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 if (GEPLHS->hasAllZeroIndices())
687 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
688 ICmpInst::getSwappedPredicate(Cond), I);
690 // If the other GEP has all zero indices, recurse.
691 if (GEPRHS->hasAllZeroIndices())
692 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
694 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
695 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
696 // If the GEPs only differ by one index, compare it.
697 unsigned NumDifferences = 0; // Keep track of # differences.
698 unsigned DiffOperand = 0; // The operand that differs.
699 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
700 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
701 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
702 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
703 // Irreconcilable differences.
707 if (NumDifferences++) break;
712 if (NumDifferences == 0) // SAME GEP?
713 return ReplaceInstUsesWith(I, // No comparison is needed here.
714 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
716 else if (NumDifferences == 1 && GEPsInBounds) {
717 Value *LHSV = GEPLHS->getOperand(DiffOperand);
718 Value *RHSV = GEPRHS->getOperand(DiffOperand);
719 // Make sure we do a signed comparison here.
720 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
724 // Only lower this if the icmp is the only user of the GEP or if we expect
725 // the result to fold to a constant!
728 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
729 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
730 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
731 Value *L = EmitGEPOffset(GEPLHS);
732 Value *R = EmitGEPOffset(GEPRHS);
733 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
739 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
740 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
741 Value *X, ConstantInt *CI,
742 ICmpInst::Predicate Pred) {
743 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
744 // so the values can never be equal. Similarly for all other "or equals"
747 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
748 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
749 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
750 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
752 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
753 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
756 // (X+1) >u X --> X <u (0-1) --> X != 255
757 // (X+2) >u X --> X <u (0-2) --> X <u 254
758 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
759 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
760 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
762 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
763 ConstantInt *SMax = ConstantInt::get(X->getContext(),
764 APInt::getSignedMaxValue(BitWidth));
766 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
767 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
768 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
769 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
770 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
771 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
772 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
773 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
775 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
776 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
777 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
778 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
779 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
780 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
782 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
783 Constant *C = Builder->getInt(CI->getValue()-1);
784 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
787 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
788 /// and CmpRHS are both known to be integer constants.
789 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
790 ConstantInt *DivRHS) {
791 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
792 const APInt &CmpRHSV = CmpRHS->getValue();
794 // FIXME: If the operand types don't match the type of the divide
795 // then don't attempt this transform. The code below doesn't have the
796 // logic to deal with a signed divide and an unsigned compare (and
797 // vice versa). This is because (x /s C1) <s C2 produces different
798 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
799 // (x /u C1) <u C2. Simply casting the operands and result won't
800 // work. :( The if statement below tests that condition and bails
802 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
803 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
805 if (DivRHS->isZero())
806 return nullptr; // The ProdOV computation fails on divide by zero.
807 if (DivIsSigned && DivRHS->isAllOnesValue())
808 return nullptr; // The overflow computation also screws up here
809 if (DivRHS->isOne()) {
810 // This eliminates some funny cases with INT_MIN.
811 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
815 // Compute Prod = CI * DivRHS. We are essentially solving an equation
816 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
817 // C2 (CI). By solving for X we can turn this into a range check
818 // instead of computing a divide.
819 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
821 // Determine if the product overflows by seeing if the product is
822 // not equal to the divide. Make sure we do the same kind of divide
823 // as in the LHS instruction that we're folding.
824 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
825 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
827 // Get the ICmp opcode
828 ICmpInst::Predicate Pred = ICI.getPredicate();
830 /// If the division is known to be exact, then there is no remainder from the
831 /// divide, so the covered range size is unit, otherwise it is the divisor.
832 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
834 // Figure out the interval that is being checked. For example, a comparison
835 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
836 // Compute this interval based on the constants involved and the signedness of
837 // the compare/divide. This computes a half-open interval, keeping track of
838 // whether either value in the interval overflows. After analysis each
839 // overflow variable is set to 0 if it's corresponding bound variable is valid
840 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
841 int LoOverflow = 0, HiOverflow = 0;
842 Constant *LoBound = nullptr, *HiBound = nullptr;
844 if (!DivIsSigned) { // udiv
845 // e.g. X/5 op 3 --> [15, 20)
847 HiOverflow = LoOverflow = ProdOV;
849 // If this is not an exact divide, then many values in the range collapse
850 // to the same result value.
851 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
854 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
855 if (CmpRHSV == 0) { // (X / pos) op 0
856 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
857 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
859 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
860 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
861 HiOverflow = LoOverflow = ProdOV;
863 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
864 } else { // (X / pos) op neg
865 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
866 HiBound = AddOne(Prod);
867 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
869 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
870 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
873 } else if (DivRHS->isNegative()) { // Divisor is < 0.
875 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
876 if (CmpRHSV == 0) { // (X / neg) op 0
877 // e.g. X/-5 op 0 --> [-4, 5)
878 LoBound = AddOne(RangeSize);
879 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
880 if (HiBound == DivRHS) { // -INTMIN = INTMIN
881 HiOverflow = 1; // [INTMIN+1, overflow)
882 HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
884 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
885 // e.g. X/-5 op 3 --> [-19, -14)
886 HiBound = AddOne(Prod);
887 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
889 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
890 } else { // (X / neg) op neg
891 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
892 LoOverflow = HiOverflow = ProdOV;
894 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
897 // Dividing by a negative swaps the condition. LT <-> GT
898 Pred = ICmpInst::getSwappedPredicate(Pred);
901 Value *X = DivI->getOperand(0);
903 default: llvm_unreachable("Unhandled icmp opcode!");
904 case ICmpInst::ICMP_EQ:
905 if (LoOverflow && HiOverflow)
906 return ReplaceInstUsesWith(ICI, Builder->getFalse());
908 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
909 ICmpInst::ICMP_UGE, X, LoBound);
911 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
912 ICmpInst::ICMP_ULT, X, HiBound);
913 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
915 case ICmpInst::ICMP_NE:
916 if (LoOverflow && HiOverflow)
917 return ReplaceInstUsesWith(ICI, Builder->getTrue());
919 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
920 ICmpInst::ICMP_ULT, X, LoBound);
922 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
923 ICmpInst::ICMP_UGE, X, HiBound);
924 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
925 DivIsSigned, false));
926 case ICmpInst::ICMP_ULT:
927 case ICmpInst::ICMP_SLT:
928 if (LoOverflow == +1) // Low bound is greater than input range.
929 return ReplaceInstUsesWith(ICI, Builder->getTrue());
930 if (LoOverflow == -1) // Low bound is less than input range.
931 return ReplaceInstUsesWith(ICI, Builder->getFalse());
932 return new ICmpInst(Pred, X, LoBound);
933 case ICmpInst::ICMP_UGT:
934 case ICmpInst::ICMP_SGT:
935 if (HiOverflow == +1) // High bound greater than input range.
936 return ReplaceInstUsesWith(ICI, Builder->getFalse());
937 if (HiOverflow == -1) // High bound less than input range.
938 return ReplaceInstUsesWith(ICI, Builder->getTrue());
939 if (Pred == ICmpInst::ICMP_UGT)
940 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
941 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
945 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
946 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
947 ConstantInt *ShAmt) {
948 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
950 // Check that the shift amount is in range. If not, don't perform
951 // undefined shifts. When the shift is visited it will be
953 uint32_t TypeBits = CmpRHSV.getBitWidth();
954 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
955 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
958 if (!ICI.isEquality()) {
959 // If we have an unsigned comparison and an ashr, we can't simplify this.
960 // Similarly for signed comparisons with lshr.
961 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
964 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
965 // by a power of 2. Since we already have logic to simplify these,
966 // transform to div and then simplify the resultant comparison.
967 if (Shr->getOpcode() == Instruction::AShr &&
968 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
971 // Revisit the shift (to delete it).
975 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
978 Shr->getOpcode() == Instruction::AShr ?
979 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
980 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
982 ICI.setOperand(0, Tmp);
984 // If the builder folded the binop, just return it.
985 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
989 // Otherwise, fold this div/compare.
