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/ADT/Statistic.h"
16 #include "llvm/Analysis/ConstantFolding.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/MemoryBuiltins.h"
19 #include "llvm/IR/ConstantRange.h"
20 #include "llvm/IR/DataLayout.h"
21 #include "llvm/IR/GetElementPtrTypeIterator.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/PatternMatch.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Target/TargetLibraryInfo.h"
29 using namespace PatternMatch;
31 #define DEBUG_TYPE "instcombine"
33 // How many times is a select replaced by one of its operands?
34 STATISTIC(NumSel, "Number of select opts");
36 // Initialization Routines
38 static ConstantInt *getOne(Constant *C) {
39 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
42 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
43 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
46 static bool HasAddOverflow(ConstantInt *Result,
47 ConstantInt *In1, ConstantInt *In2,
50 return Result->getValue().ult(In1->getValue());
52 if (In2->isNegative())
53 return Result->getValue().sgt(In1->getValue());
54 return Result->getValue().slt(In1->getValue());
57 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
58 /// overflowed for this type.
59 static bool AddWithOverflow(Constant *&Result, Constant *In1,
60 Constant *In2, bool IsSigned = false) {
61 Result = ConstantExpr::getAdd(In1, In2);
63 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
64 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
65 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
66 if (HasAddOverflow(ExtractElement(Result, Idx),
67 ExtractElement(In1, Idx),
68 ExtractElement(In2, Idx),
75 return HasAddOverflow(cast<ConstantInt>(Result),
76 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
80 static bool HasSubOverflow(ConstantInt *Result,
81 ConstantInt *In1, ConstantInt *In2,
84 return Result->getValue().ugt(In1->getValue());
86 if (In2->isNegative())
87 return Result->getValue().slt(In1->getValue());
89 return Result->getValue().sgt(In1->getValue());
92 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
93 /// overflowed for this type.
94 static bool SubWithOverflow(Constant *&Result, Constant *In1,
95 Constant *In2, bool IsSigned = false) {
96 Result = ConstantExpr::getSub(In1, In2);
98 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
99 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
100 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
101 if (HasSubOverflow(ExtractElement(Result, Idx),
102 ExtractElement(In1, Idx),
103 ExtractElement(In2, Idx),
110 return HasSubOverflow(cast<ConstantInt>(Result),
111 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
115 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
116 /// comparison only checks the sign bit. If it only checks the sign bit, set
117 /// TrueIfSigned if the result of the comparison is true when the input value is
119 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
120 bool &TrueIfSigned) {
122 case ICmpInst::ICMP_SLT: // True if LHS s< 0
124 return RHS->isZero();
125 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
127 return RHS->isAllOnesValue();
128 case ICmpInst::ICMP_SGT: // True if LHS s> -1
129 TrueIfSigned = false;
130 return RHS->isAllOnesValue();
131 case ICmpInst::ICMP_UGT:
132 // True if LHS u> RHS and RHS == high-bit-mask - 1
134 return RHS->isMaxValue(true);
135 case ICmpInst::ICMP_UGE:
136 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
138 return RHS->getValue().isSignBit();
144 /// Returns true if the exploded icmp can be expressed as a signed comparison
145 /// to zero and updates the predicate accordingly.
146 /// The signedness of the comparison is preserved.
147 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
148 if (!ICmpInst::isSigned(pred))
152 return ICmpInst::isRelational(pred);
155 if (pred == ICmpInst::ICMP_SLT) {
156 pred = ICmpInst::ICMP_SLE;
159 } else if (RHS->isAllOnesValue()) {
160 if (pred == ICmpInst::ICMP_SGT) {
161 pred = ICmpInst::ICMP_SGE;
169 // isHighOnes - Return true if the constant is of the form 1+0+.
170 // This is the same as lowones(~X).
171 static bool isHighOnes(const ConstantInt *CI) {
172 return (~CI->getValue() + 1).isPowerOf2();
175 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
176 /// set of known zero and one bits, compute the maximum and minimum values that
177 /// could have the specified known zero and known one bits, returning them in
179 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
180 const APInt& KnownOne,
181 APInt& Min, APInt& Max) {
182 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
183 KnownZero.getBitWidth() == Min.getBitWidth() &&
184 KnownZero.getBitWidth() == Max.getBitWidth() &&
185 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
186 APInt UnknownBits = ~(KnownZero|KnownOne);
188 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
189 // bit if it is unknown.
191 Max = KnownOne|UnknownBits;
193 if (UnknownBits.isNegative()) { // Sign bit is unknown
194 Min.setBit(Min.getBitWidth()-1);
195 Max.clearBit(Max.getBitWidth()-1);
199 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
200 // a set of known zero and one bits, compute the maximum and minimum values that
201 // could have the specified known zero and known one bits, returning them in
203 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
204 const APInt &KnownOne,
205 APInt &Min, APInt &Max) {
206 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
207 KnownZero.getBitWidth() == Min.getBitWidth() &&
208 KnownZero.getBitWidth() == Max.getBitWidth() &&
209 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
210 APInt UnknownBits = ~(KnownZero|KnownOne);
212 // The minimum value is when the unknown bits are all zeros.
214 // The maximum value is when the unknown bits are all ones.
215 Max = KnownOne|UnknownBits;
220 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
221 /// cmp pred (load (gep GV, ...)), cmpcst
222 /// where GV is a global variable with a constant initializer. Try to simplify
223 /// this into some simple computation that does not need the load. For example
224 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
226 /// If AndCst is non-null, then the loaded value is masked with that constant
227 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
228 Instruction *InstCombiner::
229 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
230 CmpInst &ICI, ConstantInt *AndCst) {
231 // We need TD information to know the pointer size unless this is inbounds.
232 if (!GEP->isInBounds() && !DL)
235 Constant *Init = GV->getInitializer();
236 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
239 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
240 if (ArrayElementCount > 1024) return nullptr; // Don't blow up on huge arrays.
242 // There are many forms of this optimization we can handle, for now, just do
243 // the simple index into a single-dimensional array.
245 // Require: GEP GV, 0, i {{, constant indices}}
246 if (GEP->getNumOperands() < 3 ||
247 !isa<ConstantInt>(GEP->getOperand(1)) ||
248 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
249 isa<Constant>(GEP->getOperand(2)))
252 // Check that indices after the variable are constants and in-range for the
253 // type they index. Collect the indices. This is typically for arrays of
255 SmallVector<unsigned, 4> LaterIndices;
257 Type *EltTy = Init->getType()->getArrayElementType();
258 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
259 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
260 if (!Idx) return nullptr; // Variable index.
262 uint64_t IdxVal = Idx->getZExtValue();
263 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
265 if (StructType *STy = dyn_cast<StructType>(EltTy))
266 EltTy = STy->getElementType(IdxVal);
267 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
268 if (IdxVal >= ATy->getNumElements()) return nullptr;
269 EltTy = ATy->getElementType();
271 return nullptr; // Unknown type.
274 LaterIndices.push_back(IdxVal);
277 enum { Overdefined = -3, Undefined = -2 };
279 // Variables for our state machines.
281 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
282 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
283 // and 87 is the second (and last) index. FirstTrueElement is -2 when
284 // undefined, otherwise set to the first true element. SecondTrueElement is
285 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
286 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
288 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
289 // form "i != 47 & i != 87". Same state transitions as for true elements.
290 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
292 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
293 /// define a state machine that triggers for ranges of values that the index
294 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
295 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
296 /// index in the range (inclusive). We use -2 for undefined here because we
297 /// use relative comparisons and don't want 0-1 to match -1.
298 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
300 // MagicBitvector - This is a magic bitvector where we set a bit if the
301 // comparison is true for element 'i'. If there are 64 elements or less in
302 // the array, this will fully represent all the comparison results.
303 uint64_t MagicBitvector = 0;
306 // Scan the array and see if one of our patterns matches.
307 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
308 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
309 Constant *Elt = Init->getAggregateElement(i);
310 if (!Elt) return nullptr;
312 // If this is indexing an array of structures, get the structure element.
313 if (!LaterIndices.empty())
314 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
316 // If the element is masked, handle it.
317 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
319 // Find out if the comparison would be true or false for the i'th element.
320 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
321 CompareRHS, DL, TLI);
322 // If the result is undef for this element, ignore it.
323 if (isa<UndefValue>(C)) {
324 // Extend range state machines to cover this element in case there is an
325 // undef in the middle of the range.
326 if (TrueRangeEnd == (int)i-1)
328 if (FalseRangeEnd == (int)i-1)
333 // If we can't compute the result for any of the elements, we have to give
334 // up evaluating the entire conditional.
335 if (!isa<ConstantInt>(C)) return nullptr;
337 // Otherwise, we know if the comparison is true or false for this element,
338 // update our state machines.
339 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
341 // State machine for single/double/range index comparison.
343 // Update the TrueElement state machine.
344 if (FirstTrueElement == Undefined)
345 FirstTrueElement = TrueRangeEnd = i; // First true element.
347 // Update double-compare state machine.
348 if (SecondTrueElement == Undefined)
349 SecondTrueElement = i;
351 SecondTrueElement = Overdefined;
353 // Update range state machine.
354 if (TrueRangeEnd == (int)i-1)
357 TrueRangeEnd = Overdefined;
360 // Update the FalseElement state machine.
