulInstr->getOperand(1));
2057 assert(LHS->getOpcode() == Instruction::ZExt);
2058 assert(RHS->getOpcode() == Instruction::ZExt);
2059 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
2061 // Calculate type and width of the result produced by mul.with.overflow.
2062 Type *TyA = A->getType(), *TyB = B->getType();
2063 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
2064 WidthB = TyB->getPrimitiveSizeInBits();
2067 if (WidthB > WidthA) {
2075 // In order to replace the original mul with a narrower mul.with.overflow,
2076 // all uses must ignore upper bits of the product. The number of used low
2077 // bits must be not greater than the width of mul.with.overflow.
2078 if (MulVal->hasNUsesOrMore(2))
2079 for (User *U : MulVal->users()) {
2082 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2083 // Check if truncation ignores bits above MulWidth.
2084 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
2085 if (TruncWidth > MulWidth)
2087 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2088 // Check if AND ignores bits above MulWidth.
2089 if (BO->getOpcode() != Instruction::And)
2091 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
2092 const APInt &CVal = CI->getValue();
2093 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
2097 // Other uses prohibit this transformation.
2102 // Recognize patterns
2103 switch (I.getPredicate()) {
2104 case ICmpInst::ICMP_EQ:
2105 case ICmpInst::ICMP_NE:
2106 // Recognize pattern:
2107 // mulval = mul(zext A, zext B)
2108 // cmp eq/neq mulval, zext trunc mulval
2109 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
2110 if (Zext->hasOneUse()) {
2111 Value *ZextArg = Zext->getOperand(0);
2112 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
2113 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
2117 // Recognize pattern:
2118 // mulval = mul(zext A, zext B)
2119 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
2122 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
2123 if (ValToMask != MulVal)
2125 const APInt &CVal = CI->getValue() + 1;
2126 if (CVal.isPowerOf2()) {
2127 unsigned MaskWidth = CVal.logBase2();
2128 if (MaskWidth == MulWidth)
2129 break; // Recognized
2134 case ICmpInst::ICMP_UGT:
2135 // Recognize pattern:
2136 // mulval = mul(zext A, zext B)
2137 // cmp ugt mulval, max
2138 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2139 APInt MaxVal = APInt::getMaxValue(MulWidth);
2140 MaxVal = MaxVal.zext(CI->getBitWidth());
2141 if (MaxVal.eq(CI->getValue()))
2142 break; // Recognized
2146 case ICmpInst::ICMP_UGE:
2147 // Recognize pattern:
2148 // mulval = mul(zext A, zext B)
2149 // cmp uge mulval, max+1
2150 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2151 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2152 if (MaxVal.eq(CI->getValue()))
2153 break; // Recognized
2157 case ICmpInst::ICMP_ULE:
2158 // Recognize pattern:
2159 // mulval = mul(zext A, zext B)
2160 // cmp ule mulval, max
2161 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2162 APInt MaxVal = APInt::getMaxValue(MulWidth);
2163 MaxVal = MaxVal.zext(CI->getBitWidth());
2164 if (MaxVal.eq(CI->getValue()))
2165 break; // Recognized
2169 case ICmpInst::ICMP_ULT:
2170 // Recognize pattern:
2171 // mulval = mul(zext A, zext B)
2172 // cmp ule mulval, max + 1
2173 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2174 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2175 if (MaxVal.eq(CI->getValue()))
2176 break; // Recognized
2184 InstCombiner::BuilderTy *Builder = IC.Builder;
2185 Builder->SetInsertPoint(MulInstr);
2186 Module *M = I.getParent()->getParent()->getParent();
2188 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
2189 Value *MulA = A, *MulB = B;
2190 if (WidthA < MulWidth)
2191 MulA = Builder->CreateZExt(A, MulType);
2192 if (WidthB < MulWidth)
2193 MulB = Builder->CreateZExt(B, MulType);
2195 Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
2196 CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
2197 IC.Worklist.Add(MulInstr);
2199 // If there are uses of mul result other than the comparison, we know that
2200 // they are truncation or binary AND. Change them to use result of
2201 // mul.with.overflow and adjust properly mask/size.
2202 if (MulVal->hasNUsesOrMore(2)) {
2203 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
2204 for (User *U : MulVal->users()) {
2205 if (U == &I || U == OtherVal)
2207 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2208 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
2209 IC.ReplaceInstUsesWith(*TI, Mul);
2211 TI->setOperand(0, Mul);
2212 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2213 assert(BO->getOpcode() == Instruction::And);
2214 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
2215 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
2216 APInt ShortMask = CI->getValue().trunc(MulWidth);
2217 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
2219 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
2220 IC.Worklist.Add(Zext);
2221 IC.ReplaceInstUsesWith(*BO, Zext);
2223 llvm_unreachable("Unexpected Binary operation");
2225 IC.Worklist.Add(cast<Instruction>(U));
2228 if (isa<Instruction>(OtherVal))
2229 IC.Worklist.Add(cast<Instruction>(OtherVal));
2231 // The original icmp gets replaced with the overflow value, maybe inverted
2232 // depending on predicate.