990 assert(TheDiv->getOpcode() == Instruction::SDiv ||
991 TheDiv->getOpcode() == Instruction::UDiv);
993 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
994 assert(Res && "This div/cst should have folded!");
999 // If we are comparing against bits always shifted out, the
1000 // comparison cannot succeed.
1001 APInt Comp = CmpRHSV << ShAmtVal;
1002 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1003 if (Shr->getOpcode() == Instruction::LShr)
1004 Comp = Comp.lshr(ShAmtVal);
1006 Comp = Comp.ashr(ShAmtVal);
1008 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1009 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1010 Constant *Cst = Builder->getInt1(IsICMP_NE);
1011 return ReplaceInstUsesWith(ICI, Cst);
1014 // Otherwise, check to see if the bits shifted out are known to be zero.
1015 // If so, we can compare against the unshifted value:
1016 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1017 if (Shr->hasOneUse() && Shr->isExact())
1018 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1020 if (Shr->hasOneUse()) {
1021 // Otherwise strength reduce the shift into an and.
1022 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1023 Constant *Mask = Builder->getInt(Val);
1025 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1026 Mask, Shr->getName()+".mask");
1027 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1032 /// FoldICmpCstShrCst - Handle "(icmp eq/ne (ashr/lshr const2, A), const1)" ->
1033 /// (icmp eq/ne A, Log2(const2/const1)) ->
1034 /// (icmp eq/ne A, Log2(const2) - Log2(const1)).
1035 Instruction *InstCombiner::FoldICmpCstShrCst(ICmpInst &I, Value *Op, Value *A,
1038 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1040 auto getConstant = [&I, this](bool IsTrue) {
1041 if (I.getPredicate() == I.ICMP_NE)
1043 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1046 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1047 if (I.getPredicate() == I.ICMP_NE)
1048 Pred = CmpInst::getInversePredicate(Pred);
1049 return new ICmpInst(Pred, LHS, RHS);
1052 APInt AP1 = CI1->getValue();
1053 APInt AP2 = CI2->getValue();
1055 // Don't bother doing any work for cases which InstSimplify handles.
1058 bool IsAShr = isa<AShrOperator>(Op);
1060 if (AP2.isAllOnesValue())
1062 if (AP2.isNegative() != AP1.isNegative())
1069 // 'A' must be large enough to shift out the highest set bit.
1070 return getICmp(I.ICMP_UGT, A,
1071 ConstantInt::get(A->getType(), AP2.logBase2()));
1074 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1076 // Get the distance between the highest bit that's set.
1078 // Both the constants are negative, take their positive to calculate log.
1079 if (IsAShr && AP1.isNegative())
1080 // Get the ones' complement of AP2 and AP1 when computing the distance.
1081 Shift = (~AP2).logBase2() - (~AP1).logBase2();
1083 Shift = AP2.logBase2() - AP1.logBase2();
1086 if (IsAShr ? AP1 == AP2.ashr(Shift) : AP1 == AP2.lshr(Shift))
1087 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1089 // Shifting const2 will never be equal to const1.
1090 return getConstant(false);
1093 /// FoldICmpCstShlCst - Handle "(icmp eq/ne (shl const2, A), const1)" ->
1094 /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)).
1095 Instruction *InstCombiner::FoldICmpCstShlCst(ICmpInst &I, Value *Op, Value *A,
1098 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1100 auto getConstant = [&I, this](bool IsTrue) {
1101 if (I.getPredicate() == I.ICMP_NE)
1103 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1106 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1107 if (I.getPredicate() == I.ICMP_NE)
1108 Pred = CmpInst::getInversePredicate(Pred);
1109 return new ICmpInst(Pred, LHS, RHS);
1112 APInt AP1 = CI1->getValue();
1113 APInt AP2 = CI2->getValue();
1115 // Don't bother doing any work for cases which InstSimplify handles.
1119 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1121 if (!AP1 && AP2TrailingZeros != 0)
1122 return getICmp(I.ICMP_UGE, A,
1123 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1126 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1128 // Get the distance between the lowest bits that are set.
1129 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1131 if (Shift > 0 && AP2.shl(Shift) == AP1)
1132 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1134 // Shifting const2 will never be equal to const1.
1135 return getConstant(false);
1138 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1140 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1143 const APInt &RHSV = RHS->getValue();
1145 switch (LHSI->getOpcode()) {
1146 case Instruction::Trunc:
1147 if (ICI.isEquality() && LHSI->hasOneUse()) {
1148 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1149 // of the high bits truncated out of x are known.
1150 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1151 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1152 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1153 computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne, 0, &ICI);
1155 // If all the high bits are known, we can do this xform.
1156 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1157 // Pull in the high bits from known-ones set.
1158 APInt NewRHS = RHS->getValue().zext(SrcBits);
1159 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1160 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1161 Builder->getInt(NewRHS));
1166 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1167 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1168 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1170 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1171 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1172 Value *CompareVal = LHSI->getOperand(0);
1174 // If the sign bit of the XorCst is not set, there is no change to
1175 // the operation, just stop using the Xor.
1176 if (!XorCst->isNegative()) {
1177 ICI.setOperand(0, CompareVal);
1182 // Was the old condition true if the operand is positive?
1183 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1185 // If so, the new one isn't.
1186 isTrueIfPositive ^= true;
1188 if (isTrueIfPositive)
1189 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1192 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1196 if (LHSI->hasOneUse()) {
1197 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1198 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1199 const APInt &SignBit = XorCst->getValue();
1200 ICmpInst::Predicate Pred = ICI.isSigned()
1201 ? ICI.getUnsignedPredicate()
1202 : ICI.getSignedPredicate();
1203 return new ICmpInst(Pred, LHSI->getOperand(0),
1204 Builder->getInt(RHSV ^ SignBit));
1207 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1208 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1209 const APInt &NotSignBit = XorCst->getValue();
1210 ICmpInst::Predicate Pred = ICI.isSigned()
1211 ? ICI.getUnsignedPredicate()
1212 : ICI.getSignedPredicate();
1213 Pred = ICI.getSwappedPredicate(Pred);
1214 return new ICmpInst(Pred, LHSI->getOperand(0),
1215 Builder->getInt(RHSV ^ NotSignBit));
1219 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1220 // iff -C is a power of 2
1221 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1222 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1223 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1225 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1226 // iff -C is a power of 2
1227 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1228 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1229 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1232 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1233 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1234 LHSI->getOperand(0)->hasOneUse()) {
1235 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1237 // If the LHS is an AND of a truncating cast, we can widen the
1238 // and/compare to be the input width without changing the value
1239 // produced, eliminating a cast.
1240 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1241 // We can do this transformation if either the AND constant does not
1242 // have its sign bit set or if it is an equality comparison.
1243 // Extending a relational comparison when we're checking the sign
1244 // bit would not work.
1245 if (ICI.isEquality() ||
1246 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1248 Builder->CreateAnd(Cast->getOperand(0),
1249 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1250 NewAnd->takeName(LHSI);
1251 return new ICmpInst(ICI.getPredicate(), NewAnd,
1252 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1256 // If the LHS is an AND of a zext, and we have an equality compare, we can
1257 // shrink the and/compare to the smaller type, eliminating the cast.
1258 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1259 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1260 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1261 // should fold the icmp to true/false in that case.
1262 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1264 Builder->CreateAnd(Cast->getOperand(0),
1265 ConstantExpr::getTrunc(AndCst, Ty));
1266 NewAnd->takeName(LHSI);
1267 return new ICmpInst(ICI.getPredicate(), NewAnd,
1268 ConstantExpr::getTrunc(RHS, Ty));
1272 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1273 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1274 // happens a LOT in code produced by the C front-end, for bitfield
1276 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1277 if (Shift && !Shift->isShift())
1281 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
1283 // This seemingly simple opportunity to fold away a shift turns out to
1284 // be rather complicated. See PR17827
1285 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1287 bool CanFold = false;
1288 unsigned ShiftOpcode = Shift->getOpcode();
1289 if (ShiftOpcode == Instruction::AShr) {
1290 // There may be some constraints that make this possible,
1291 // but nothing simple has been discovered yet.