361 if (FirstFalseElement == Undefined)
362 FirstFalseElement = FalseRangeEnd = i; // First false element.
364 // Update double-compare state machine.
365 if (SecondFalseElement == Undefined)
366 SecondFalseElement = i;
368 SecondFalseElement = Overdefined;
370 // Update range state machine.
371 if (FalseRangeEnd == (int)i-1)
374 FalseRangeEnd = Overdefined;
379 // If this element is in range, update our magic bitvector.
380 if (i < 64 && IsTrueForElt)
381 MagicBitvector |= 1ULL << i;
383 // If all of our states become overdefined, bail out early. Since the
384 // predicate is expensive, only check it every 8 elements. This is only
385 // really useful for really huge arrays.
386 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
387 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
388 FalseRangeEnd == Overdefined)
392 // Now that we've scanned the entire array, emit our new comparison(s). We
393 // order the state machines in complexity of the generated code.
394 Value *Idx = GEP->getOperand(2);
396 // If the index is larger than the pointer size of the target, truncate the
397 // index down like the GEP would do implicitly. We don't have to do this for
398 // an inbounds GEP because the index can't be out of range.
399 if (!GEP->isInBounds()) {
400 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
401 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
402 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
403 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
406 // If the comparison is only true for one or two elements, emit direct
408 if (SecondTrueElement != Overdefined) {
409 // None true -> false.
410 if (FirstTrueElement == Undefined)
411 return ReplaceInstUsesWith(ICI, Builder->getFalse());
413 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
415 // True for one element -> 'i == 47'.
416 if (SecondTrueElement == Undefined)
417 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
419 // True for two elements -> 'i == 47 | i == 72'.
420 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
421 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
422 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
423 return BinaryOperator::CreateOr(C1, C2);
426 // If the comparison is only false for one or two elements, emit direct
428 if (SecondFalseElement != Overdefined) {
429 // None false -> true.
430 if (FirstFalseElement == Undefined)
431 return ReplaceInstUsesWith(ICI, Builder->getTrue());
433 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
435 // False for one element -> 'i != 47'.
436 if (SecondFalseElement == Undefined)
437 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
439 // False for two elements -> 'i != 47 & i != 72'.
440 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
441 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
442 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
443 return BinaryOperator::CreateAnd(C1, C2);
446 // If the comparison can be replaced with a range comparison for the elements
447 // where it is true, emit the range check.
448 if (TrueRangeEnd != Overdefined) {
449 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
451 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
452 if (FirstTrueElement) {
453 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
454 Idx = Builder->CreateAdd(Idx, Offs);
457 Value *End = ConstantInt::get(Idx->getType(),
458 TrueRangeEnd-FirstTrueElement+1);
459 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
462 // False range check.
463 if (FalseRangeEnd != Overdefined) {
464 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
465 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
466 if (FirstFalseElement) {
467 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
468 Idx = Builder->CreateAdd(Idx, Offs);
471 Value *End = ConstantInt::get(Idx->getType(),
472 FalseRangeEnd-FirstFalseElement);
473 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
477 // If a magic bitvector captures the entire comparison state
478 // of this load, replace it with computation that does:
479 // ((magic_cst >> i) & 1) != 0
483 // Look for an appropriate type:
484 // - The type of Idx if the magic fits
485 // - The smallest fitting legal type if we have a DataLayout
487 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
490 Ty = DL->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
491 else if (ArrayElementCount <= 32)
492 Ty = Type::getInt32Ty(Init->getContext());
495 Value *V = Builder->CreateIntCast(Idx, Ty, false);
496 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
497 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
498 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
506 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
507 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
508 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
509 /// be complex, and scales are involved. The above expression would also be
510 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
511 /// This later form is less amenable to optimization though, and we are allowed
512 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
514 /// If we can't emit an optimized form for this expression, this returns null.
516 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
517 const DataLayout &DL = *IC.getDataLayout();
518 gep_type_iterator GTI = gep_type_begin(GEP);
520 // Check to see if this gep only has a single variable index. If so, and if
521 // any constant indices are a multiple of its scale, then we can compute this
522 // in terms of the scale of the variable index. For example, if the GEP
523 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
524 // because the expression will cross zero at the same point.
525 unsigned i, e = GEP->getNumOperands();
527 for (i = 1; i != e; ++i, ++GTI) {
528 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
529 // Compute the aggregate offset of constant indices.
530 if (CI->isZero()) continue;
532 // Handle a struct index, which adds its field offset to the pointer.
533 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
534 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
536 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
537 Offset += Size*CI->getSExtValue();
540 // Found our variable index.
545 // If there are no variable indices, we must have a constant offset, just
546 // evaluate it the general way.
547 if (i == e) return nullptr;
549 Value *VariableIdx = GEP->getOperand(i);
550 // Determine the scale factor of the variable element. For example, this is
551 // 4 if the variable index is into an array of i32.
552 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
554 // Verify that there are no other variable indices. If so, emit the hard way.
555 for (++i, ++GTI; i != e; ++i, ++GTI) {
556 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
557 if (!CI) return nullptr;
559 // Compute the aggregate offset of constant indices.
560 if (CI->isZero()) continue;
562 // Handle a struct index, which adds its field offset to the pointer.
563 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
564 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
566 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
567 Offset += Size*CI->getSExtValue();
573 // Okay, we know we have a single variable index, which must be a
574 // pointer/array/vector index. If there is no offset, life is simple, return
576 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
577 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
579 // Cast to intptrty in case a truncation occurs. If an extension is needed,
580 // we don't need to bother extending: the extension won't affect where the
581 // computation crosses zero.
582 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
583 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
588 // Otherwise, there is an index. The computation we will do will be modulo
589 // the pointer size, so get it.
590 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
592 Offset &= PtrSizeMask;
593 VariableScale &= PtrSizeMask;
595 // To do this transformation, any constant index must be a multiple of the
596 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
597 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
598 // multiple of the variable scale.
599 int64_t NewOffs = Offset / (int64_t)VariableScale;
600 if (Offset != NewOffs*(int64_t)VariableScale)
603 // Okay, we can do this evaluation. Start by converting the index to intptr.
604 if (VariableIdx->getType() != IntPtrTy)
605 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
607 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
608 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
611 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
612 /// else. At this point we know that the GEP is on the LHS of the comparison.
613 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
614 ICmpInst::Predicate Cond,
616 // Don't transform signed compares of GEPs into index compares. Even if the
617 // GEP is inbounds, the final add of the base pointer can have signed overflow
618 // and would change the result of the icmp.
619 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
620 // the maximum signed value for the pointer type.
621 if (ICmpInst::isSigned(Cond))
624 // Look through bitcasts and addrspacecasts. We do not however want to remove
626 if (!isa<GetElementPtrInst>(RHS))
627 RHS = RHS->stripPointerCasts();
629 Value *PtrBase = GEPLHS->getOperand(0);
630 if (DL && PtrBase == RHS && GEPLHS->isInBounds()) {
631 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
632 // This transformation (ignoring the base and scales) is valid because we
633 // know pointers can't overflow since the gep is inbounds. See if we can
634 // output an optimized form.
635 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
637 // If not, synthesize the offset the hard way.
639 Offset = EmitGEPOffset(GEPLHS);
640 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
641 Constant::getNullValue(Offset->getType()));
642 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
643 // If the base pointers are different, but the indices are the same, just
644 // compare the base pointer.
645 if (PtrBase != GEPRHS->getOperand(0)) {
646 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
647 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
648 GEPRHS->getOperand(0)->getType();
650 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
651 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
652 IndicesTheSame = false;
656 // If all indices are the same, just compare the base pointers.
658 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
660 // If we're comparing GEPs with two base pointers that only differ in type
661 // and both GEPs have only constant indices or just one use, then fold
662 // the compare with the adjusted indices.
663 if (DL && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
664 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
665 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
666 PtrBase->stripPointerCasts() ==
667 GEPRHS->getOperand(0)->stripPointerCasts()) {
668 Value *LOffset = EmitGEPOffset(GEPLHS);
669 Value *ROffset = EmitGEPOffset(GEPRHS);
671 // If we looked through an addrspacecast between different sized address
672 // spaces, the LHS and RHS pointers are different sized
673 // integers. Truncate to the smaller one.
674 Type *LHSIndexTy = LOffset->getType();
675 Type *RHSIndexTy = ROffset->getType();
676 if (LHSIndexTy != RHSIndexTy) {
677 if (LHSIndexTy->getPrimitiveSizeInBits() <
678 RHSIndexTy->getPrimitiveSizeInBits()) {
679 ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy);
681 LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy);
684 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
686 return ReplaceInstUsesWith(I, Cmp);
689 // Otherwise, the base pointers are different and the indices are
690 // different, bail out.
694 // If one of the GEPs has all zero indices, recurse.
695 if (GEPLHS->hasAllZeroIndices())
696 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
697 ICmpInst::getSwappedPredicate(Cond), I);
699 // If the other GEP has all zero indices, recurse.
700 if (GEPRHS->hasAllZeroIndices())
701 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
703 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
704 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
705 // If the GEPs only differ by one index, compare it.
706 unsigned NumDifferences = 0; // Keep track of # differences.