2233 bool Inverse = false;
2234 switch (I.getPredicate()) {
2235 case ICmpInst::ICMP_NE:
2237 case ICmpInst::ICMP_EQ:
2240 case ICmpInst::ICMP_UGT:
2241 case ICmpInst::ICMP_UGE:
2242 if (I.getOperand(0) == MulVal)
2246 case ICmpInst::ICMP_ULT:
2247 case ICmpInst::ICMP_ULE:
2248 if (I.getOperand(1) == MulVal)
2253 llvm_unreachable("Unexpected predicate");
2256 Value *Res = Builder->CreateExtractValue(Call, 1);
2257 return BinaryOperator::CreateNot(Res);
2260 return ExtractValueInst::Create(Call, 1);
2263 // DemandedBitsLHSMask - When performing a comparison against a constant,
2264 // it is possible that not all the bits in the LHS are demanded. This helper
2265 // method computes the mask that IS demanded.
2266 static APInt DemandedBitsLHSMask(ICmpInst &I,
2267 unsigned BitWidth, bool isSignCheck) {
2269 return APInt::getSignBit(BitWidth);
2271 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2272 if (!CI) return APInt::getAllOnesValue(BitWidth);
2273 const APInt &RHS = CI->getValue();
2275 switch (I.getPredicate()) {
2276 // For a UGT comparison, we don't care about any bits that
2277 // correspond to the trailing ones of the comparand. The value of these
2278 // bits doesn't impact the outcome of the comparison, because any value
2279 // greater than the RHS must differ in a bit higher than these due to carry.
2280 case ICmpInst::ICMP_UGT: {
2281 unsigned trailingOnes = RHS.countTrailingOnes();
2282 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2286 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2287 // Any value less than the RHS must differ in a higher bit because of carries.
2288 case ICmpInst::ICMP_ULT: {
2289 unsigned trailingZeros = RHS.countTrailingZeros();
2290 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2295 return APInt::getAllOnesValue(BitWidth);
2300 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2301 /// should be swapped.
2302 /// The decision is based on how many times these two operands are reused
2303 /// as subtract operands and their positions in those instructions.
2304 /// The rational is that several architectures use the same instruction for
2305 /// both subtract and cmp, thus it is better if the order of those operands
2307 /// \return true if Op0 and Op1 should be swapped.
2308 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2309 const Value * Op1) {
2310 // Filter out pointer value as those cannot appears directly in subtract.
2311 // FIXME: we may want to go through inttoptrs or bitcasts.
2312 if (Op0->getType()->isPointerTy())
2314 // Count every uses of both Op0 and Op1 in a subtract.
2315 // Each time Op0 is the first operand, count -1: swapping is bad, the
2316 // subtract has already the same layout as the compare.
2317 // Each time Op0 is the second operand, count +1: swapping is good, the
2318 // subtract has a different layout as the compare.
2319 // At the end, if the benefit is greater than 0, Op0 should come second to
2320 // expose more CSE opportunities.
2321 int GlobalSwapBenefits = 0;
2322 for (const User *U : Op0->users()) {
2323 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
2324 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2326 // If Op0 is the first argument, this is not beneficial to swap the
2328 int LocalSwapBenefits = -1;
2329 unsigned Op1Idx = 1;
2330 if (BinOp->getOperand(Op1Idx) == Op0) {
2332 LocalSwapBenefits = 1;
2334 if (BinOp->getOperand(Op1Idx) != Op1)
2336 GlobalSwapBenefits += LocalSwapBenefits;
2338 return GlobalSwapBenefits > 0;
2341 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2342 bool Changed = false;
2343 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2344 unsigned Op0Cplxity = getComplexity(Op0);
2345 unsigned Op1Cplxity = getComplexity(Op1);
2347 /// Orders the operands of the compare so that they are listed from most
2348 /// complex to least complex. This puts constants before unary operators,
2349 /// before binary operators.
2350 if (Op0Cplxity < Op1Cplxity ||
2351 (Op0Cplxity == Op1Cplxity &&
2352 swapMayExposeCSEOpportunities(Op0, Op1))) {
2354 std::swap(Op0, Op1);
2358 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL))
2359 return ReplaceInstUsesWith(I, V);
2361 // comparing -val or val with non-zero is the same as just comparing val
2362 // ie, abs(val) != 0 -> val != 0
2363 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2365 Value *Cond, *SelectTrue, *SelectFalse;
2366 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2367 m_Value(SelectFalse)))) {
2368 if (Value *V = dyn_castNegVal(SelectTrue)) {
2369 if (V == SelectFalse)
2370 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2372 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2373 if (V == SelectTrue)
2374 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2379 Type *Ty = Op0->getType();
2381 // icmp's with boolean values can always be turned into bitwise operations
2382 if (Ty->isIntegerTy(1)) {
2383 switch (I.getPredicate()) {
2384 default: llvm_unreachable("Invalid icmp instruction!");
2385 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2386 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2387 return BinaryOperator::CreateNot(Xor);
2389 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2390 return BinaryOperator::CreateXor(Op0, Op1);
2392 case ICmpInst::ICMP_UGT:
2393 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2395 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2396 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2397 return BinaryOperator::CreateAnd(Not, Op1);
2399 case ICmpInst::ICMP_SGT:
2400 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2402 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2403 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2404 return BinaryOperator::CreateAnd(Not, Op0);
2406 case ICmpInst::ICMP_UGE:
2407 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2409 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2410 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2411 return BinaryOperator::CreateOr(Not, Op1);
2413 case ICmpInst::ICMP_SGE:
2414 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2416 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2417 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2418 return BinaryOperator::CreateOr(Not, Op0);
2423 unsigned BitWidth = 0;
2424 if (Ty->isIntOrIntVectorTy())
2425 BitWidth = Ty->getScalarSizeInBits();
2426 else if (DL) // Pointers require DL info to get their size.
2427 BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
2429 bool isSignBit = false;
2431 // See if we are doing a comparison with a constant.