1293 } else if (ShiftOpcode == Instruction::Shl) {
1294 // For a left shift, we can fold if the comparison is not signed.
1295 // We can also fold a signed comparison if the mask value and
1296 // comparison value are not negative. These constraints may not be
1297 // obvious, but we can prove that they are correct using an SMT
1299 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1301 } else if (ShiftOpcode == Instruction::LShr) {
1302 // For a logical right shift, we can fold if the comparison is not
1303 // signed. We can also fold a signed comparison if the shifted mask
1304 // value and the shifted comparison value are not negative.
1305 // These constraints may not be obvious, but we can prove that they
1306 // are correct using an SMT solver.
1307 if (!ICI.isSigned())
1310 ConstantInt *ShiftedAndCst =
1311 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1312 ConstantInt *ShiftedRHSCst =
1313 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1315 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1322 if (ShiftOpcode == Instruction::Shl)
1323 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1325 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1327 // Check to see if we are shifting out any of the bits being
1329 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1330 // If we shifted bits out, the fold is not going to work out.
1331 // As a special case, check to see if this means that the
1332 // result is always true or false now.
1333 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1334 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1335 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1336 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1338 ICI.setOperand(1, NewCst);
1339 Constant *NewAndCst;
1340 if (ShiftOpcode == Instruction::Shl)
1341 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1343 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1344 LHSI->setOperand(1, NewAndCst);
1345 LHSI->setOperand(0, Shift->getOperand(0));
1346 Worklist.Add(Shift); // Shift is dead.
1352 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1353 // preferable because it allows the C<<Y expression to be hoisted out
1354 // of a loop if Y is invariant and X is not.
1355 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1356 ICI.isEquality() && !Shift->isArithmeticShift() &&
1357 !isa<Constant>(Shift->getOperand(0))) {
1360 if (Shift->getOpcode() == Instruction::LShr) {
1361 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1363 // Insert a logical shift.
1364 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1367 // Compute X & (C << Y).
1369 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1371 ICI.setOperand(0, NewAnd);
1375 // (icmp pred (and (or (lshr X, Y), X), 1), 0) -->
1376 // (icmp pred (and X, (or (shl 1, Y), 1), 0))
1378 // iff pred isn't signed
1380 Value *X, *Y, *LShr;
1381 if (!ICI.isSigned() && RHSV == 0) {
1382 if (match(LHSI->getOperand(1), m_One())) {
1383 Constant *One = cast<Constant>(LHSI->getOperand(1));
1384 Value *Or = LHSI->getOperand(0);
1385 if (match(Or, m_Or(m_Value(LShr), m_Value(X))) &&
1386 match(LShr, m_LShr(m_Specific(X), m_Value(Y)))) {
1387 unsigned UsesRemoved = 0;
1388 if (LHSI->hasOneUse())
1390 if (Or->hasOneUse())
1392 if (LShr->hasOneUse())
1394 Value *NewOr = nullptr;
1395 // Compute X & ((1 << Y) | 1)
1396 if (auto *C = dyn_cast<Constant>(Y)) {
1397 if (UsesRemoved >= 1)
1399 ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1401 if (UsesRemoved >= 3)
1402 NewOr = Builder->CreateOr(Builder->CreateShl(One, Y,
1405 One, Or->getName());
1408 Value *NewAnd = Builder->CreateAnd(X, NewOr, LHSI->getName());
1409 ICI.setOperand(0, NewAnd);
1417 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1418 // bit set in (X & AndCst) will produce a result greater than RHSV.
1419 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1420 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1421 if ((NTZ < AndCst->getBitWidth()) &&
1422 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1423 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1424 Constant::getNullValue(RHS->getType()));
1428 // Try to optimize things like "A[i]&42 == 0" to index computations.
1429 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1430 if (GetElementPtrInst *GEP =
1431 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1432 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1433 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1434 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1435 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1436 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1441 // X & -C == -C -> X > u ~C
1442 // X & -C != -C -> X <= u ~C
1443 // iff C is a power of 2
1444 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1445 return new ICmpInst(
1446 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1447 : ICmpInst::ICMP_ULE,
1448 LHSI->getOperand(0), SubOne(RHS));
1451 case Instruction::Or: {
1452 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1455 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1456 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1457 // -> and (icmp eq P, null), (icmp eq Q, null).
1458 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1459 Constant::getNullValue(P->getType()));
1460 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1461 Constant::getNullValue(Q->getType()));
1463 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1464 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1466 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1472 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1473 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1476 // If this is a signed comparison to 0 and the mul is sign preserving,
1477 // use the mul LHS operand instead.
1478 ICmpInst::Predicate pred = ICI.getPredicate();
1479 if (isSignTest(pred, RHS) && !Val->isZero() &&
1480 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1481 return new ICmpInst(Val->isNegative() ?
1482 ICmpInst::getSwappedPredicate(pred) : pred,
1483 LHSI->getOperand(0),
1484 Constant::getNullValue(RHS->getType()));
1489 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1490 uint32_t TypeBits = RHSV.getBitWidth();
1491 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1494 // (1 << X) pred P2 -> X pred Log2(P2)
1495 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1496 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1497 ICmpInst::Predicate Pred = ICI.getPredicate();
1498 if (ICI.isUnsigned()) {
1499 if (!RHSVIsPowerOf2) {
1500 // (1 << X) < 30 -> X <= 4
1501 // (1 << X) <= 30 -> X <= 4
1502 // (1 << X) >= 30 -> X > 4
1503 // (1 << X) > 30 -> X > 4
1504 if (Pred == ICmpInst::ICMP_ULT)
1505 Pred = ICmpInst::ICMP_ULE;
1506 else if (Pred == ICmpInst::ICMP_UGE)
1507 Pred = ICmpInst::ICMP_UGT;
1509 unsigned RHSLog2 = RHSV.logBase2();
1511 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1512 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1513 if (RHSLog2 == TypeBits-1) {
1514 if (Pred == ICmpInst::ICMP_UGE)
1515 Pred = ICmpInst::ICMP_EQ;
1516 else if (Pred == ICmpInst::ICMP_ULT)
1517 Pred = ICmpInst::ICMP_NE;
1520 return new ICmpInst(Pred, X,
1521 ConstantInt::get(RHS->getType(), RHSLog2));
1522 } else if (ICI.isSigned()) {
1523 if (RHSV.isAllOnesValue()) {
1524 // (1 << X) <= -1 -> X == 31
1525 if (Pred == ICmpInst::ICMP_SLE)
1526 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1527 ConstantInt::get(RHS->getType(), TypeBits-1));
1529 // (1 << X) > -1 -> X != 31
1530 if (Pred == ICmpInst::ICMP_SGT)
1531 return new ICmpInst(ICmpInst::ICMP_NE, X,
1532 ConstantInt::get(RHS->getType(), TypeBits-1));
1534 // (1 << X) < 0 -> X == 31
1535 // (1 << X) <= 0 -> X == 31
1536 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1537 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1538 ConstantInt::get(RHS->getType(), TypeBits-1));
1540 // (1 << X) >= 0 -> X != 31
1541 // (1 << X) > 0 -> X != 31
1542 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1543 return new ICmpInst(ICmpInst::ICMP_NE, X,
1544 ConstantInt::get(RHS->getType(), TypeBits-1));
1546 } else if (ICI.isEquality()) {
1548 return new ICmpInst(
1549 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1555 // Check that the shift amount is in range. If not, don't perform
1556 // undefined shifts. When the shift is visited it will be
1558 if (ShAmt->uge(TypeBits))
1561 if (ICI.isEquality()) {
1562 // If we are comparing against bits always shifted out, the
1563 // comparison cannot succeed.
1565 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1567 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1568 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1569 Constant *Cst = Builder->getInt1(IsICMP_NE);
1570 return ReplaceInstUsesWith(ICI, Cst);
1573 // If the shift is NUW, then it is just shifting out zeros, no need for an
1575 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1576 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1577 ConstantExpr::getLShr(RHS, ShAmt));
1579 // If the shift is NSW and we compare to 0, then it is just shifting out
1580 // sign bits, no need for an AND either.