707 unsigned DiffOperand = 0; // The operand that differs.
708 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
709 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
710 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
711 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
712 // Irreconcilable differences.
716 if (NumDifferences++) break;
721 if (NumDifferences == 0) // SAME GEP?
722 return ReplaceInstUsesWith(I, // No comparison is needed here.
723 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
725 else if (NumDifferences == 1 && GEPsInBounds) {
726 Value *LHSV = GEPLHS->getOperand(DiffOperand);
727 Value *RHSV = GEPRHS->getOperand(DiffOperand);
728 // Make sure we do a signed comparison here.
729 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
733 // Only lower this if the icmp is the only user of the GEP or if we expect
734 // the result to fold to a constant!
737 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
738 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
739 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
740 Value *L = EmitGEPOffset(GEPLHS);
741 Value *R = EmitGEPOffset(GEPRHS);
742 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
748 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
749 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
750 Value *X, ConstantInt *CI,
751 ICmpInst::Predicate Pred) {
752 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
753 // so the values can never be equal. Similarly for all other "or equals"
756 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
757 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
758 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
759 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
761 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
762 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
765 // (X+1) >u X --> X <u (0-1) --> X != 255
766 // (X+2) >u X --> X <u (0-2) --> X <u 254
767 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
768 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
769 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
771 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
772 ConstantInt *SMax = ConstantInt::get(X->getContext(),
773 APInt::getSignedMaxValue(BitWidth));
775 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
776 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
777 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
778 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
779 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
780 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
781 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
782 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
784 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
785 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
786 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
787 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
788 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
789 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
791 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
792 Constant *C = Builder->getInt(CI->getValue()-1);
793 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
796 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
797 /// and CmpRHS are both known to be integer constants.
798 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
799 ConstantInt *DivRHS) {
800 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
801 const APInt &CmpRHSV = CmpRHS->getValue();
803 // FIXME: If the operand types don't match the type of the divide
804 // then don't attempt this transform. The code below doesn't have the
805 // logic to deal with a signed divide and an unsigned compare (and
806 // vice versa). This is because (x /s C1) <s C2 produces different
807 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
808 // (x /u C1) <u C2. Simply casting the operands and result won't
809 // work. :( The if statement below tests that condition and bails
811 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
812 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
814 if (DivRHS->isZero())
815 return nullptr; // The ProdOV computation fails on divide by zero.
816 if (DivIsSigned && DivRHS->isAllOnesValue())
817 return nullptr; // The overflow computation also screws up here
818 if (DivRHS->isOne()) {
819 // This eliminates some funny cases with INT_MIN.
820 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
824 // Compute Prod = CI * DivRHS. We are essentially solving an equation
825 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
826 // C2 (CI). By solving for X we can turn this into a range check
827 // instead of computing a divide.
828 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
830 // Determine if the product overflows by seeing if the product is
831 // not equal to the divide. Make sure we do the same kind of divide
832 // as in the LHS instruction that we're folding.
833 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
834 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
836 // Get the ICmp opcode
837 ICmpInst::Predicate Pred = ICI.getPredicate();
839 /// If the division is known to be exact, then there is no remainder from the
840 /// divide, so the covered range size is unit, otherwise it is the divisor.
841 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
843 // Figure out the interval that is being checked. For example, a comparison
844 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
845 // Compute this interval based on the constants involved and the signedness of
846 // the compare/divide. This computes a half-open interval, keeping track of
847 // whether either value in the interval overflows. After analysis each
848 // overflow variable is set to 0 if it's corresponding bound variable is valid
849 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
850 int LoOverflow = 0, HiOverflow = 0;
851 Constant *LoBound = nullptr, *HiBound = nullptr;
853 if (!DivIsSigned) { // udiv
854 // e.g. X/5 op 3 --> [15, 20)
856 HiOverflow = LoOverflow = ProdOV;
858 // If this is not an exact divide, then many values in the range collapse
859 // to the same result value.
860 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
863 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
864 if (CmpRHSV == 0) { // (X / pos) op 0
865 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
866 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
868 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
869 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
870 HiOverflow = LoOverflow = ProdOV;
872 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
873 } else { // (X / pos) op neg
874 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
875 HiBound = AddOne(Prod);
876 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
878 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
879 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
882 } else if (DivRHS->isNegative()) { // Divisor is < 0.
884 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
885 if (CmpRHSV == 0) { // (X / neg) op 0
886 // e.g. X/-5 op 0 --> [-4, 5)
887 LoBound = AddOne(RangeSize);
888 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
889 if (HiBound == DivRHS) { // -INTMIN = INTMIN
890 HiOverflow = 1; // [INTMIN+1, overflow)
891 HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN
893 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
894 // e.g. X/-5 op 3 --> [-19, -14)
895 HiBound = AddOne(Prod);
896 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
898 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
899 } else { // (X / neg) op neg
900 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
901 LoOverflow = HiOverflow = ProdOV;
903 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
906 // Dividing by a negative swaps the condition. LT <-> GT
907 Pred = ICmpInst::getSwappedPredicate(Pred);
910 Value *X = DivI->getOperand(0);
912 default: llvm_unreachable("Unhandled icmp opcode!");
913 case ICmpInst::ICMP_EQ:
914 if (LoOverflow && HiOverflow)
915 return ReplaceInstUsesWith(ICI, Builder->getFalse());
917 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
918 ICmpInst::ICMP_UGE, X, LoBound);
920 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
921 ICmpInst::ICMP_ULT, X, HiBound);
922 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
924 case ICmpInst::ICMP_NE:
925 if (LoOverflow && HiOverflow)
926 return ReplaceInstUsesWith(ICI, Builder->getTrue());
928 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
929 ICmpInst::ICMP_ULT, X, LoBound);
931 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
932 ICmpInst::ICMP_UGE, X, HiBound);
933 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
934 DivIsSigned, false));
935 case ICmpInst::ICMP_ULT:
936 case ICmpInst::ICMP_SLT:
937 if (LoOverflow == +1) // Low bound is greater than input range.
938 return ReplaceInstUsesWith(ICI, Builder->getTrue());
939 if (LoOverflow == -1) // Low bound is less than input range.
940 return ReplaceInstUsesWith(ICI, Builder->getFalse());
941 return new ICmpInst(Pred, X, LoBound);
942 case ICmpInst::ICMP_UGT:
943 case ICmpInst::ICMP_SGT:
944 if (HiOverflow == +1) // High bound greater than input range.
945 return ReplaceInstUsesWith(ICI, Builder->getFalse());
946 if (HiOverflow == -1) // High bound less than input range.
947 return ReplaceInstUsesWith(ICI, Builder->getTrue());
948 if (Pred == ICmpInst::ICMP_UGT)
949 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
950 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
954 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
955 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
956 ConstantInt *ShAmt) {
957 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
959 // Check that the shift amount is in range. If not, don't perform
960 // undefined shifts. When the shift is visited it will be
962 uint32_t TypeBits = CmpRHSV.getBitWidth();
963 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
964 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
967 if (!ICI.isEquality()) {
968 // If we have an unsigned comparison and an ashr, we can't simplify this.
969 // Similarly for signed comparisons with lshr.
970 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
973 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
974 // by a power of 2. Since we already have logic to simplify these,
975 // transform to div and then simplify the resultant comparison.
976 if (Shr->getOpcode() == Instruction::AShr &&
977 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
980 // Revisit the shift (to delete it).
984 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
987 Shr->getOpcode() == Instruction::AShr ?
988 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
989 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
991 ICI.setOperand(0, Tmp);
993 // If the builder folded the binop, just return it.
994 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
998 // Otherwise, fold this div/compare.
999 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1000 TheDiv->getOpcode() == Instruction::UDiv);
1002 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1003 assert(Res && "This div/cst should have folded!");
1008 // If we are comparing against bits always shifted out, the
1009 // comparison cannot succeed.
1010 APInt Comp = CmpRHSV << ShAmtVal;
1011 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1012 if (Shr->getOpcode() == Instruction::LShr)
1013 Comp = Comp.lshr(ShAmtVal);
1015 Comp = Comp.ashr(ShAmtVal);
1017 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1018 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1019 Constant *Cst = Builder->getInt1(IsICMP_NE);
1020 return ReplaceInstUsesWith(ICI, Cst);
1023 // Otherwise, check to see if the bits shifted out are known to be zero.
1024 // If so, we can compare against the unshifted value:
1025 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1026 if (Shr->hasOneUse() && Shr->isExact())
1027 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1029 if (Shr->hasOneUse()) {
1030 // Otherwise strength reduce the shift into an and.
1031 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1032 Constant *Mask = Builder->getInt(Val);
1034 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1035 Mask, Shr->getName()+".mask");
1036 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1041 /// FoldICmpCstShrCst - Handle "(icmp eq/ne (ashr/lshr const2, A), const1)" ->
1042 /// (icmp eq/ne A, Log2(const2/const1)) ->
1043 /// (icmp eq/ne A, Log2(const2) - Log2(const1)).