2432 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2433 Value *A = nullptr, *B = nullptr;
2435 // Match the following pattern, which is a common idiom when writing
2436 // overflow-safe integer arithmetic function. The source performs an
2437 // addition in wider type, and explicitly checks for overflow using
2438 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2439 // sadd_with_overflow intrinsic.
2441 // TODO: This could probably be generalized to handle other overflow-safe
2442 // operations if we worked out the formulas to compute the appropriate
2446 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2448 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2449 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2450 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2451 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2455 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2456 if (I.isEquality() && CI->isZero() &&
2457 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2458 // (icmp cond A B) if cond is equality
2459 return new ICmpInst(I.getPredicate(), A, B);
2462 // If we have an icmp le or icmp ge instruction, turn it into the
2463 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2464 // them being folded in the code below. The SimplifyICmpInst code has
2465 // already handled the edge cases for us, so we just assert on them.
2466 switch (I.getPredicate()) {
2468 case ICmpInst::ICMP_ULE:
2469 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2470 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2471 Builder->getInt(CI->getValue()+1));
2472 case ICmpInst::ICMP_SLE:
2473 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2474 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2475 Builder->getInt(CI->getValue()+1));
2476 case ICmpInst::ICMP_UGE:
2477 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2478 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2479 Builder->getInt(CI->getValue()-1));
2480 case ICmpInst::ICMP_SGE:
2481 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2482 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2483 Builder->getInt(CI->getValue()-1));
2486 // If this comparison is a normal comparison, it demands all
2487 // bits, if it is a sign bit comparison, it only demands the sign bit.
2489 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2492 // See if we can fold the comparison based on range information we can get
2493 // by checking whether bits are known to be zero or one in the input.
2494 if (BitWidth != 0) {
2495 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2496 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2498 if (SimplifyDemandedBits(I.getOperandUse(0),
2499 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2500 Op0KnownZero, Op0KnownOne, 0))
2502 if (SimplifyDemandedBits(I.getOperandUse(1),
2503 APInt::getAllOnesValue(BitWidth),
2504 Op1KnownZero, Op1KnownOne, 0))
2507 // Given the known and unknown bits, compute a range that the LHS could be
2508 // in. Compute the Min, Max and RHS values based on the known bits. For the
2509 // EQ and NE we use unsigned values.
2510 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2511 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2513 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2515 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2518 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2520 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2524 // If Min and Max are known to be the same, then SimplifyDemandedBits
2525 // figured out that the LHS is a constant. Just constant fold this now so
2526 // that code below can assume that Min != Max.
2527 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2528 return new ICmpInst(I.getPredicate(),
2529 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2530 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2531 return new ICmpInst(I.getPredicate(), Op0,
2532 ConstantInt::get(Op1->getType(), Op1Min));
2534 // Based on the range information we know about the LHS, see if we can
2535 // simplify this comparison. For example, (x&4) < 8 is always true.
2536 switch (I.getPredicate()) {
2537 default: llvm_unreachable("Unknown icmp opcode!");
2538 case ICmpInst::ICMP_EQ: {
2539 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2540 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2542 // If all bits are known zero except for one, then we know at most one
2543 // bit is set. If the comparison is against zero, then this is a check
2544 // to see if *that* bit is set.
2545 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2546 if (~Op1KnownZero == 0) {
2547 // If the LHS is an AND with the same constant, look through it.
2548 Value *LHS = nullptr;
2549 ConstantInt *LHSC = nullptr;
2550 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2551 LHSC->getValue() != Op0KnownZeroInverted)
2554 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2555 // then turn "((1 << x)&8) == 0" into "x != 3".
2556 // or turn "((1 << x)&7) == 0" into "x > 2".
2558 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2559 APInt ValToCheck = Op0KnownZeroInverted;
2560 if (ValToCheck.isPowerOf2()) {
2561 unsigned CmpVal = ValToCheck.countTrailingZeros();
2562 return new ICmpInst(ICmpInst::ICMP_NE, X,
2563 ConstantInt::get(X->getType(), CmpVal));
2564 } else if ((++ValToCheck).isPowerOf2()) {
2565 unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
2566 return new ICmpInst(ICmpInst::ICMP_UGT, X,
2567 ConstantInt::get(X->getType(), CmpVal));
2571 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2572 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2574 if (Op0KnownZeroInverted == 1 &&
2575 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2576 return new ICmpInst(ICmpInst::ICMP_NE, X,
2577 ConstantInt::get(X->getType(),
2578 CI->countTrailingZeros()));
2583 case ICmpInst::ICMP_NE: {
2584 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2585 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2587 // If all bits are known zero except for one, then we know at most one
2588 // bit is set. If the comparison is against zero, then this is a check
2589 // to see if *that* bit is set.
2590 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2591 if (~Op1KnownZero == 0) {
2592 // If the LHS is an AND with the same constant, look through it.
2593 Value *LHS = nullptr;
2594 ConstantInt *LHSC = nullptr;
2595 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2596 LHSC->getValue() != Op0KnownZeroInverted)
2599 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2600 // then turn "((1 << x)&8) != 0" into "x == 3".
2601 // or turn "((1 << x)&7) != 0" into "x < 3".