1581 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1582 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1583 ConstantExpr::getLShr(RHS, ShAmt));
1585 if (LHSI->hasOneUse()) {
1586 // Otherwise strength reduce the shift into an and.
1587 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1588 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1589 TypeBits - ShAmtVal));
1592 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1593 return new ICmpInst(ICI.getPredicate(), And,
1594 ConstantExpr::getLShr(RHS, ShAmt));
1598 // If this is a signed comparison to 0 and the shift is sign preserving,
1599 // use the shift LHS operand instead.
1600 ICmpInst::Predicate pred = ICI.getPredicate();
1601 if (isSignTest(pred, RHS) &&
1602 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1603 return new ICmpInst(pred,
1604 LHSI->getOperand(0),
1605 Constant::getNullValue(RHS->getType()));
1607 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1608 bool TrueIfSigned = false;
1609 if (LHSI->hasOneUse() &&
1610 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1611 // (X << 31) <s 0 --> (X&1) != 0
1612 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1613 APInt::getOneBitSet(TypeBits,
1614 TypeBits-ShAmt->getZExtValue()-1));
1616 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1617 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1618 And, Constant::getNullValue(And->getType()));
1621 // Transform (icmp pred iM (shl iM %v, N), CI)
1622 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1623 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1624 // This enables to get rid of the shift in favor of a trunc which can be
1625 // free on the target. It has the additional benefit of comparing to a
1626 // smaller constant, which will be target friendly.
1627 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1628 if (LHSI->hasOneUse() &&
1629 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1630 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1631 Constant *NCI = ConstantExpr::getTrunc(
1632 ConstantExpr::getAShr(RHS,
1633 ConstantInt::get(RHS->getType(), Amt)),
1635 return new ICmpInst(ICI.getPredicate(),
1636 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1643 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1644 case Instruction::AShr: {
1645 // Handle equality comparisons of shift-by-constant.
1646 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1647 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1648 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1652 // Handle exact shr's.
1653 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1654 if (RHSV.isMinValue())
1655 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1660 case Instruction::SDiv:
1661 case Instruction::UDiv:
1662 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1663 // Fold this div into the comparison, producing a range check.
1664 // Determine, based on the divide type, what the range is being
1665 // checked. If there is an overflow on the low or high side, remember
1666 // it, otherwise compute the range [low, hi) bounding the new value.
1667 // See: InsertRangeTest above for the kinds of replacements possible.
1668 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1669 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1674 case Instruction::Sub: {
1675 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1677 const APInt &LHSV = LHSC->getValue();
1679 // C1-X <u C2 -> (X|(C2-1)) == C1
1680 // iff C1 & (C2-1) == C2-1
1681 // C2 is a power of 2
1682 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1683 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1684 return new ICmpInst(ICmpInst::ICMP_EQ,
1685 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1688 // C1-X >u C2 -> (X|C2) != C1
1689 // iff C1 & C2 == C2
1690 // C2+1 is a power of 2
1691 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1692 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1693 return new ICmpInst(ICmpInst::ICMP_NE,
1694 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1698 case Instruction::Add:
1699 // Fold: icmp pred (add X, C1), C2
1700 if (!ICI.isEquality()) {
1701 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1703 const APInt &LHSV = LHSC->getValue();
1705 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1708 if (ICI.isSigned()) {
1709 if (CR.getLower().isSignBit()) {
1710 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1711 Builder->getInt(CR.getUpper()));
1712 } else if (CR.getUpper().isSignBit()) {
1713 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1714 Builder->getInt(CR.getLower()));
1717 if (CR.getLower().isMinValue()) {
1718 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1719 Builder->getInt(CR.getUpper()));
1720 } else if (CR.getUpper().isMinValue()) {
1721 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1722 Builder->getInt(CR.getLower()));
1726 // X-C1 <u C2 -> (X & -C2) == C1
1727 // iff C1 & (C2-1) == 0
1728 // C2 is a power of 2
1729 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1730 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1731 return new ICmpInst(ICmpInst::ICMP_EQ,
1732 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1733 ConstantExpr::getNeg(LHSC));
1735 // X-C1 >u C2 -> (X & ~C2) != C1
1737 // C2+1 is a power of 2
1738 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1739 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1740 return new ICmpInst(ICmpInst::ICMP_NE,
1741 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1742 ConstantExpr::getNeg(LHSC));
1747 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1748 if (ICI.isEquality()) {
1749 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1751 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1752 // the second operand is a constant, simplify a bit.
1753 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1754 switch (BO->getOpcode()) {
1755 case Instruction::SRem:
1756 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1757 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1758 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1759 if (V.sgt(1) && V.isPowerOf2()) {
1761 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1763 return new ICmpInst(ICI.getPredicate(), NewRem,
1764 Constant::getNullValue(BO->getType()));
1768 case Instruction::Add:
1769 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1770 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1771 if (BO->hasOneUse())
1772 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1773 ConstantExpr::getSub(RHS, BOp1C));
1774 } else if (RHSV == 0) {
1775 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1776 // efficiently invertible, or if the add has just this one use.
1777 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1779 if (Value *NegVal = dyn_castNegVal(BOp1))
1780 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1781 if (Value *NegVal = dyn_castNegVal(BOp0))
1782 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1783 if (BO->hasOneUse()) {
1784 Value *Neg = Builder->CreateNeg(BOp1);
1786 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1790 case Instruction::Xor:
1791 // For the xor case, we can xor two constants together, eliminating
1792 // the explicit xor.
1793 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1794 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1795 ConstantExpr::getXor(RHS, BOC));
1796 } else if (RHSV == 0) {
1797 // Replace ((xor A, B) != 0) with (A != B)
1798 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1802 case Instruction::Sub:
1803 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1804 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1805 if (BO->hasOneUse())
1806 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1807 ConstantExpr::getSub(BOp0C, RHS));
1808 } else if (RHSV == 0) {
1809 // Replace ((sub A, B) != 0) with (A != B)
1810 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1814 case Instruction::Or:
1815 // If bits are being or'd in that are not present in the constant we
1816 // are comparing against, then the comparison could never succeed!
1817 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1818 Constant *NotCI = ConstantExpr::getNot(RHS);
1819 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1820 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1824 case Instruction::And:
1825 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1826 // If bits are being compared against that are and'd out, then the
1827 // comparison can never succeed!
1828 if ((RHSV & ~BOC->getValue()) != 0)
1829 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1831 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1832 if (RHS == BOC && RHSV.isPowerOf2())
1833 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1834 ICmpInst::ICMP_NE, LHSI,
1835 Constant::getNullValue(RHS->getType()));
1837 // Don't perform the following transforms if the AND has multiple uses
1838 if (!BO->hasOneUse())
1841 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1842 if (BOC->getValue().isSignBit()) {
1843 Value *X = BO->getOperand(0);
1844 Constant *Zero = Constant::getNullValue(X->getType());
1845 ICmpInst::Predicate pred = isICMP_NE ?
1846 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1847 return new ICmpInst(pred, X, Zero);
1850 // ((X & ~7) == 0) --> X < 8
1851 if (RHSV == 0 && isHighOnes(BOC)) {
1852 Value *X = BO->getOperand(0);
1853 Constant *NegX = ConstantExpr::getNeg(BOC);
1854 ICmpInst::Predicate pred = isICMP_NE ?
1855 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1856 return new ICmpInst(pred, X, NegX);
1860 case Instruction::Mul:
1861 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1862 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1863 // The trivial case (mul X, 0) is handled by InstSimplify
1864 // General case : (mul X, C) != 0 iff X != 0
1865 // (mul X, C) == 0 iff X == 0
1867 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1868 Constant::getNullValue(RHS->getType()));
1874 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1875 // Handle icmp {eq|ne} <intrinsic>, intcst.