1044 Instruction *InstCombiner::FoldICmpCstShrCst(ICmpInst &I, Value *Op, Value *A,
1047 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1049 auto getConstant = [&I, this](bool IsTrue) {
1050 if (I.getPredicate() == I.ICMP_NE)
1052 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1055 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1056 if (I.getPredicate() == I.ICMP_NE)
1057 Pred = CmpInst::getInversePredicate(Pred);
1058 return new ICmpInst(Pred, LHS, RHS);
1061 APInt AP1 = CI1->getValue();
1062 APInt AP2 = CI2->getValue();
1064 // Don't bother doing any work for cases which InstSimplify handles.
1067 bool IsAShr = isa<AShrOperator>(Op);
1069 if (AP2.isAllOnesValue())
1071 if (AP2.isNegative() != AP1.isNegative())
1078 // 'A' must be large enough to shift out the highest set bit.
1079 return getICmp(I.ICMP_UGT, A,
1080 ConstantInt::get(A->getType(), AP2.logBase2()));
1083 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1085 // Get the distance between the highest bit that's set.
1087 // Both the constants are negative, take their positive to calculate log.
1088 if (IsAShr && AP1.isNegative())
1089 // Get the ones' complement of AP2 and AP1 when computing the distance.
1090 Shift = (~AP2).logBase2() - (~AP1).logBase2();
1092 Shift = AP2.logBase2() - AP1.logBase2();
1095 if (IsAShr ? AP1 == AP2.ashr(Shift) : AP1 == AP2.lshr(Shift))
1096 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1098 // Shifting const2 will never be equal to const1.
1099 return getConstant(false);
1102 /// FoldICmpCstShlCst - Handle "(icmp eq/ne (shl const2, A), const1)" ->
1103 /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)).
1104 Instruction *InstCombiner::FoldICmpCstShlCst(ICmpInst &I, Value *Op, Value *A,
1107 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1109 auto getConstant = [&I, this](bool IsTrue) {
1110 if (I.getPredicate() == I.ICMP_NE)
1112 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), IsTrue));
1115 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1116 if (I.getPredicate() == I.ICMP_NE)
1117 Pred = CmpInst::getInversePredicate(Pred);
1118 return new ICmpInst(Pred, LHS, RHS);
1121 APInt AP1 = CI1->getValue();
1122 APInt AP2 = CI2->getValue();
1124 // Don't bother doing any work for cases which InstSimplify handles.
1128 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1130 if (!AP1 && AP2TrailingZeros != 0)
1131 return getICmp(I.ICMP_UGE, A,
1132 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1135 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1137 // Get the distance between the lowest bits that are set.
1138 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1140 if (Shift > 0 && AP2.shl(Shift) == AP1)
1141 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1143 // Shifting const2 will never be equal to const1.
1144 return getConstant(false);
1147 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1149 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1152 const APInt &RHSV = RHS->getValue();
1154 switch (LHSI->getOpcode()) {
1155 case Instruction::Trunc:
1156 if (ICI.isEquality() && LHSI->hasOneUse()) {
1157 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1158 // of the high bits truncated out of x are known.
1159 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1160 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1161 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1162 computeKnownBits(LHSI->getOperand(0), KnownZero, KnownOne, 0, &ICI);
1164 // If all the high bits are known, we can do this xform.
1165 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1166 // Pull in the high bits from known-ones set.
1167 APInt NewRHS = RHS->getValue().zext(SrcBits);
1168 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1169 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1170 Builder->getInt(NewRHS));
1175 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1176 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1177 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1179 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1180 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1181 Value *CompareVal = LHSI->getOperand(0);
1183 // If the sign bit of the XorCst is not set, there is no change to
1184 // the operation, just stop using the Xor.
1185 if (!XorCst->isNegative()) {
1186 ICI.setOperand(0, CompareVal);
1191 // Was the old condition true if the operand is positive?
1192 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1194 // If so, the new one isn't.
1195 isTrueIfPositive ^= true;
1197 if (isTrueIfPositive)
1198 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1201 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1205 if (LHSI->hasOneUse()) {
1206 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1207 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1208 const APInt &SignBit = XorCst->getValue();
1209 ICmpInst::Predicate Pred = ICI.isSigned()
1210 ? ICI.getUnsignedPredicate()
1211 : ICI.getSignedPredicate();
1212 return new ICmpInst(Pred, LHSI->getOperand(0),
1213 Builder->getInt(RHSV ^ SignBit));
1216 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1217 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1218 const APInt &NotSignBit = XorCst->getValue();
1219 ICmpInst::Predicate Pred = ICI.isSigned()
1220 ? ICI.getUnsignedPredicate()
1221 : ICI.getSignedPredicate();
1222 Pred = ICI.getSwappedPredicate(Pred);
1223 return new ICmpInst(Pred, LHSI->getOperand(0),
1224 Builder->getInt(RHSV ^ NotSignBit));
1228 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1229 // iff -C is a power of 2
1230 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1231 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1232 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1234 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1235 // iff -C is a power of 2
1236 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1237 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1238 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1241 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1242 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1243 LHSI->getOperand(0)->hasOneUse()) {
1244 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1246 // If the LHS is an AND of a truncating cast, we can widen the
1247 // and/compare to be the input width without changing the value
1248 // produced, eliminating a cast.
1249 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1250 // We can do this transformation if either the AND constant does not
1251 // have its sign bit set or if it is an equality comparison.
1252 // Extending a relational comparison when we're checking the sign
1253 // bit would not work.
1254 if (ICI.isEquality() ||
1255 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1257 Builder->CreateAnd(Cast->getOperand(0),
1258 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1259 NewAnd->takeName(LHSI);
1260 return new ICmpInst(ICI.getPredicate(), NewAnd,
1261 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1265 // If the LHS is an AND of a zext, and we have an equality compare, we can
1266 // shrink the and/compare to the smaller type, eliminating the cast.
1267 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1268 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1269 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1270 // should fold the icmp to true/false in that case.
1271 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1273 Builder->CreateAnd(Cast->getOperand(0),
1274 ConstantExpr::getTrunc(AndCst, Ty));
1275 NewAnd->takeName(LHSI);
1276 return new ICmpInst(ICI.getPredicate(), NewAnd,
1277 ConstantExpr::getTrunc(RHS, Ty));
1281 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1282 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1283 // happens a LOT in code produced by the C front-end, for bitfield
1285 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1286 if (Shift && !Shift->isShift())
1290 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : nullptr;
1292 // This seemingly simple opportunity to fold away a shift turns out to
1293 // be rather complicated. See PR17827
1294 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1296 bool CanFold = false;
1297 unsigned ShiftOpcode = Shift->getOpcode();
1298 if (ShiftOpcode == Instruction::AShr) {
1299 // There may be some constraints that make this possible,
1300 // but nothing simple has been discovered yet.
1302 } else if (ShiftOpcode == Instruction::Shl) {
1303 // For a left shift, we can fold if the comparison is not signed.
1304 // We can also fold a signed comparison if the mask value and
1305 // comparison value are not negative. These constraints may not be
1306 // obvious, but we can prove that they are correct using an SMT
1308 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1310 } else if (ShiftOpcode == Instruction::LShr) {
1311 // For a logical right shift, we can fold if the comparison is not
1312 // signed. We can also fold a signed comparison if the shifted mask
1313 // value and the shifted comparison value are not negative.
1314 // These constraints may not be obvious, but we can prove that they
1315 // are correct using an SMT solver.
1316 if (!ICI.isSigned())
1319 ConstantInt *ShiftedAndCst =
1320 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1321 ConstantInt *ShiftedRHSCst =
1322 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1324 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1331 if (ShiftOpcode == Instruction::Shl)
1332 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1334 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1336 // Check to see if we are shifting out any of the bits being
1338 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1339 // If we shifted bits out, the fold is not going to work out.
1340 // As a special case, check to see if this means that the
1341 // result is always true or false now.
1342 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1343 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1344 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1345 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1347 ICI.setOperand(1, NewCst);
1348 Constant *NewAndCst;
1349 if (ShiftOpcode == Instruction::Shl)
1350 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1352 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1353 LHSI->setOperand(1, NewAndCst);
1354 LHSI->setOperand(0, Shift->getOperand(0));
1355 Worklist.Add(Shift); // Shift is dead.
1361 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1362 // preferable because it allows the C<<Y expression to be hoisted out
1363 // of a loop if Y is invariant and X is not.
1364 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1365 ICI.isEquality() && !Shift->isArithmeticShift() &&
1366 !isa<Constant>(Shift->getOperand(0))) {
1369 if (Shift->getOpcode() == Instruction::LShr) {
1370 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1372 // Insert a logical shift.
1373 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1376 // Compute X & (C << Y).