2603 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2604 APInt ValToCheck = Op0KnownZeroInverted;
2605 if (ValToCheck.isPowerOf2()) {
2606 unsigned CmpVal = ValToCheck.countTrailingZeros();
2607 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2608 ConstantInt::get(X->getType(), CmpVal));
2609 } else if ((++ValToCheck).isPowerOf2()) {
2610 unsigned CmpVal = ValToCheck.countTrailingZeros();
2611 return new ICmpInst(ICmpInst::ICMP_ULT, X,
2612 ConstantInt::get(X->getType(), CmpVal));
2616 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2617 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2619 if (Op0KnownZeroInverted == 1 &&
2620 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2621 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2622 ConstantInt::get(X->getType(),
2623 CI->countTrailingZeros()));
2628 case ICmpInst::ICMP_ULT:
2629 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2630 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2631 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2632 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2633 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2634 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2635 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2636 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2637 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2638 Builder->getInt(CI->getValue()-1));
2640 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2641 if (CI->isMinValue(true))
2642 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2643 Constant::getAllOnesValue(Op0->getType()));
2646 case ICmpInst::ICMP_UGT:
2647 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2648 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2649 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2650 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2652 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2653 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2654 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2655 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2656 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2657 Builder->getInt(CI->getValue()+1));
2659 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2660 if (CI->isMaxValue(true))
2661 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2662 Constant::getNullValue(Op0->getType()));
2665 case ICmpInst::ICMP_SLT:
2666 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2667 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2668 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2669 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2670 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2671 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2672 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2673 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2674 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2675 Builder->getInt(CI->getValue()-1));
2678 case ICmpInst::ICMP_SGT:
2679 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2680 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2681 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2682 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2684 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2685 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2686 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2687 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2688 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2689 Builder->getInt(CI->getValue()+1));
2692 case ICmpInst::ICMP_SGE:
2693 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2694 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2695 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2696 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2697 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2699 case ICmpInst::ICMP_SLE:
2700 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2701 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2702 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2703 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2704 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2706 case ICmpInst::ICMP_UGE:
2707 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2708 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2709 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2710 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2711 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2713 case ICmpInst::ICMP_ULE:
2714 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2715 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2716 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2717 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2718 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2722 // Turn a signed comparison into an unsigned one if both operands
2723 // are known to have the same sign.
2725 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2726 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2727 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2730 // Test if the ICmpInst instruction is used exclusively by a select as
2731 // part of a minimum or maximum operation. If so, refrain from doing
2732 // any other folding. This helps out other analyses which understand
2733 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2734 // and CodeGen. And in this case, at least one of the comparison
2735 // operands has at least one user besides the compare (the select),
2736 // which would often largely negate the benefit of folding anyway.
2738 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
2739 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2740 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2743 // See if we are doing a comparison between a constant and an instruction that
2744 // can be folded into the comparison.
2745 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2746 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2747 // instruction, see if that instruction also has constants so that the
2748 // instruction can be folded into the icmp
2749 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2750 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2754 // Handle icmp with constant (but not simple integer constant) RHS
2755 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2756 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2757 switch (LHSI->getOpcode()) {
2758 case Instruction::GetElementPtr:
2759 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2760 if (RHSC->isNullValue() &&
2761 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2762 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2763 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2765 case Instruction::PHI:
2766 // Only fold icmp into the PHI if the phi and icmp are in the same
2767 // block. If in the same block, we're encouraging jump threading. If
2768 // not, we are just pessimizing the code by making an i1 phi.
2769 if (LHSI->getParent() == I.getParent())
2770 if (Instruction *NV = FoldOpIntoPhi(I))
2773 case Instruction::Select: {
2774 // If either operand of the select is a constant, we can fold the
2775 // comparison into the select arms, which will cause one to be
2776 // constant folded and the select turned into a bitwise or.
2777 Value *Op1 = nullptr, *Op2 = nullptr;
2778 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2779 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2780 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2781 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2783 // We only want to perform this transformation if it will not lead to
2784 // additional code. This is true if either both sides of the select
2785 // fold to a constant (in which case the icmp is replaced with a select
2786 // which will usually simplify) or this is the only user of the
2787 // select (in which case we are trading a select+icmp for a simpler
2789 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2791 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2794 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2796 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2800 case Instruction::IntToPtr:
2801 // icmp pred inttoptr(X), null -> icmp pred X, 0
2802 if (RHSC->isNullValue() && DL &&
2803 DL->getIntPtrType(RHSC->getType()) ==
2804 LHSI->getOperand(0)->getType())
2805 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2806 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2809 case Instruction::Load:
2810 // Try to optimize things like "A[i] > 4" to index computations.
2811 if (GetElementPtrInst *GEP =
2812 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2813 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2814 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2815 !cast<LoadInst>(LHSI)->isVolatile())
2816 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2823 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2824 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2825 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2827 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2828 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2829 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2832 // Test to see if the operands of the icmp are casted versions of other
2833 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2835 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2836 if (Op0->getType()->isPointerTy() &&
2837 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2838 // We keep moving the cast from the left operand over to the right
2839 // operand, where it can often be eliminated completely.
2840 Op0 = CI->getOperand(0);
2842 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2843 // so eliminate it as well.
2844 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2845 Op1 = CI2->getOperand(0);
2847 // If Op1 is a constant, we can fold the cast into the constant.
2848 if (Op0->getType() != Op1->getType()) {
2849 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2850 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2852 // Otherwise, cast the RHS right before the icmp
2853 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2856 return new ICmpInst(I.getPredicate(), Op0, Op1);
2860 if (isa<CastInst>(Op0)) {
2861 // Handle the special case of: icmp (cast bool to X), <cst>
2862 // This comes up when you have code like
2865 // For generality, we handle any zero-extension of any operand comparison
2866 // with a constant or another cast from the same type.
2867 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2868 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2872 // Special logic for binary operators.