1876 switch (II->getIntrinsicID()) {
1877 case Intrinsic::bswap:
1879 ICI.setOperand(0, II->getArgOperand(0));
1880 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1882 case Intrinsic::ctlz:
1883 case Intrinsic::cttz:
1884 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1885 if (RHSV == RHS->getType()->getBitWidth()) {
1887 ICI.setOperand(0, II->getArgOperand(0));
1888 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1892 case Intrinsic::ctpop:
1893 // popcount(A) == 0 -> A == 0 and likewise for !=
1894 if (RHS->isZero()) {
1896 ICI.setOperand(0, II->getArgOperand(0));
1897 ICI.setOperand(1, RHS);
1909 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1910 /// We only handle extending casts so far.
1912 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1913 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1914 Value *LHSCIOp = LHSCI->getOperand(0);
1915 Type *SrcTy = LHSCIOp->getType();
1916 Type *DestTy = LHSCI->getType();
1919 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1920 // integer type is the same size as the pointer type.
1921 if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
1922 DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1923 Value *RHSOp = nullptr;
1924 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1925 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1926 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1927 RHSOp = RHSC->getOperand(0);
1928 // If the pointer types don't match, insert a bitcast.
1929 if (LHSCIOp->getType() != RHSOp->getType())
1930 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1934 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1937 // The code below only handles extension cast instructions, so far.
1939 if (LHSCI->getOpcode() != Instruction::ZExt &&
1940 LHSCI->getOpcode() != Instruction::SExt)
1943 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1944 bool isSignedCmp = ICI.isSigned();
1946 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1947 // Not an extension from the same type?
1948 RHSCIOp = CI->getOperand(0);
1949 if (RHSCIOp->getType() != LHSCIOp->getType())
1952 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1953 // and the other is a zext), then we can't handle this.
1954 if (CI->getOpcode() != LHSCI->getOpcode())
1957 // Deal with equality cases early.
1958 if (ICI.isEquality())
1959 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1961 // A signed comparison of sign extended values simplifies into a
1962 // signed comparison.
1963 if (isSignedCmp && isSignedExt)
1964 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1966 // The other three cases all fold into an unsigned comparison.
1967 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1970 // If we aren't dealing with a constant on the RHS, exit early
1971 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1975 // Compute the constant that would happen if we truncated to SrcTy then
1976 // reextended to DestTy.
1977 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1978 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1981 // If the re-extended constant didn't change...
1983 // Deal with equality cases early.
1984 if (ICI.isEquality())
1985 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1987 // A signed comparison of sign extended values simplifies into a
1988 // signed comparison.
1989 if (isSignedExt && isSignedCmp)
1990 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1992 // The other three cases all fold into an unsigned comparison.
1993 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1996 // The re-extended constant changed so the constant cannot be represented
1997 // in the shorter type. Consequently, we cannot emit a simple comparison.
1998 // All the cases that fold to true or false will have already been handled
1999 // by SimplifyICmpInst, so only deal with the tricky case.
2001 if (isSignedCmp || !isSignedExt)
2004 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
2005 // should have been folded away previously and not enter in here.
2007 // We're performing an unsigned comp with a sign extended value.
2008 // This is true if the input is >= 0. [aka >s -1]
2009 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
2010 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
2012 // Finally, return the value computed.
2013 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
2014 return ReplaceInstUsesWith(ICI, Result);
2016 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
2017 return BinaryOperator::CreateNot(Result);
2020 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
2021 /// I = icmp ugt (add (add A, B), CI2), CI1
2022 /// If this is of the form:
2024 /// if (sum+128 >u 255)
2025 /// Then replace it with llvm.sadd.with.overflow.i8.
2027 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
2028 ConstantInt *CI2, ConstantInt *CI1,
2030 // The transformation we're trying to do here is to transform this into an
2031 // llvm.sadd.with.overflow. To do this, we have to replace the original add
2032 // with a narrower add, and discard the add-with-constant that is part of the
2033 // range check (if we can't eliminate it, this isn't profitable).
2035 // In order to eliminate the add-with-constant, the compare can be its only
2037 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
2038 if (!AddWithCst->hasOneUse()) return nullptr;
2040 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
2041 if (!CI2->getValue().isPowerOf2()) return nullptr;
2042 unsigned NewWidth = CI2->getValue().countTrailingZeros();
2043 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
2045 // The width of the new add formed is 1 more than the bias.
2048 // Check to see that CI1 is an all-ones value with NewWidth bits.
2049 if (CI1->getBitWidth() == NewWidth ||
2050 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
2053 // This is only really a signed overflow check if the inputs have been
2054 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
2055 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
2056 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
2057 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
2058 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
2061 // In order to replace the original add with a narrower
2062 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
2063 // and truncates that discard the high bits of the add. Verify that this is
2065 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
2066 for (User *U : OrigAdd->users()) {
2067 if (U == AddWithCst) continue;
2069 // Only accept truncates for now. We would really like a nice recursive
2070 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
2071 // chain to see which bits of a value are actually demanded. If the
2072 // original add had another add which was then immediately truncated, we
2073 // could still do the transformation.
2074 TruncInst *TI = dyn_cast<TruncInst>(U);
2075 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
2079 // If the pattern matches, truncate the inputs to the narrower type and
2080 // use the sadd_with_overflow intrinsic to efficiently compute both the
2081 // result and the overflow bit.
2082 Module *M = I.getParent()->getParent()->getParent();
2084 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
2085 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
2088 InstCombiner::BuilderTy *Builder = IC.Builder;
2090 // Put the new code above the original add, in case there are any uses of the
2091 // add between the add and the compare.
2092 Builder->SetInsertPoint(OrigAdd);
2094 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
2095 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
2096 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
2097 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
2098 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
2100 // The inner add was the result of the narrow add, zero extended to the
2101 // wider type. Replace it with the result computed by the intrinsic.
2102 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
2104 // The original icmp gets replaced with the overflow value.
2105 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
2108 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
2110 // Don't bother doing this transformation for pointers, don't do it for
2112 if (!isa<IntegerType>(OrigAddV->getType())) return nullptr;
2114 // If the add is a constant expr, then we don't bother transforming it.
2115 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
2116 if (!OrigAdd) return nullptr;
2118 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
2120 // Put the new code above the original add, in case there are any uses of the
2121 // add between the add and the compare.
2122 InstCombiner::BuilderTy *Builder = IC.Builder;
2123 Builder->SetInsertPoint(OrigAdd);
2125 Module *M = I.getParent()->getParent()->getParent();
2126 Type *Ty = LHS->getType();
2127 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
2128 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
2129 Value *Add = Builder->CreateExtractValue(Call, 0);
2131 IC.ReplaceInstUsesWith(*OrigAdd, Add);
2133 // The original icmp gets replaced with the overflow value.
2134 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2137 /// \brief Recognize and process idiom involving test for multiplication
2140 /// The caller has matched a pattern of the form:
2141 /// I = cmp u (mul(zext A, zext B), V
2142 /// The function checks if this is a test for overflow and if so replaces
2143 /// multiplication with call to 'mul.with.overflow' intrinsic.
2145 /// \param I Compare instruction.
2146 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
2147 /// the compare instruction. Must be of integer type.
2148 /// \param OtherVal The other argument of compare instruction.
2149 /// \returns Instruction which must replace the compare instruction, NULL if no
2150 /// replacement required.
2151 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
2152 Value *OtherVal, InstCombiner &IC) {
2153 // Don't bother doing this transformation for pointers, don't do it for
2155 if (!isa<IntegerType>(MulVal->getType()))
2158 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
2159 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
2160 Instruction *MulInstr = cast<Instruction>(MulVal);
2161 assert(MulInstr->getOpcode() == Instruction::Mul);
2163 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
2164 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
2165 assert(LHS->getOpcode() == Instruction::ZExt);
2166 assert(RHS->getOpcode() == Instruction::ZExt);
2167 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
2169 // Calculate type and width of the result produced by mul.with.overflow.
2170 Type *TyA = A->getType(), *TyB = B->getType();
2171 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
2172 WidthB = TyB->getPrimitiveSizeInBits();
2175 if (WidthB > WidthA) {
2183 // In order to replace the original mul with a narrower mul.with.overflow,
2184 // all uses must ignore upper bits of the product. The number of used low
2185 // bits must be not greater than the width of mul.with.overflow.