1378 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1380 ICI.setOperand(0, NewAnd);
1384 // (icmp pred (and (or (lshr X, Y), X), 1), 0) -->
1385 // (icmp pred (and X, (or (shl 1, Y), 1), 0))
1387 // iff pred isn't signed
1389 Value *X, *Y, *LShr;
1390 if (!ICI.isSigned() && RHSV == 0) {
1391 if (match(LHSI->getOperand(1), m_One())) {
1392 Constant *One = cast<Constant>(LHSI->getOperand(1));
1393 Value *Or = LHSI->getOperand(0);
1394 if (match(Or, m_Or(m_Value(LShr), m_Value(X))) &&
1395 match(LShr, m_LShr(m_Specific(X), m_Value(Y)))) {
1396 unsigned UsesRemoved = 0;
1397 if (LHSI->hasOneUse())
1399 if (Or->hasOneUse())
1401 if (LShr->hasOneUse())
1403 Value *NewOr = nullptr;
1404 // Compute X & ((1 << Y) | 1)
1405 if (auto *C = dyn_cast<Constant>(Y)) {
1406 if (UsesRemoved >= 1)
1408 ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1410 if (UsesRemoved >= 3)
1411 NewOr = Builder->CreateOr(Builder->CreateShl(One, Y,
1414 One, Or->getName());
1417 Value *NewAnd = Builder->CreateAnd(X, NewOr, LHSI->getName());
1418 ICI.setOperand(0, NewAnd);
1426 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1427 // bit set in (X & AndCst) will produce a result greater than RHSV.
1428 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1429 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1430 if ((NTZ < AndCst->getBitWidth()) &&
1431 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1432 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1433 Constant::getNullValue(RHS->getType()));
1437 // Try to optimize things like "A[i]&42 == 0" to index computations.
1438 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1439 if (GetElementPtrInst *GEP =
1440 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1441 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1442 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1443 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1444 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1445 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1450 // X & -C == -C -> X > u ~C
1451 // X & -C != -C -> X <= u ~C
1452 // iff C is a power of 2
1453 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1454 return new ICmpInst(
1455 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1456 : ICmpInst::ICMP_ULE,
1457 LHSI->getOperand(0), SubOne(RHS));
1460 case Instruction::Or: {
1461 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1464 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1465 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1466 // -> and (icmp eq P, null), (icmp eq Q, null).
1467 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1468 Constant::getNullValue(P->getType()));
1469 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1470 Constant::getNullValue(Q->getType()));
1472 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1473 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1475 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1481 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1482 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1485 // If this is a signed comparison to 0 and the mul is sign preserving,
1486 // use the mul LHS operand instead.
1487 ICmpInst::Predicate pred = ICI.getPredicate();
1488 if (isSignTest(pred, RHS) && !Val->isZero() &&
1489 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1490 return new ICmpInst(Val->isNegative() ?
1491 ICmpInst::getSwappedPredicate(pred) : pred,
1492 LHSI->getOperand(0),
1493 Constant::getNullValue(RHS->getType()));
1498 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1499 uint32_t TypeBits = RHSV.getBitWidth();
1500 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1503 // (1 << X) pred P2 -> X pred Log2(P2)
1504 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1505 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1506 ICmpInst::Predicate Pred = ICI.getPredicate();
1507 if (ICI.isUnsigned()) {
1508 if (!RHSVIsPowerOf2) {
1509 // (1 << X) < 30 -> X <= 4
1510 // (1 << X) <= 30 -> X <= 4
1511 // (1 << X) >= 30 -> X > 4
1512 // (1 << X) > 30 -> X > 4
1513 if (Pred == ICmpInst::ICMP_ULT)
1514 Pred = ICmpInst::ICMP_ULE;
1515 else if (Pred == ICmpInst::ICMP_UGE)
1516 Pred = ICmpInst::ICMP_UGT;
1518 unsigned RHSLog2 = RHSV.logBase2();
1520 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1521 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1522 if (RHSLog2 == TypeBits-1) {
1523 if (Pred == ICmpInst::ICMP_UGE)
1524 Pred = ICmpInst::ICMP_EQ;
1525 else if (Pred == ICmpInst::ICMP_ULT)
1526 Pred = ICmpInst::ICMP_NE;
1529 return new ICmpInst(Pred, X,
1530 ConstantInt::get(RHS->getType(), RHSLog2));
1531 } else if (ICI.isSigned()) {
1532 if (RHSV.isAllOnesValue()) {
1533 // (1 << X) <= -1 -> X == 31
1534 if (Pred == ICmpInst::ICMP_SLE)
1535 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1536 ConstantInt::get(RHS->getType(), TypeBits-1));
1538 // (1 << X) > -1 -> X != 31
1539 if (Pred == ICmpInst::ICMP_SGT)
1540 return new ICmpInst(ICmpInst::ICMP_NE, X,
1541 ConstantInt::get(RHS->getType(), TypeBits-1));
1543 // (1 << X) < 0 -> X == 31
1544 // (1 << X) <= 0 -> X == 31
1545 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1546 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1547 ConstantInt::get(RHS->getType(), TypeBits-1));
1549 // (1 << X) >= 0 -> X != 31
1550 // (1 << X) > 0 -> X != 31
1551 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1552 return new ICmpInst(ICmpInst::ICMP_NE, X,
1553 ConstantInt::get(RHS->getType(), TypeBits-1));
1555 } else if (ICI.isEquality()) {
1557 return new ICmpInst(
1558 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1564 // Check that the shift amount is in range. If not, don't perform
1565 // undefined shifts. When the shift is visited it will be
1567 if (ShAmt->uge(TypeBits))
1570 if (ICI.isEquality()) {
1571 // If we are comparing against bits always shifted out, the
1572 // comparison cannot succeed.
1574 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1576 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1577 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1578 Constant *Cst = Builder->getInt1(IsICMP_NE);
1579 return ReplaceInstUsesWith(ICI, Cst);
1582 // If the shift is NUW, then it is just shifting out zeros, no need for an
1584 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1585 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1586 ConstantExpr::getLShr(RHS, ShAmt));
1588 // If the shift is NSW and we compare to 0, then it is just shifting out
1589 // sign bits, no need for an AND either.
1590 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1591 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1592 ConstantExpr::getLShr(RHS, ShAmt));
1594 if (LHSI->hasOneUse()) {
1595 // Otherwise strength reduce the shift into an and.
1596 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1597 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1598 TypeBits - ShAmtVal));
1601 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1602 return new ICmpInst(ICI.getPredicate(), And,
1603 ConstantExpr::getLShr(RHS, ShAmt));
1607 // If this is a signed comparison to 0 and the shift is sign preserving,
1608 // use the shift LHS operand instead.
1609 ICmpInst::Predicate pred = ICI.getPredicate();
1610 if (isSignTest(pred, RHS) &&
1611 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1612 return new ICmpInst(pred,
1613 LHSI->getOperand(0),
1614 Constant::getNullValue(RHS->getType()));
1616 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1617 bool TrueIfSigned = false;
1618 if (LHSI->hasOneUse() &&
1619 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1620 // (X << 31) <s 0 --> (X&1) != 0
1621 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1622 APInt::getOneBitSet(TypeBits,
1623 TypeBits-ShAmt->getZExtValue()-1));
1625 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1626 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1627 And, Constant::getNullValue(And->getType()));
1630 // Transform (icmp pred iM (shl iM %v, N), CI)
1631 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1632 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1633 // This enables to get rid of the shift in favor of a trunc which can be
1634 // free on the target. It has the additional benefit of comparing to a
1635 // smaller constant, which will be target friendly.
1636 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1637 if (LHSI->hasOneUse() &&
1638 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1639 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1640 Constant *NCI = ConstantExpr::getTrunc(
1641 ConstantExpr::getAShr(RHS,
1642 ConstantInt::get(RHS->getType(), Amt)),
1644 return new ICmpInst(ICI.getPredicate(),
1645 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1652 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1653 case Instruction::AShr: {
1654 // Handle equality comparisons of shift-by-constant.
1655 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1656 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1657 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1661 // Handle exact shr's.
1662 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1663 if (RHSV.isMinValue())
1664 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1669 case Instruction::SDiv:
1670 case Instruction::UDiv:
1671 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1672 // Fold this div into the comparison, producing a range check.
1673 // Determine, based on the divide type, what the range is being
1674 // checked. If there is an overflow on the low or high side, remember
1675 // it, otherwise compute the range [low, hi) bounding the new value.
1676 // See: InsertRangeTest above for the kinds of replacements possible.
1677 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1678 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1683 case Instruction::Sub: {
1684 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1686 const APInt &LHSV = LHSC->getValue();
1688 // C1-X <u C2 -> (X|(C2-1)) == C1
1689 // iff C1 & (C2-1) == C2-1
1690 // C2 is a power of 2
1691 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1692 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1693 return new ICmpInst(ICmpInst::ICMP_EQ,
1694 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1697 // C1-X >u C2 -> (X|C2) != C1
1698 // iff C1 & C2 == C2
1699 // C2+1 is a power of 2
1700 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1701 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1702 return new ICmpInst(ICmpInst::ICMP_NE,
1703 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1707 case Instruction::Add:
1708 // Fold: icmp pred (add X, C1), C2
1709 if (!ICI.isEquality()) {
1710 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1712 const APInt &LHSV = LHSC->getValue();
1714 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1717 if (ICI.isSigned()) {
1718 if (CR.getLower().isSignBit()) {
1719 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1720 Builder->getInt(CR.getUpper()));
1721 } else if (CR.getUpper().isSignBit()) {
1722 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1723 Builder->getInt(CR.getLower()));
1726 if (CR.getLower().isMinValue()) {
1727 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1728 Builder->getInt(CR.getUpper()));
1729 } else if (CR.getUpper().isMinValue()) {
1730 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1731 Builder->getInt(CR.getLower()));
1735 // X-C1 <u C2 -> (X & -C2) == C1
1736 // iff C1 & (C2-1) == 0
1737 // C2 is a power of 2
1738 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1739 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1740 return new ICmpInst(ICmpInst::ICMP_EQ,
1741 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1742 ConstantExpr::getNeg(LHSC));
1744 // X-C1 >u C2 -> (X & ~C2) != C1
1746 // C2+1 is a power of 2
1747 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1748 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1749 return new ICmpInst(ICmpInst::ICMP_NE,
1750 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1751 ConstantExpr::getNeg(LHSC));
1756 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1757 if (ICI.isEquality()) {
1758 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1760 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1761 // the second operand is a constant, simplify a bit.