2873 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2874 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2876 CmpInst::Predicate Pred = I.getPredicate();
2877 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2878 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2879 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2880 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2881 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2882 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2883 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2884 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2885 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2887 // Analyze the case when either Op0 or Op1 is an add instruction.
2888 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2889 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2890 if (BO0 && BO0->getOpcode() == Instruction::Add)
2891 A = BO0->getOperand(0), B = BO0->getOperand(1);
2892 if (BO1 && BO1->getOpcode() == Instruction::Add)
2893 C = BO1->getOperand(0), D = BO1->getOperand(1);
2895 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2896 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2897 return new ICmpInst(Pred, A == Op1 ? B : A,
2898 Constant::getNullValue(Op1->getType()));
2900 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2901 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2902 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2905 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2906 if (A && C && (A == C || A == D || B == C || B == D) &&
2907 NoOp0WrapProblem && NoOp1WrapProblem &&
2908 // Try not to increase register pressure.
2909 BO0->hasOneUse() && BO1->hasOneUse()) {
2910 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2913 // C + B == C + D -> B == D
2916 } else if (A == D) {
2917 // D + B == C + D -> B == C
2920 } else if (B == C) {
2921 // A + C == C + D -> A == D
2926 // A + D == C + D -> A == C
2930 return new ICmpInst(Pred, Y, Z);
2933 // icmp slt (X + -1), Y -> icmp sle X, Y
2934 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2935 match(B, m_AllOnes()))
2936 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2938 // icmp sge (X + -1), Y -> icmp sgt X, Y
2939 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2940 match(B, m_AllOnes()))
2941 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2943 // icmp sle (X + 1), Y -> icmp slt X, Y
2944 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
2946 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2948 // icmp sgt (X + 1), Y -> icmp sge X, Y
2949 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
2951 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2953 // if C1 has greater magnitude than C2:
2954 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2955 // s.t. C3 = C1 - C2
2957 // if C2 has greater magnitude than C1:
2958 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2959 // s.t. C3 = C2 - C1
2960 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2961 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2962 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2963 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2964 const APInt &AP1 = C1->getValue();
2965 const APInt &AP2 = C2->getValue();
2966 if (AP1.isNegative() == AP2.isNegative()) {
2967 APInt AP1Abs = C1->getValue().abs();
2968 APInt AP2Abs = C2->getValue().abs();
2969 if (AP1Abs.uge(AP2Abs)) {
2970 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2971 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2972 return new ICmpInst(Pred, NewAdd, C);
2974 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2975 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2976 return new ICmpInst(Pred, A, NewAdd);
2982 // Analyze the case when either Op0 or Op1 is a sub instruction.
2983 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2984 A = nullptr; B = nullptr; C = nullptr; D = nullptr;
2985 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2986 A = BO0->getOperand(0), B = BO0->getOperand(1);
2987 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2988 C = BO1->getOperand(0), D = BO1->getOperand(1);
2990 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2991 if (A == Op1 && NoOp0WrapProblem)
2992 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2994 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2995 if (C == Op0 && NoOp1WrapProblem)
2996 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2998 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2999 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3000 // Try not to increase register pressure.
3001 BO0->hasOneUse() && BO1->hasOneUse())
3002 return new ICmpInst(Pred, A, C);
3004 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3005 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3006 // Try not to increase register pressure.
3007 BO0->hasOneUse() && BO1->hasOneUse())
3008 return new ICmpInst(Pred, D, B);
3010 // icmp (0-X) < cst --> x > -cst
3011 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3013 if (match(BO0, m_Neg(m_Value(X))))
3014 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3015 if (!RHSC->isMinValue(/*isSigned=*/true))
3016 return new ICmpInst(I.getSwappedPredicate(), X,
3017 ConstantExpr::getNeg(RHSC));
3020 BinaryOperator *SRem = nullptr;
3021 // icmp (srem X, Y), Y
3022 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
3023 Op1 == BO0->getOperand(1))
3025 // icmp Y, (srem X, Y)
3026 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3027 Op0 == BO1->getOperand(1))
3030 // We don't check hasOneUse to avoid increasing register pressure because
3031 // the value we use is the same value this instruction was already using.
3032 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3034 case ICmpInst::ICMP_EQ:
3035 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3036 case ICmpInst::ICMP_NE:
3037 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3038 case ICmpInst::ICMP_SGT:
3039 case ICmpInst::ICMP_SGE:
3040 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3041 Constant::getAllOnesValue(SRem->getType()));
3042 case ICmpInst::ICMP_SLT:
3043 case ICmpInst::ICMP_SLE:
3044 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3045 Constant::getNullValue(SRem->getType()));
3049 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
3050 BO0->hasOneUse() && BO1->hasOneUse() &&
3051 BO0->getOperand(1) == BO1->getOperand(1)) {
3052 switch (BO0->getOpcode()) {
3054 case Instruction::Add:
3055 case Instruction::Sub:
3056 case Instruction::Xor:
3057 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3058 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3059 BO1->getOperand(0));
3060 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
3061 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3062 if (CI->getValue().isSignBit()) {
3063 ICmpInst::Predicate Pred = I.isSigned()
3064 ? I.getUnsignedPredicate()
3065 : I.getSignedPredicate();
3066 return new ICmpInst(Pred, BO0->getOperand(0),
3067 BO1->getOperand(0));
3070 if (CI->isMaxValue(true)) {
3071 ICmpInst::Predicate Pred = I.isSigned()
3072 ? I.getUnsignedPredicate()
3073 : I.getSignedPredicate();
3074 Pred = I.getSwappedPredicate(Pred);
3075 return new ICmpInst(Pred, BO0->getOperand(0),
3076 BO1->getOperand(0));
3080 case Instruction::Mul:
3081 if (!I.isEquality())
3084 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3085 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
3086 // Mask = -1 >> count-trailing-zeros(Cst).