2186 if (MulVal->hasNUsesOrMore(2))
2187 for (User *U : MulVal->users()) {
2190 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2191 // Check if truncation ignores bits above MulWidth.
2192 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
2193 if (TruncWidth > MulWidth)
2195 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2196 // Check if AND ignores bits above MulWidth.
2197 if (BO->getOpcode() != Instruction::And)
2199 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
2200 const APInt &CVal = CI->getValue();
2201 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
2205 // Other uses prohibit this transformation.
2210 // Recognize patterns
2211 switch (I.getPredicate()) {
2212 case ICmpInst::ICMP_EQ:
2213 case ICmpInst::ICMP_NE:
2214 // Recognize pattern:
2215 // mulval = mul(zext A, zext B)
2216 // cmp eq/neq mulval, zext trunc mulval
2217 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
2218 if (Zext->hasOneUse()) {
2219 Value *ZextArg = Zext->getOperand(0);
2220 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
2221 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
2225 // Recognize pattern:
2226 // mulval = mul(zext A, zext B)
2227 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
2230 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
2231 if (ValToMask != MulVal)
2233 const APInt &CVal = CI->getValue() + 1;
2234 if (CVal.isPowerOf2()) {
2235 unsigned MaskWidth = CVal.logBase2();
2236 if (MaskWidth == MulWidth)
2237 break; // Recognized
2242 case ICmpInst::ICMP_UGT:
2243 // Recognize pattern:
2244 // mulval = mul(zext A, zext B)
2245 // cmp ugt mulval, max
2246 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2247 APInt MaxVal = APInt::getMaxValue(MulWidth);
2248 MaxVal = MaxVal.zext(CI->getBitWidth());
2249 if (MaxVal.eq(CI->getValue()))
2250 break; // Recognized
2254 case ICmpInst::ICMP_UGE:
2255 // Recognize pattern:
2256 // mulval = mul(zext A, zext B)
2257 // cmp uge mulval, max+1
2258 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2259 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2260 if (MaxVal.eq(CI->getValue()))
2261 break; // Recognized
2265 case ICmpInst::ICMP_ULE:
2266 // Recognize pattern:
2267 // mulval = mul(zext A, zext B)
2268 // cmp ule mulval, max
2269 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2270 APInt MaxVal = APInt::getMaxValue(MulWidth);
2271 MaxVal = MaxVal.zext(CI->getBitWidth());
2272 if (MaxVal.eq(CI->getValue()))
2273 break; // Recognized
2277 case ICmpInst::ICMP_ULT:
2278 // Recognize pattern:
2279 // mulval = mul(zext A, zext B)
2280 // cmp ule mulval, max + 1
2281 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2282 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2283 if (MaxVal.eq(CI->getValue()))
2284 break; // Recognized
2292 InstCombiner::BuilderTy *Builder = IC.Builder;
2293 Builder->SetInsertPoint(MulInstr);
2294 Module *M = I.getParent()->getParent()->getParent();
2296 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
2297 Value *MulA = A, *MulB = B;
2298 if (WidthA < MulWidth)
2299 MulA = Builder->CreateZExt(A, MulType);
2300 if (WidthB < MulWidth)
2301 MulB = Builder->CreateZExt(B, MulType);
2303 Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
2304 CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
2305 IC.Worklist.Add(MulInstr);
2307 // If there are uses of mul result other than the comparison, we know that
2308 // they are truncation or binary AND. Change them to use result of
2309 // mul.with.overflow and adjust properly mask/size.
2310 if (MulVal->hasNUsesOrMore(2)) {
2311 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
2312 for (User *U : MulVal->users()) {
2313 if (U == &I || U == OtherVal)
2315 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2316 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
2317 IC.ReplaceInstUsesWith(*TI, Mul);
2319 TI->setOperand(0, Mul);
2320 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2321 assert(BO->getOpcode() == Instruction::And);
2322 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
2323 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
2324 APInt ShortMask = CI->getValue().trunc(MulWidth);
2325 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
2327 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
2328 IC.Worklist.Add(Zext);
2329 IC.ReplaceInstUsesWith(*BO, Zext);
2331 llvm_unreachable("Unexpected Binary operation");
2333 IC.Worklist.Add(cast<Instruction>(U));
2336 if (isa<Instruction>(OtherVal))
2337 IC.Worklist.Add(cast<Instruction>(OtherVal));
2339 // The original icmp gets replaced with the overflow value, maybe inverted
2340 // depending on predicate.
2341 bool Inverse = false;
2342 switch (I.getPredicate()) {
2343 case ICmpInst::ICMP_NE:
2345 case ICmpInst::ICMP_EQ:
2348 case ICmpInst::ICMP_UGT:
2349 case ICmpInst::ICMP_UGE:
2350 if (I.getOperand(0) == MulVal)
2354 case ICmpInst::ICMP_ULT:
2355 case ICmpInst::ICMP_ULE:
2356 if (I.getOperand(1) == MulVal)
2361 llvm_unreachable("Unexpected predicate");
2364 Value *Res = Builder->CreateExtractValue(Call, 1);
2365 return BinaryOperator::CreateNot(Res);
2368 return ExtractValueInst::Create(Call, 1);
2371 // DemandedBitsLHSMask - When performing a comparison against a constant,
2372 // it is possible that not all the bits in the LHS are demanded. This helper
2373 // method computes the mask that IS demanded.
2374 static APInt DemandedBitsLHSMask(ICmpInst &I,
2375 unsigned BitWidth, bool isSignCheck) {
2377 return APInt::getSignBit(BitWidth);
2379 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2380 if (!CI) return APInt::getAllOnesValue(BitWidth);
2381 const APInt &RHS = CI->getValue();
2383 switch (I.getPredicate()) {
2384 // For a UGT comparison, we don't care about any bits that
2385 // correspond to the trailing ones of the comparand. The value of these
2386 // bits doesn't impact the outcome of the comparison, because any value
2387 // greater than the RHS must differ in a bit higher than these due to carry.
2388 case ICmpInst::ICMP_UGT: {
2389 unsigned trailingOnes = RHS.countTrailingOnes();
2390 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2394 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2395 // Any value less than the RHS must differ in a higher bit because of carries.
2396 case ICmpInst::ICMP_ULT: {
2397 unsigned trailingZeros = RHS.countTrailingZeros();
2398 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2403 return APInt::getAllOnesValue(BitWidth);
2408 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2409 /// should be swapped.
2410 /// The decision is based on how many times these two operands are reused
2411 /// as subtract operands and their positions in those instructions.
2412 /// The rational is that several architectures use the same instruction for
2413 /// both subtract and cmp, thus it is better if the order of those operands
2415 /// \return true if Op0 and Op1 should be swapped.
2416 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2417 const Value * Op1) {
2418 // Filter out pointer value as those cannot appears directly in subtract.
2419 // FIXME: we may want to go through inttoptrs or bitcasts.
2420 if (Op0->getType()->isPointerTy())
2422 // Count every uses of both Op0 and Op1 in a subtract.
2423 // Each time Op0 is the first operand, count -1: swapping is bad, the
2424 // subtract has already the same layout as the compare.
2425 // Each time Op0 is the second operand, count +1: swapping is good, the
2426 // subtract has a different layout as the compare.
2427 // At the end, if the benefit is greater than 0, Op0 should come second to
2428 // expose more CSE opportunities.
2429 int GlobalSwapBenefits = 0;
2430 for (const User *U : Op0->users()) {
2431 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
2432 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2434 // If Op0 is the first argument, this is not beneficial to swap the
2436 int LocalSwapBenefits = -1;
2437 unsigned Op1Idx = 1;
2438 if (BinOp->getOperand(Op1Idx) == Op0) {
2440 LocalSwapBenefits = 1;
2442 if (BinOp->getOperand(Op1Idx) != Op1)
2444 GlobalSwapBenefits += LocalSwapBenefits;
2446 return GlobalSwapBenefits > 0;
2449 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2450 bool Changed = false;
2451 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2452 unsigned Op0Cplxity = getComplexity(Op0);
2453 unsigned Op1Cplxity = getComplexity(Op1);
2455 /// Orders the operands of the compare so that they are listed from most
2456 /// complex to least complex. This puts constants before unary operators,
2457 /// before binary operators.