1762 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1763 switch (BO->getOpcode()) {
1764 case Instruction::SRem:
1765 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1766 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1767 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1768 if (V.sgt(1) && V.isPowerOf2()) {
1770 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1772 return new ICmpInst(ICI.getPredicate(), NewRem,
1773 Constant::getNullValue(BO->getType()));
1777 case Instruction::Add:
1778 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1779 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1780 if (BO->hasOneUse())
1781 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1782 ConstantExpr::getSub(RHS, BOp1C));
1783 } else if (RHSV == 0) {
1784 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1785 // efficiently invertible, or if the add has just this one use.
1786 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1788 if (Value *NegVal = dyn_castNegVal(BOp1))
1789 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1790 if (Value *NegVal = dyn_castNegVal(BOp0))
1791 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1792 if (BO->hasOneUse()) {
1793 Value *Neg = Builder->CreateNeg(BOp1);
1795 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1799 case Instruction::Xor:
1800 // For the xor case, we can xor two constants together, eliminating
1801 // the explicit xor.
1802 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1803 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1804 ConstantExpr::getXor(RHS, BOC));
1805 } else if (RHSV == 0) {
1806 // Replace ((xor A, B) != 0) with (A != B)
1807 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1811 case Instruction::Sub:
1812 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1813 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1814 if (BO->hasOneUse())
1815 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1816 ConstantExpr::getSub(BOp0C, RHS));
1817 } else if (RHSV == 0) {
1818 // Replace ((sub A, B) != 0) with (A != B)
1819 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1823 case Instruction::Or:
1824 // If bits are being or'd in that are not present in the constant we
1825 // are comparing against, then the comparison could never succeed!
1826 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1827 Constant *NotCI = ConstantExpr::getNot(RHS);
1828 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1829 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1833 case Instruction::And:
1834 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1835 // If bits are being compared against that are and'd out, then the
1836 // comparison can never succeed!
1837 if ((RHSV & ~BOC->getValue()) != 0)
1838 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1840 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1841 if (RHS == BOC && RHSV.isPowerOf2())
1842 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1843 ICmpInst::ICMP_NE, LHSI,
1844 Constant::getNullValue(RHS->getType()));
1846 // Don't perform the following transforms if the AND has multiple uses
1847 if (!BO->hasOneUse())
1850 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1851 if (BOC->getValue().isSignBit()) {
1852 Value *X = BO->getOperand(0);
1853 Constant *Zero = Constant::getNullValue(X->getType());
1854 ICmpInst::Predicate pred = isICMP_NE ?
1855 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1856 return new ICmpInst(pred, X, Zero);
1859 // ((X & ~7) == 0) --> X < 8
1860 if (RHSV == 0 && isHighOnes(BOC)) {
1861 Value *X = BO->getOperand(0);
1862 Constant *NegX = ConstantExpr::getNeg(BOC);
1863 ICmpInst::Predicate pred = isICMP_NE ?
1864 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1865 return new ICmpInst(pred, X, NegX);
1869 case Instruction::Mul:
1870 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1871 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1872 // The trivial case (mul X, 0) is handled by InstSimplify
1873 // General case : (mul X, C) != 0 iff X != 0
1874 // (mul X, C) == 0 iff X == 0
1876 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1877 Constant::getNullValue(RHS->getType()));
1883 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1884 // Handle icmp {eq|ne} <intrinsic>, intcst.
1885 switch (II->getIntrinsicID()) {
1886 case Intrinsic::bswap:
1888 ICI.setOperand(0, II->getArgOperand(0));
1889 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1891 case Intrinsic::ctlz:
1892 case Intrinsic::cttz:
1893 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1894 if (RHSV == RHS->getType()->getBitWidth()) {
1896 ICI.setOperand(0, II->getArgOperand(0));
1897 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1901 case Intrinsic::ctpop:
1902 // popcount(A) == 0 -> A == 0 and likewise for !=
1903 if (RHS->isZero()) {
1905 ICI.setOperand(0, II->getArgOperand(0));
1906 ICI.setOperand(1, RHS);
1918 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1919 /// We only handle extending casts so far.
1921 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1922 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1923 Value *LHSCIOp = LHSCI->getOperand(0);
1924 Type *SrcTy = LHSCIOp->getType();
1925 Type *DestTy = LHSCI->getType();
1928 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1929 // integer type is the same size as the pointer type.
1930 if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
1931 DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1932 Value *RHSOp = nullptr;
1933 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1934 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1935 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1936 RHSOp = RHSC->getOperand(0);
1937 // If the pointer types don't match, insert a bitcast.
1938 if (LHSCIOp->getType() != RHSOp->getType())
1939 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1943 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1946 // The code below only handles extension cast instructions, so far.
1948 if (LHSCI->getOpcode() != Instruction::ZExt &&
1949 LHSCI->getOpcode() != Instruction::SExt)
1952 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1953 bool isSignedCmp = ICI.isSigned();
1955 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1956 // Not an extension from the same type?
1957 RHSCIOp = CI->getOperand(0);
1958 if (RHSCIOp->getType() != LHSCIOp->getType())
1961 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1962 // and the other is a zext), then we can't handle this.
1963 if (CI->getOpcode() != LHSCI->getOpcode())
1966 // Deal with equality cases early.
1967 if (ICI.isEquality())
1968 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1970 // A signed comparison of sign extended values simplifies into a
1971 // signed comparison.
1972 if (isSignedCmp && isSignedExt)
1973 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1975 // The other three cases all fold into an unsigned comparison.
1976 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1979 // If we aren't dealing with a constant on the RHS, exit early
1980 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1984 // Compute the constant that would happen if we truncated to SrcTy then
1985 // reextended to DestTy.
1986 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1987 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1990 // If the re-extended constant didn't change...
1992 // Deal with equality cases early.
1993 if (ICI.isEquality())
1994 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1996 // A signed comparison of sign extended values simplifies into a
1997 // signed comparison.
1998 if (isSignedExt && isSignedCmp)
1999 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
2001 // The other three cases all fold into an unsigned comparison.
2002 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
2005 // The re-extended constant changed so the constant cannot be represented
2006 // in the shorter type. Consequently, we cannot emit a simple comparison.
2007 // All the cases that fold to true or false will have already been handled
2008 // by SimplifyICmpInst, so only deal with the tricky case.
2010 if (isSignedCmp || !isSignedExt)
2013 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
2014 // should have been folded away previously and not enter in here.
2016 // We're performing an unsigned comp with a sign extended value.
2017 // This is true if the input is >= 0. [aka >s -1]
2018 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
2019 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
2021 // Finally, return the value computed.
2022 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
2023 return ReplaceInstUsesWith(ICI, Result);
2025 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
2026 return BinaryOperator::CreateNot(Result);
2029 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
2030 /// I = icmp ugt (add (add A, B), CI2), CI1
2031 /// If this is of the form:
2033 /// if (sum+128 >u 255)
2034 /// Then replace it with llvm.sadd.with.overflow.i8.
2036 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
2037 ConstantInt *CI2, ConstantInt *CI1,
2039 // The transformation we're trying to do here is to transform this into an
2040 // llvm.sadd.with.overflow. To do this, we have to replace the original add
2041 // with a narrower add, and discard the add-with-constant that is part of the
2042 // range check (if we can't eliminate it, this isn't profitable).
2044 // In order to eliminate the add-with-constant, the compare can be its only
2046 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
2047 if (!AddWithCst->hasOneUse()) return nullptr;
2049 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
2050 if (!CI2->getValue().isPowerOf2()) return nullptr;
2051 unsigned NewWidth = CI2->getValue().countTrailingZeros();
2052 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return nullptr;
2054 // The width of the new add formed is 1 more than the bias.
2057 // Check to see that CI1 is an all-ones value with NewWidth bits.
2058 if (CI1->getBitWidth() == NewWidth ||
2059 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
2062 // This is only really a signed overflow check if the inputs have been
2063 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
2064 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
2065 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
2066 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
2067 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
2070 // In order to replace the original add with a narrower
2071 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
2072 // and truncates that discard the high bits of the add. Verify that this is
2074 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
2075 for (User *U : OrigAdd->users()) {
2076 if (U == AddWithCst) continue;
2078 // Only accept truncates for now. We would really like a nice recursive
2079 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
2080 // chain to see which bits of a value are actually demanded. If the
2081 // original add had another add which was then immediately truncated, we
2082 // could still do the transformation.
2083 TruncInst *TI = dyn_cast<TruncInst>(U);
2084 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
2088 // If the pattern matches, truncate the inputs to the narrower type and
2089 // use the sadd_with_overflow intrinsic to efficiently compute both the
2090 // result and the overflow bit.