3087 if (!CI->isZero() && !CI->isOne()) {
3088 const APInt &AP = CI->getValue();
3089 ConstantInt *Mask = ConstantInt::get(I.getContext(),
3090 APInt::getLowBitsSet(AP.getBitWidth(),
3092 AP.countTrailingZeros()));
3093 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
3094 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
3095 return new ICmpInst(I.getPredicate(), And1, And2);
3099 case Instruction::UDiv:
3100 case Instruction::LShr:
3104 case Instruction::SDiv:
3105 case Instruction::AShr:
3106 if (!BO0->isExact() || !BO1->isExact())
3108 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3109 BO1->getOperand(0));
3110 case Instruction::Shl: {
3111 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3112 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3115 if (!NSW && I.isSigned())
3117 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3118 BO1->getOperand(0));
3125 // Transform (A & ~B) == 0 --> (A & B) != 0
3126 // and (A & ~B) != 0 --> (A & B) == 0
3127 // if A is a power of 2.
3128 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
3129 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
3130 return new ICmpInst(I.getInversePredicate(),
3131 Builder->CreateAnd(A, B),
3134 // ~x < ~y --> y < x
3135 // ~x < cst --> ~cst < x
3136 if (match(Op0, m_Not(m_Value(A)))) {
3137 if (match(Op1, m_Not(m_Value(B))))
3138 return new ICmpInst(I.getPredicate(), B, A);
3139 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3140 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
3143 // (a+b) <u a --> llvm.uadd.with.overflow.
3144 // (a+b) <u b --> llvm.uadd.with.overflow.
3145 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
3146 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
3147 (Op1 == A || Op1 == B))
3148 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
3151 // a >u (a+b) --> llvm.uadd.with.overflow.
3152 // b >u (a+b) --> llvm.uadd.with.overflow.
3153 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
3154 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
3155 (Op0 == A || Op0 == B))
3156 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
3159 // (zext a) * (zext b) --> llvm.umul.with.overflow.
3160 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3161 if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
3164 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3165 if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
3170 if (I.isEquality()) {
3171 Value *A, *B, *C, *D;
3173 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3174 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3175 Value *OtherVal = A == Op1 ? B : A;
3176 return new ICmpInst(I.getPredicate(), OtherVal,
3177 Constant::getNullValue(A->getType()));
3180 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3181 // A^c1 == C^c2 --> A == C^(c1^c2)
3182 ConstantInt *C1, *C2;
3183 if (match(B, m_ConstantInt(C1)) &&
3184 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
3185 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3186 Value *Xor = Builder->CreateXor(C, NC);
3187 return new ICmpInst(I.getPredicate(), A, Xor);
3190 // A^B == A^D -> B == D
3191 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
3192 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
3193 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
3194 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
3198 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
3199 (A == Op0 || B == Op0)) {
3200 // A == (A^B) -> B == 0
3201 Value *OtherVal = A == Op0 ? B : A;
3202 return new ICmpInst(I.getPredicate(), OtherVal,
3203 Constant::getNullValue(A->getType()));
3206 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3207 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3208 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3209 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3212 X = B; Y = D; Z = A;
3213 } else if (A == D) {
3214 X = B; Y = C; Z = A;
3215 } else if (B == C) {
3216 X = A; Y = D; Z = B;
3217 } else if (B == D) {
3218 X = A; Y = C; Z = B;
3221 if (X) { // Build (X^Y) & Z
3222 Op1 = Builder->CreateXor(X, Y);
3223 Op1 = Builder->CreateAnd(Op1, Z);
3224 I.setOperand(0, Op1);
3225 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3230 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3231 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3233 if ((Op0->hasOneUse() &&
3234 match(Op0, m_ZExt(m_Value(A))) &&
3235 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3236 (Op1->hasOneUse() &&
3237 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3238 match(Op1, m_ZExt(m_Value(A))))) {
3239 APInt Pow2 = Cst1->getValue() + 1;
3240 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3241 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3242 return new ICmpInst(I.getPredicate(), A,
3243 Builder->CreateTrunc(B, A->getType()));
3246 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3247 // For lshr and ashr pairs.
3248 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3249 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3250 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3251 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3252 unsigned TypeBits = Cst1->getBitWidth();
3253 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3254 if (ShAmt < TypeBits && ShAmt != 0) {
3255 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
3256 ? ICmpInst::ICMP_UGE
3257 : ICmpInst::ICMP_ULT;
3258 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3259 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3260 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
3264 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3265 // "icmp (and X, mask), cst"
3267 if (Op0->hasOneUse() &&
3268 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
3269 m_ConstantInt(ShAmt))))) &&
3270 match(Op1, m_ConstantInt(Cst1)) &&
3271 // Only do this when A has multiple uses. This is most important to do
3272 // when it exposes other optimizations.
3274 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3276 if (ShAmt < ASize) {
3278 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3281 APInt CmpV = Cst1->getValue().zext(ASize);
3284 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3285 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3291 Value *X; ConstantInt *Cst;
3293 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3294 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3297 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3298 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3300 return Changed ? &I : nullptr;
3303 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3305 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3308 if (!isa<ConstantFP>(RHSC)) return nullptr;
3309 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3311 // Get the width of the mantissa. We don't want to hack on conversions that
3312 // might lose information from the integer, e.g. "i64 -> float"
3313 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3314 if (MantissaWidth == -1) return nullptr; // Unknown.