2458 if (Op0Cplxity < Op1Cplxity ||
2459 (Op0Cplxity == Op1Cplxity &&
2460 swapMayExposeCSEOpportunities(Op0, Op1))) {
2462 std::swap(Op0, Op1);
2466 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
2467 return ReplaceInstUsesWith(I, V);
2469 // comparing -val or val with non-zero is the same as just comparing val
2470 // ie, abs(val) != 0 -> val != 0
2471 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2473 Value *Cond, *SelectTrue, *SelectFalse;
2474 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2475 m_Value(SelectFalse)))) {
2476 if (Value *V = dyn_castNegVal(SelectTrue)) {
2477 if (V == SelectFalse)
2478 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2480 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2481 if (V == SelectTrue)
2482 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2487 Type *Ty = Op0->getType();
2489 // icmp's with boolean values can always be turned into bitwise operations
2490 if (Ty->isIntegerTy(1)) {
2491 switch (I.getPredicate()) {
2492 default: llvm_unreachable("Invalid icmp instruction!");
2493 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2494 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2495 return BinaryOperator::CreateNot(Xor);
2497 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2498 return BinaryOperator::CreateXor(Op0, Op1);
2500 case ICmpInst::ICMP_UGT:
2501 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2503 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2504 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2505 return BinaryOperator::CreateAnd(Not, Op1);
2507 case ICmpInst::ICMP_SGT:
2508 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2510 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2511 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2512 return BinaryOperator::CreateAnd(Not, Op0);
2514 case ICmpInst::ICMP_UGE:
2515 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2517 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2518 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2519 return BinaryOperator::CreateOr(Not, Op1);
2521 case ICmpInst::ICMP_SGE:
2522 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2524 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2525 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2526 return BinaryOperator::CreateOr(Not, Op0);
2531 unsigned BitWidth = 0;
2532 if (Ty->isIntOrIntVectorTy())
2533 BitWidth = Ty->getScalarSizeInBits();
2534 else if (DL) // Pointers require DL info to get their size.
2535 BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
2537 bool isSignBit = false;
2539 // See if we are doing a comparison with a constant.
2540 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2541 Value *A = nullptr, *B = nullptr;
2543 // Match the following pattern, which is a common idiom when writing
2544 // overflow-safe integer arithmetic function. The source performs an
2545 // addition in wider type, and explicitly checks for overflow using
2546 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2547 // sadd_with_overflow intrinsic.
2549 // TODO: This could probably be generalized to handle other overflow-safe
2550 // operations if we worked out the formulas to compute the appropriate
2554 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2556 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2557 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2558 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2559 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2563 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2564 if (I.isEquality() && CI->isZero() &&
2565 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2566 // (icmp cond A B) if cond is equality
2567 return new ICmpInst(I.getPredicate(), A, B);
2570 // If we have an icmp le or icmp ge instruction, turn it into the
2571 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2572 // them being folded in the code below. The SimplifyICmpInst code has
2573 // already handled the edge cases for us, so we just assert on them.
2574 switch (I.getPredicate()) {
2576 case ICmpInst::ICMP_ULE:
2577 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2578 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2579 Builder->getInt(CI->getValue()+1));
2580 case ICmpInst::ICMP_SLE:
2581 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2582 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2583 Builder->getInt(CI->getValue()+1));
2584 case ICmpInst::ICMP_UGE:
2585 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2586 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2587 Builder->getInt(CI->getValue()-1));
2588 case ICmpInst::ICMP_SGE:
2589 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2590 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2591 Builder->getInt(CI->getValue()-1));
2594 if (I.isEquality()) {
2596 if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
2597 match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
2598 // (icmp eq/ne (ashr/lshr const2, A), const1)
2599 if (Instruction *Inst = FoldICmpCstShrCst(I, Op0, A, CI, CI2))
2602 if (match(Op0, m_Shl(m_ConstantInt(CI2), m_Value(A)))) {
2603 // (icmp eq/ne (shl const2, A), const1)
2604 if (Instruction *Inst = FoldICmpCstShlCst(I, Op0, A, CI, CI2))
2609 // If this comparison is a normal comparison, it demands all
2610 // bits, if it is a sign bit comparison, it only demands the sign bit.
2612 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2615 // See if we can fold the comparison based on range information we can get
2616 // by checking whether bits are known to be zero or one in the input.
2617 if (BitWidth != 0) {
2618 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2619 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2621 if (SimplifyDemandedBits(I.getOperandUse(0),
2622 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2623 Op0KnownZero, Op0KnownOne, 0))
2625 if (SimplifyDemandedBits(I.getOperandUse(1),
2626 APInt::getAllOnesValue(BitWidth),
2627 Op1KnownZero, Op1KnownOne, 0))
2630 // Given the known and unknown bits, compute a range that the LHS could be
2631 // in. Compute the Min, Max and RHS values based on the known bits. For the
2632 // EQ and NE we use unsigned values.
2633 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2634 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2636 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2638 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2641 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2643 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2647 // If Min and Max are known to be the same, then SimplifyDemandedBits
2648 // figured out that the LHS is a constant. Just constant fold this now so
2649 // that code below can assume that Min != Max.
2650 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2651 return new ICmpInst(I.getPredicate(),
2652 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2653 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2654 return new ICmpInst(I.getPredicate(), Op0,
2655 ConstantInt::get(Op1->getType(), Op1Min));
2657 // Based on the range information we know about the LHS, see if we can
2658 // simplify this comparison. For example, (x&4) < 8 is always true.
2659 switch (I.getPredicate()) {
2660 default: llvm_unreachable("Unknown icmp opcode!");
2661 case ICmpInst::ICMP_EQ: {
2662 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2663 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2665 // If all bits are known zero except for one, then we know at most one
2666 // bit is set. If the comparison is against zero, then this is a check
2667 // to see if *that* bit is set.
2668 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2669 if (~Op1KnownZero == 0) {
2670 // If the LHS is an AND with the same constant, look through it.
2671 Value *LHS = nullptr;
2672 ConstantInt *LHSC = nullptr;
2673 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2674 LHSC->getValue() != Op0KnownZeroInverted)
2677 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2678 // then turn "((1 << x)&8) == 0" into "x != 3".
2679 // or turn "((1 << x)&7) == 0" into "x > 2".
2681 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2682 APInt ValToCheck = Op0KnownZeroInverted;
2683 if (ValToCheck.isPowerOf2()) {
2684 unsigned CmpVal = ValToCheck.countTrailingZeros();
2685 return new ICmpInst(ICmpInst::ICMP_NE, X,
2686 ConstantInt::get(X->getType(), CmpVal));
2687 } else if ((++ValToCheck).isPowerOf2()) {
2688 unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
2689 return new ICmpInst(ICmpInst::ICMP_UGT, X,
2690 ConstantInt::get(X->getType(), CmpVal));
2694 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2695 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2697 if (Op0KnownZeroInverted == 1 &&
2698 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2699 return new ICmpInst(ICmpInst::ICMP_NE, X,
2700 ConstantInt::get(X->getType(),
2701 CI->countTrailingZeros()));
2706 case ICmpInst::ICMP_NE: {
2707 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2708 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2710 // If all bits are known zero except for one, then we know at most one
2711 // bit is set. If the comparison is against zero, then this is a check
2712 // to see if *that* bit is set.
2713 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2714 if (~Op1KnownZero == 0) {
2715 // If the LHS is an AND with the same constant, look through it.
2716 Value *LHS = nullptr;
2717 ConstantInt *LHSC = nullptr;
2718 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2719 LHSC->getValue() != Op0KnownZeroInverted)
2722 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2723 // then turn "((1 << x)&8) != 0" into "x == 3".