2091 Module *M = I.getParent()->getParent()->getParent();
2093 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
2094 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
2097 InstCombiner::BuilderTy *Builder = IC.Builder;
2099 // Put the new code above the original add, in case there are any uses of the
2100 // add between the add and the compare.
2101 Builder->SetInsertPoint(OrigAdd);
2103 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
2104 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
2105 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
2106 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
2107 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
2109 // The inner add was the result of the narrow add, zero extended to the
2110 // wider type. Replace it with the result computed by the intrinsic.
2111 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
2113 // The original icmp gets replaced with the overflow value.
2114 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
2117 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
2119 // Don't bother doing this transformation for pointers, don't do it for
2121 if (!isa<IntegerType>(OrigAddV->getType())) return nullptr;
2123 // If the add is a constant expr, then we don't bother transforming it.
2124 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
2125 if (!OrigAdd) return nullptr;
2127 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
2129 // Put the new code above the original add, in case there are any uses of the
2130 // add between the add and the compare.
2131 InstCombiner::BuilderTy *Builder = IC.Builder;
2132 Builder->SetInsertPoint(OrigAdd);
2134 Module *M = I.getParent()->getParent()->getParent();
2135 Type *Ty = LHS->getType();
2136 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
2137 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
2138 Value *Add = Builder->CreateExtractValue(Call, 0);
2140 IC.ReplaceInstUsesWith(*OrigAdd, Add);
2142 // The original icmp gets replaced with the overflow value.
2143 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2146 /// \brief Recognize and process idiom involving test for multiplication
2149 /// The caller has matched a pattern of the form:
2150 /// I = cmp u (mul(zext A, zext B), V
2151 /// The function checks if this is a test for overflow and if so replaces
2152 /// multiplication with call to 'mul.with.overflow' intrinsic.
2154 /// \param I Compare instruction.
2155 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
2156 /// the compare instruction. Must be of integer type.
2157 /// \param OtherVal The other argument of compare instruction.
2158 /// \returns Instruction which must replace the compare instruction, NULL if no
2159 /// replacement required.
2160 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
2161 Value *OtherVal, InstCombiner &IC) {
2162 // Don't bother doing this transformation for pointers, don't do it for
2164 if (!isa<IntegerType>(MulVal->getType()))
2167 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
2168 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
2169 Instruction *MulInstr = cast<Instruction>(MulVal);
2170 assert(MulInstr->getOpcode() == Instruction::Mul);
2172 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
2173 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
2174 assert(LHS->getOpcode() == Instruction::ZExt);
2175 assert(RHS->getOpcode() == Instruction::ZExt);
2176 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
2178 // Calculate type and width of the result produced by mul.with.overflow.
2179 Type *TyA = A->getType(), *TyB = B->getType();
2180 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
2181 WidthB = TyB->getPrimitiveSizeInBits();
2184 if (WidthB > WidthA) {
2192 // In order to replace the original mul with a narrower mul.with.overflow,
2193 // all uses must ignore upper bits of the product. The number of used low
2194 // bits must be not greater than the width of mul.with.overflow.
2195 if (MulVal->hasNUsesOrMore(2))
2196 for (User *U : MulVal->users()) {
2199 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2200 // Check if truncation ignores bits above MulWidth.
2201 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
2202 if (TruncWidth > MulWidth)
2204 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2205 // Check if AND ignores bits above MulWidth.
2206 if (BO->getOpcode() != Instruction::And)
2208 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
2209 const APInt &CVal = CI->getValue();
2210 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
2214 // Other uses prohibit this transformation.
2219 // Recognize patterns
2220 switch (I.getPredicate()) {
2221 case ICmpInst::ICMP_EQ:
2222 case ICmpInst::ICMP_NE:
2223 // Recognize pattern:
2224 // mulval = mul(zext A, zext B)
2225 // cmp eq/neq mulval, zext trunc mulval
2226 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
2227 if (Zext->hasOneUse()) {
2228 Value *ZextArg = Zext->getOperand(0);
2229 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
2230 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
2234 // Recognize pattern:
2235 // mulval = mul(zext A, zext B)
2236 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
2239 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
2240 if (ValToMask != MulVal)
2242 const APInt &CVal = CI->getValue() + 1;
2243 if (CVal.isPowerOf2()) {
2244 unsigned MaskWidth = CVal.logBase2();
2245 if (MaskWidth == MulWidth)
2246 break; // Recognized
2251 case ICmpInst::ICMP_UGT:
2252 // Recognize pattern:
2253 // mulval = mul(zext A, zext B)
2254 // cmp ugt mulval, max
2255 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2256 APInt MaxVal = APInt::getMaxValue(MulWidth);
2257 MaxVal = MaxVal.zext(CI->getBitWidth());
2258 if (MaxVal.eq(CI->getValue()))
2259 break; // Recognized
2263 case ICmpInst::ICMP_UGE:
2264 // Recognize pattern:
2265 // mulval = mul(zext A, zext B)
2266 // cmp uge mulval, max+1
2267 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2268 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2269 if (MaxVal.eq(CI->getValue()))
2270 break; // Recognized
2274 case ICmpInst::ICMP_ULE:
2275 // Recognize pattern:
2276 // mulval = mul(zext A, zext B)
2277 // cmp ule mulval, max
2278 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2279 APInt MaxVal = APInt::getMaxValue(MulWidth);
2280 MaxVal = MaxVal.zext(CI->getBitWidth());
2281 if (MaxVal.eq(CI->getValue()))
2282 break; // Recognized
2286 case ICmpInst::ICMP_ULT:
2287 // Recognize pattern:
2288 // mulval = mul(zext A, zext B)
2289 // cmp ule mulval, max + 1
2290 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2291 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2292 if (MaxVal.eq(CI->getValue()))
2293 break; // Recognized
2301 InstCombiner::BuilderTy *Builder = IC.Builder;
2302 Builder->SetInsertPoint(MulInstr);
2303 Module *M = I.getParent()->getParent()->getParent();
2305 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
2306 Value *MulA = A, *MulB = B;
2307 if (WidthA < MulWidth)
2308 MulA = Builder->CreateZExt(A, MulType);
2309 if (WidthB < MulWidth)
2310 MulB = Builder->CreateZExt(B, MulType);
2312 Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
2313 CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
2314 IC.Worklist.Add(MulInstr);
2316 // If there are uses of mul result other than the comparison, we know that
2317 // they are truncation or binary AND. Change them to use result of
2318 // mul.with.overflow and adjust properly mask/size.
2319 if (MulVal->hasNUsesOrMore(2)) {
2320 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
2321 for (User *U : MulVal->users()) {
2322 if (U == &I || U == OtherVal)
2324 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2325 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
2326 IC.ReplaceInstUsesWith(*TI, Mul);
2328 TI->setOperand(0, Mul);
2329 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2330 assert(BO->getOpcode() == Instruction::And);
2331 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
2332 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
2333 APInt ShortMask = CI->getValue().trunc(MulWidth);
2334 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
2336 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
2337 IC.Worklist.Add(Zext);
2338 IC.ReplaceInstUsesWith(*BO, Zext);
2340 llvm_unreachable("Unexpected Binary operation");
2342 IC.Worklist.Add(cast<Instruction>(U));
2345 if (isa<Instruction>(OtherVal))
2346 IC.Worklist.Add(cast<Instruction>(OtherVal));
2348 // The original icmp gets replaced with the overflow value, maybe inverted
2349 // depending on predicate.
2350 bool Inverse = false;
2351 switch (I.getPredicate()) {
2352 case ICmpInst::ICMP_NE:
2354 case ICmpInst::ICMP_EQ:
2357 case ICmpInst::ICMP_UGT:
2358 case ICmpInst::ICMP_UGE:
2359 if (I.getOperand(0) == MulVal)
2363 case ICmpInst::ICMP_ULT:
2364 case ICmpInst::ICMP_ULE:
2365 if (I.getOperand(1) == MulVal)
2370 llvm_unreachable("Unexpected predicate");
2373 Value *Res = Builder->CreateExtractValue(Call, 1);
2374 return BinaryOperator::CreateNot(Res);
2377 return ExtractValueInst::Create(Call, 1);
2380 // DemandedBitsLHSMask - When performing a comparison against a constant,
2381 // it is possible that not all the bits in the LHS are demanded. This helper
2382 // method computes the mask that IS demanded.
2383 static APInt DemandedBitsLHSMask(ICmpInst &I,
2384 unsigned BitWidth, bool isSignCheck) {
2386 return APInt::getSignBit(BitWidth);
2388 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2389 if (!CI) return APInt::getAllOnesValue(BitWidth);
2390 const APInt &RHS = CI->getValue();
2392 switch (I.getPredicate()) {
2393 // For a UGT comparison, we don't care about any bits that
2394 // correspond to the trailing ones of the comparand. The value of these
2395 // bits doesn't impact the outcome of the comparison, because any value
2396 // greater than the RHS must differ in a bit higher than these due to carry.
2397 case ICmpInst::ICMP_UGT: {
2398 unsigned trailingOnes = RHS.countTrailingOnes();
2399 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2403 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2404 // Any value less than the RHS must differ in a higher bit because of carries.