3316 // Check to see that the input is converted from an integer type that is small
3317 // enough that preserves all bits. TODO: check here for "known" sign bits.
3318 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3319 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
3321 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3322 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3326 // If the conversion would lose info, don't hack on this.
3327 if ((int)InputSize > MantissaWidth)
3330 // Otherwise, we can potentially simplify the comparison. We know that it
3331 // will always come through as an integer value and we know the constant is
3332 // not a NAN (it would have been previously simplified).
3333 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3335 ICmpInst::Predicate Pred;
3336 switch (I.getPredicate()) {
3337 default: llvm_unreachable("Unexpected predicate!");
3338 case FCmpInst::FCMP_UEQ:
3339 case FCmpInst::FCMP_OEQ:
3340 Pred = ICmpInst::ICMP_EQ;
3342 case FCmpInst::FCMP_UGT:
3343 case FCmpInst::FCMP_OGT:
3344 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3346 case FCmpInst::FCMP_UGE:
3347 case FCmpInst::FCMP_OGE:
3348 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3350 case FCmpInst::FCMP_ULT:
3351 case FCmpInst::FCMP_OLT:
3352 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3354 case FCmpInst::FCMP_ULE:
3355 case FCmpInst::FCMP_OLE:
3356 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3358 case FCmpInst::FCMP_UNE:
3359 case FCmpInst::FCMP_ONE:
3360 Pred = ICmpInst::ICMP_NE;
3362 case FCmpInst::FCMP_ORD:
3363 return ReplaceInstUsesWith(I, Builder->getTrue());
3364 case FCmpInst::FCMP_UNO:
3365 return ReplaceInstUsesWith(I, Builder->getFalse());
3368 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3370 // Now we know that the APFloat is a normal number, zero or inf.
3372 // See if the FP constant is too large for the integer. For example,
3373 // comparing an i8 to 300.0.
3374 unsigned IntWidth = IntTy->getScalarSizeInBits();
3377 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3378 // and large values.
3379 APFloat SMax(RHS.getSemantics());
3380 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3381 APFloat::rmNearestTiesToEven);
3382 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3383 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3384 Pred == ICmpInst::ICMP_SLE)
3385 return ReplaceInstUsesWith(I, Builder->getTrue());
3386 return ReplaceInstUsesWith(I, Builder->getFalse());
3389 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3390 // +INF and large values.
3391 APFloat UMax(RHS.getSemantics());
3392 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3393 APFloat::rmNearestTiesToEven);
3394 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3395 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3396 Pred == ICmpInst::ICMP_ULE)
3397 return ReplaceInstUsesWith(I, Builder->getTrue());
3398 return ReplaceInstUsesWith(I, Builder->getFalse());
3403 // See if the RHS value is < SignedMin.
3404 APFloat SMin(RHS.getSemantics());
3405 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3406 APFloat::rmNearestTiesToEven);
3407 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3408 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3409 Pred == ICmpInst::ICMP_SGE)
3410 return ReplaceInstUsesWith(I, Builder->getTrue());
3411 return ReplaceInstUsesWith(I, Builder->getFalse());
3414 // See if the RHS value is < UnsignedMin.
3415 APFloat SMin(RHS.getSemantics());
3416 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3417 APFloat::rmNearestTiesToEven);
3418 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3419 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3420 Pred == ICmpInst::ICMP_UGE)
3421 return ReplaceInstUsesWith(I, Builder->getTrue());
3422 return ReplaceInstUsesWith(I, Builder->getFalse());
3426 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3427 // [0, UMAX], but it may still be fractional. See if it is fractional by
3428 // casting the FP value to the integer value and back, checking for equality.
3429 // Don't do this for zero, because -0.0 is not fractional.
3430 Constant *RHSInt = LHSUnsigned
3431 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3432 : ConstantExpr::getFPToSI(RHSC, IntTy);
3433 if (!RHS.isZero()) {
3434 bool Equal = LHSUnsigned
3435 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3436 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3438 // If we had a comparison against a fractional value, we have to adjust
3439 // the compare predicate and sometimes the value. RHSC is rounded towards
3440 // zero at this point.