2724 // or turn "((1 << x)&7) != 0" into "x < 3".
2726 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2727 APInt ValToCheck = Op0KnownZeroInverted;
2728 if (ValToCheck.isPowerOf2()) {
2729 unsigned CmpVal = ValToCheck.countTrailingZeros();
2730 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2731 ConstantInt::get(X->getType(), CmpVal));
2732 } else if ((++ValToCheck).isPowerOf2()) {
2733 unsigned CmpVal = ValToCheck.countTrailingZeros();
2734 return new ICmpInst(ICmpInst::ICMP_ULT, X,
2735 ConstantInt::get(X->getType(), CmpVal));
2739 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2740 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2742 if (Op0KnownZeroInverted == 1 &&
2743 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2744 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2745 ConstantInt::get(X->getType(),
2746 CI->countTrailingZeros()));
2751 case ICmpInst::ICMP_ULT:
2752 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2753 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2754 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2755 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2756 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2757 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2758 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2759 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2760 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2761 Builder->getInt(CI->getValue()-1));
2763 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2764 if (CI->isMinValue(true))
2765 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2766 Constant::getAllOnesValue(Op0->getType()));
2769 case ICmpInst::ICMP_UGT:
2770 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2771 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2772 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2773 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2775 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2776 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2777 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2778 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2779 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2780 Builder->getInt(CI->getValue()+1));
2782 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2783 if (CI->isMaxValue(true))
2784 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2785 Constant::getNullValue(Op0->getType()));
2788 case ICmpInst::ICMP_SLT:
2789 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2790 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2791 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2792 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2793 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2794 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2795 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2796 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2797 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2798 Builder->getInt(CI->getValue()-1));
2801 case ICmpInst::ICMP_SGT:
2802 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2803 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2804 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2805 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2807 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2808 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2809 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2810 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2811 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2812 Builder->getInt(CI->getValue()+1));
2815 case ICmpInst::ICMP_SGE:
2816 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2817 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2818 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2819 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2820 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2822 case ICmpInst::ICMP_SLE:
2823 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2824 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2825 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2826 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2827 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2829 case ICmpInst::ICMP_UGE:
2830 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2831 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2832 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2833 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2834 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2836 case ICmpInst::ICMP_ULE:
2837 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2838 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2839 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2840 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2841 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2845 // Turn a signed comparison into an unsigned one if both operands
2846 // are known to have the same sign.
2848 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2849 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2850 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2853 // Test if the ICmpInst instruction is used exclusively by a select as
2854 // part of a minimum or maximum operation. If so, refrain from doing
2855 // any other folding. This helps out other analyses which understand
2856 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2857 // and CodeGen. And in this case, at least one of the comparison
2858 // operands has at least one user besides the compare (the select),
2859 // which would often largely negate the benefit of folding anyway.
2861 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
2862 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2863 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2866 // See if we are doing a comparison between a constant and an instruction that
2867 // can be folded into the comparison.
2868 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2869 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2870 // instruction, see if that instruction also has constants so that the
2871 // instruction can be folded into the icmp
2872 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2873 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2877 // Handle icmp with constant (but not simple integer constant) RHS
2878 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2879 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2880 switch (LHSI->getOpcode()) {
2881 case Instruction::GetElementPtr:
2882 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2883 if (RHSC->isNullValue() &&
2884 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2885 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2886 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2888 case Instruction::PHI:
2889 // Only fold icmp into the PHI if the phi and icmp are in the same
2890 // block. If in the same block, we're encouraging jump threading. If
2891 // not, we are just pessimizing the code by making an i1 phi.
2892 if (LHSI->getParent() == I.getParent())
2893 if (Instruction *NV = FoldOpIntoPhi(I))
2896 case Instruction::Select: {
2897 // If either operand of the select is a constant, we can fold the
2898 // comparison into the select arms, which will cause one to be
2899 // constant folded and the select turned into a bitwise or.
2900 Value *Op1 = nullptr, *Op2 = nullptr;
2901 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2902 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2903 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2904 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2906 // We only want to perform this transformation if it will not lead to
2907 // additional code. This is true if either both sides of the select
2908 // fold to a constant (in which case the icmp is replaced with a select
2909 // which will usually simplify) or this is the only user of the
2910 // select (in which case we are trading a select+icmp for a simpler
2912 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2914 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2917 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2919 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2923 case Instruction::IntToPtr:
2924 // icmp pred inttoptr(X), null -> icmp pred X, 0
2925 if (RHSC->isNullValue() && DL &&
2926 DL->getIntPtrType(RHSC->getType()) ==
2927 LHSI->getOperand(0)->getType())
2928 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2929 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2932 case Instruction::Load:
2933 // Try to optimize things like "A[i] > 4" to index computations.
2934 if (GetElementPtrInst *GEP =
2935 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2936 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2937 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2938 !cast<LoadInst>(LHSI)->isVolatile())
2939 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2946 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2947 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2948 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2950 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2951 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2952 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2955 // Test to see if the operands of the icmp are casted versions of other
2956 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2958 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2959 if (Op0->getType()->isPointerTy() &&
2960 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2961 // We keep moving the cast from the left operand over to the right
2962 // operand, where it can often be eliminated completely.
2963 Op0 = CI->getOperand(0);
2965 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2966 // so eliminate it as well.
2967 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2968 Op1 = CI2->getOperand(0);
2970 // If Op1 is a constant, we can fold the cast into the constant.
2971 if (Op0->getType() != Op1->getType()) {
2972 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2973 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2975 // Otherwise, cast the RHS right before the icmp
2976 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2979 return new ICmpInst(I.getPredicate(), Op0, Op1);
2983 if (isa<CastInst>(Op0)) {
2984 // Handle the special case of: icmp (cast bool to X), <cst>
2985 // This comes up when you have code like
2988 // For generality, we handle any zero-extension of any operand comparison
2989 // with a constant or another cast from the same type.
2990 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2991 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2995 // Special logic for binary operators.
2996 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2997 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2999 CmpInst::Predicate Pred = I.getPredicate();
3000 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3001 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3002 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
3003 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3004 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3005 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3006 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
3007 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3008 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3010 // Analyze the case when either Op0 or Op1 is an add instruction.
3011 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3012 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3013 if (BO0 && BO0->getOpcode() == Instruction::Add)
3014 A = BO0->getOperand(0), B = BO0->getOperand(1);
3015 if (BO1 && BO1->getOpcode() == Instruction::Add)
3016 C = BO1->getOperand(0), D = BO1->getOperand(1);
3018 // icmp (X+cst) < 0 --> X < -cst
3019 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred) && match(Op1, m_Zero()))
3020 if (ConstantInt *RHSC = dyn_cast_or_null<ConstantInt>(B))
3021 if (!RHSC->isMinValue(/*isSigned=*/true))
3022 return new ICmpInst(Pred, A, ConstantExpr::getNeg(RHSC));
3024 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3025 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3026 return new ICmpInst(Pred, A == Op1 ? B : A,
3027 Constant::getNullValue(Op1->getType()));
3029 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3030 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3031 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3034 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3035 if (A && C && (A == C || A == D || B == C || B == D) &&
3036 NoOp0WrapProblem && NoOp1WrapProblem &&
3037 // Try not to increase register pressure.
3038 BO0->hasOneUse() && BO1->hasOneUse()) {
3039 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3042 // C + B == C + D -> B == D
3045 } else if (A == D) {
3046 // D + B == C + D -> B == C
3049 } else if (B == C) {
3050 // A + C == C + D -> A == D
3055 // A + D == C + D -> A == C
3059 return new ICmpInst(Pred, Y, Z);
3062 // icmp slt (X + -1), Y -> icmp sle X, Y
3063 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3064 match(B, m_AllOnes()))
3065 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3067 // icmp sge (X + -1), Y -> icmp sgt X, Y
3068 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3069 match(B, m_AllOnes()))
3070 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3072 // icmp sle (X + 1), Y -> icmp slt X, Y
3073 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
3075 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3077 // icmp sgt (X + 1), Y -> icmp sge X, Y
3078 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&