3442 default: llvm_unreachable("Unexpected integer comparison!");
3443 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3444 return ReplaceInstUsesWith(I, Builder->getTrue());
3445 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3446 return ReplaceInstUsesWith(I, Builder->getFalse());
3447 case ICmpInst::ICMP_ULE:
3448 // (float)int <= 4.4 --> int <= 4
3449 // (float)int <= -4.4 --> false
3450 if (RHS.isNegative())
3451 return ReplaceInstUsesWith(I, Builder->getFalse());
3453 case ICmpInst::ICMP_SLE:
3454 // (float)int <= 4.4 --> int <= 4
3455 // (float)int <= -4.4 --> int < -4
3456 if (RHS.isNegative())
3457 Pred = ICmpInst::ICMP_SLT;
3459 case ICmpInst::ICMP_ULT:
3460 // (float)int < -4.4 --> false
3461 // (float)int < 4.4 --> int <= 4
3462 if (RHS.isNegative())
3463 return ReplaceInstUsesWith(I, Builder->getFalse());
3464 Pred = ICmpInst::ICMP_ULE;
3466 case ICmpInst::ICMP_SLT:
3467 // (float)int < -4.4 --> int < -4
3468 // (float)int < 4.4 --> int <= 4
3469 if (!RHS.isNegative())
3470 Pred = ICmpInst::ICMP_SLE;
3472 case ICmpInst::ICMP_UGT:
3473 // (float)int > 4.4 --> int > 4
3474 // (float)int > -4.4 --> true
3475 if (RHS.isNegative())
3476 return ReplaceInstUsesWith(I, Builder->getTrue());
3478 case ICmpInst::ICMP_SGT:
3479 // (float)int > 4.4 --> int > 4
3480 // (float)int > -4.4 --> int >= -4
3481 if (RHS.isNegative())
3482 Pred = ICmpInst::ICMP_SGE;
3484 case ICmpInst::ICMP_UGE:
3485 // (float)int >= -4.4 --> true
3486 // (float)int >= 4.4 --> int > 4
3487 if (RHS.isNegative())
3488 return ReplaceInstUsesWith(I, Builder->getTrue());
3489 Pred = ICmpInst::ICMP_UGT;
3491 case ICmpInst::ICMP_SGE:
3492 // (float)int >= -4.4 --> int >= -4
3493 // (float)int >= 4.4 --> int > 4
3494 if (!RHS.isNegative())
3495 Pred = ICmpInst::ICMP_SGT;
3501 // Lower this FP comparison into an appropriate integer version of the
3503 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3506 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3507 bool Changed = false;
3509 /// Orders the operands of the compare so that they are listed from most
3510 /// complex to least complex. This puts constants before unary operators,
3511 /// before binary operators.
3512 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3517 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3519 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL))
3520 return ReplaceInstUsesWith(I, V);
3522 // Simplify 'fcmp pred X, X'
3524 switch (I.getPredicate()) {
3525 default: llvm_unreachable("Unknown predicate!");
3526 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3527 case FCmpInst::FCMP_ULT: // True if unordered or less than
3528 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3529 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3530 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3531 I.setPredicate(FCmpInst::FCMP_UNO);
3532 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3535 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3536 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3537 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3538 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3539 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3540 I.setPredicate(FCmpInst::FCMP_ORD);
3541 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3546 // Handle fcmp with constant RHS
3547 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3548 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3549 switch (LHSI->getOpcode()) {
3550 case Instruction::FPExt: {
3551 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3552 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3553 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3557 const fltSemantics *Sem;
3558 // FIXME: This shouldn't be here.
3559 if (LHSExt->getSrcTy()->isHalfTy())
3560 Sem = &APFloat::IEEEhalf;
3561 else if (LHSExt->getSrcTy()->isFloatTy())
3562 Sem = &APFloat::IEEEsingle;
3563 else if (LHSExt->getSrcTy()->isDoubleTy())
3564 Sem = &APFloat::IEEEdouble;
3565 else if (LHSExt->getSrcTy()->isFP128Ty())
3566 Sem = &APFloat::IEEEquad;
3567 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3568 Sem = &APFloat::x87DoubleExtended;
3569 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3570 Sem = &APFloat::PPCDoubleDouble;
3575 APFloat F = RHSF->getValueAPF();
3576 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3578 // Avoid lossy conversions and denormals. Zero is a special case
3579 // that's OK to convert.
3583 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3584 APFloat::cmpLessThan) || Fabs.isZero()))
3586 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3587 ConstantFP::get(RHSC->getContext(), F));
3590 case Instruction::PHI:
3591 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3592 // block. If in the same block, we're encouraging jump threading. If
3593 // not, we are just pessimizing the code by making an i1 phi.
3594 if (LHSI->getParent() == I.getParent())
3595 if (Instruction *NV = FoldOpIntoPhi(I))
3598 case Instruction::SIToFP:
3599 case Instruction::UIToFP:
3600 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3603 case Instruction::FSub: {
3604 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3606 if (match(LHSI, m_FNeg(m_Value(Op))))
3607 return new FCmpInst(I.getSwappedPredicate(), Op,
3608 ConstantExpr::getFNeg(RHSC));
3611 case Instruction::Load:
3612 if (GetElementPtrInst *GEP =
3613 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3614 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3615 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3616 !cast<LoadInst>(LHSI)->isVolatile())
3617 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3621 case Instruction::Call: {
3622 CallInst *CI = cast<CallInst>(LHSI);
3624 // Various optimization for fabs compared with zero.
3625 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3626 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3628 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3629 Func == LibFunc::fabsl) {
3630 switch (I.getPredicate()) {
3632 // fabs(x) < 0 --> false
3633 case FCmpInst::FCMP_OLT:
3634 return ReplaceInstUsesWith(I, Builder->getFalse());
3635 // fabs(x) > 0 --> x != 0
3636 case FCmpInst::FCMP_OGT:
3637 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3639 // fabs(x) <= 0 --> x == 0
3640 case FCmpInst::FCMP_OLE:
3641 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3643 // fabs(x) >= 0 --> !isnan(x)
3644 case FCmpInst::FCMP_OGE:
3645 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3647 // fabs(x) == 0 --> x == 0
3648 // fabs(x) != 0 --> x != 0
3649 case FCmpInst::FCMP_OEQ:
3650 case FCmpInst::FCMP_UEQ:
3651 case FCmpInst::FCMP_ONE:
3652 case FCmpInst::FCMP_UNE:
3653 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3662 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3664 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3665 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3667 // fcmp (fpext x), (fpext y) -> fcmp x, y
3668 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3669 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3670 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3671 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3672 RHSExt->getOperand(0));
3674 return Changed ? &I : nullptr;