2405 case ICmpInst::ICMP_ULT: {
2406 unsigned trailingZeros = RHS.countTrailingZeros();
2407 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2412 return APInt::getAllOnesValue(BitWidth);
2417 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2418 /// should be swapped.
2419 /// The decision is based on how many times these two operands are reused
2420 /// as subtract operands and their positions in those instructions.
2421 /// The rational is that several architectures use the same instruction for
2422 /// both subtract and cmp, thus it is better if the order of those operands
2424 /// \return true if Op0 and Op1 should be swapped.
2425 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2426 const Value * Op1) {
2427 // Filter out pointer value as those cannot appears directly in subtract.
2428 // FIXME: we may want to go through inttoptrs or bitcasts.
2429 if (Op0->getType()->isPointerTy())
2431 // Count every uses of both Op0 and Op1 in a subtract.
2432 // Each time Op0 is the first operand, count -1: swapping is bad, the
2433 // subtract has already the same layout as the compare.
2434 // Each time Op0 is the second operand, count +1: swapping is good, the
2435 // subtract has a different layout as the compare.
2436 // At the end, if the benefit is greater than 0, Op0 should come second to
2437 // expose more CSE opportunities.
2438 int GlobalSwapBenefits = 0;
2439 for (const User *U : Op0->users()) {
2440 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
2441 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2443 // If Op0 is the first argument, this is not beneficial to swap the
2445 int LocalSwapBenefits = -1;
2446 unsigned Op1Idx = 1;
2447 if (BinOp->getOperand(Op1Idx) == Op0) {
2449 LocalSwapBenefits = 1;
2451 if (BinOp->getOperand(Op1Idx) != Op1)
2453 GlobalSwapBenefits += LocalSwapBenefits;
2455 return GlobalSwapBenefits > 0;
2458 /// \brief Check that one use is in the same block as the definition and all
2459 /// other uses are in blocks dominated by a given block
2461 /// \param DI Definition
2463 /// \param DB Block that must dominate all uses of \p DI outside
2464 /// the parent block
2465 /// \return true when \p UI is the only use of \p DI in the parent block
2466 /// and all other uses of \p DI are in blocks dominated by \p DB.
2468 bool InstCombiner::dominatesAllUses(const Instruction *DI,
2469 const Instruction *UI,
2470 const BasicBlock *DB) const {
2471 assert(DI && UI && "Instruction not defined\n");
2472 // ignore incomplete definitions
2473 if (!DI->getParent())
2475 // DI and UI must be in the same block
2476 if (DI->getParent() != UI->getParent())
2478 // Protect from self-referencing blocks
2479 if (DI->getParent() == DB)
2481 // DominatorTree available?
2484 for (const User *U : DI->users()) {
2485 auto *Usr = cast<Instruction>(U);
2486 if (Usr != UI && !DT->dominates(DB, Usr->getParent()))
2493 /// true when the instruction sequence within a block is select-cmp-br.
2495 static bool isChainSelectCmpBranch(const SelectInst *SI) {
2496 const BasicBlock *BB = SI->getParent();
2499 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
2500 if (!BI || BI->getNumSuccessors() != 2)
2502 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
2503 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
2509 /// \brief True when a select result is replaced by one of its operands
2510 /// in select-icmp sequence. This will eventually result in the elimination
2513 /// \param SI Select instruction
2514 /// \param Icmp Compare instruction
2515 /// \param SIOpd Operand that replaces the select
2518 /// - The replacement is global and requires dominator information
2519 /// - The caller is responsible for the actual replacement
2524 /// %4 = select i1 %3, %C* %0, %C* null
2525 /// %5 = icmp eq %C* %4, null
2526 /// br i1 %5, label %9, label %7
2528 /// ; <label>:7 ; preds = %entry
2529 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
2532 /// can be transformed to
2534 /// %5 = icmp eq %C* %0, null
2535 /// %6 = select i1 %3, i1 %5, i1 true
2536 /// br i1 %6, label %9, label %7
2538 /// ; <label>:7 ; preds = %entry
2539 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
2541 /// Similar when the first operand of the select is a constant or/and
2542 /// the compare is for not equal rather than equal.
2544 /// NOTE: The function is only called when the select and compare constants
2545 /// are equal, the optimization can work only for EQ predicates. This is not a
2546 /// major restriction since a NE compare should be 'normalized' to an equal
2547 /// compare, which usually happens in the combiner and test case
2548 /// select-cmp-br.ll
2550 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
2551 const ICmpInst *Icmp,
2552 const unsigned SIOpd) {
2553 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
2554 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
2555 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
2556 // The check for the unique predecessor is not the best that can be
2557 // done. But it protects efficiently against cases like when SI's
2558 // home block has two successors, Succ and Succ1, and Succ1 predecessor
2559 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
2560 // replaced can be reached on either path. So the uniqueness check
2561 // guarantees that the path all uses of SI (outside SI's parent) are on
2562 // is disjoint from all other paths out of SI. But that information
2563 // is more expensive to compute, and the trade-off here is in favor
2565 if (Succ->getUniquePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
2567 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
2574 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2575 bool Changed = false;
2576 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2577 unsigned Op0Cplxity = getComplexity(Op0);
2578 unsigned Op1Cplxity = getComplexity(Op1);
2580 /// Orders the operands of the compare so that they are listed from most
2581 /// complex to least complex. This puts constants before unary operators,
2582 /// before binary operators.
2583 if (Op0Cplxity < Op1Cplxity ||
2584 (Op0Cplxity == Op1Cplxity &&
2585 swapMayExposeCSEOpportunities(Op0, Op1))) {
2587 std::swap(Op0, Op1);
2591 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
2592 return ReplaceInstUsesWith(I, V);
2594 // comparing -val or val with non-zero is the same as just comparing val
2595 // ie, abs(val) != 0 -> val != 0
2596 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2598 Value *Cond, *SelectTrue, *SelectFalse;
2599 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2600 m_Value(SelectFalse)))) {
2601 if (Value *V = dyn_castNegVal(SelectTrue)) {
2602 if (V == SelectFalse)
2603 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2605 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2606 if (V == SelectTrue)
2607 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2612 Type *Ty = Op0->getType();
2614 // icmp's with boolean values can always be turned into bitwise operations
2615 if (Ty->isIntegerTy(1)) {
2616 switch (I.getPredicate()) {
2617 default: llvm_unreachable("Invalid icmp instruction!");
2618 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2619 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2620 return BinaryOperator::CreateNot(Xor);
2622 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2623 return BinaryOperator::CreateXor(Op0, Op1);
2625 case ICmpInst::ICMP_UGT:
2626 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2628 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2629 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2630 return BinaryOperator::CreateAnd(Not, Op1);
2632 case ICmpInst::ICMP_SGT:
2633 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2635 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2636 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2637 return BinaryOperator::CreateAnd(Not, Op0);
2639 case ICmpInst::ICMP_UGE:
2640 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2642 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2643 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2644 return BinaryOperator::CreateOr(Not, Op1);
2646 case ICmpInst::ICMP_SGE:
2647 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2649 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2650 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2651 return BinaryOperator::CreateOr(Not, Op0);
2656 unsigned BitWidth = 0;
2657 if (Ty->isIntOrIntVectorTy())
2658 BitWidth = Ty->getScalarSizeInBits();
2659 else if (DL) // Pointers require DL info to get their size.
2660 BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
2662 bool isSignBit = false;
2664 // See if we are doing a comparison with a constant.
2665 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2666 Value *A = nullptr, *B = nullptr;
2668 // Match the following pattern, which is a common idiom when writing
2669 // overflow-safe integer arithmetic function. The source performs an
2670 // addition in wider type, and explicitly checks for overflow using
2671 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2672 // sadd_with_overflow intrinsic.
2674 // TODO: This could probably be generalized to handle other overflow-safe
2675 // operations if we worked out the formulas to compute the appropriate
2679 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2681 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2682 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2683 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2684 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2688 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2689 if (I.isEquality() && CI->isZero() &&
2690 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2691 // (icmp cond A B) if cond is equality
2692 return new ICmpInst(I.getPredicate(), A, B);
2695 // If we have an icmp le or icmp ge instruction, turn it into the
2696 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2697 // them being folded in the code below. The SimplifyICmpInst code has
2698 // already handled the edge cases for us, so we just assert on them.
2699 switch (I.getPredicate()) {
2701 case ICmpInst::ICMP_ULE:
2702 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2703 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2704 Builder->getInt(CI->getValue()+1));
2705 case ICmpInst::ICMP_SLE:
2706 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2707 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2708 Builder->getInt(CI->getValue()+1));
2709 case ICmpInst::ICMP_UGE:
2710 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2711 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2712 Builder->getInt(CI->getValue()-1));
2713 case ICmpInst::ICMP_SGE:
2714 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2715 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2716 Builder->getInt(CI->getValue()-1));
2719 if (I.isEquality()) {
2721 if (match(Op0, m_AShr(m_ConstantInt(CI2), m_Value(A))) ||
2722 match(Op0, m_LShr(m_ConstantInt(CI2), m_Value(A)))) {
2723 // (icmp eq/ne (ashr/lshr const2, A), const1)
2724 if (Instruction *Inst = FoldICmpCstShrCst(I, Op0, A, CI, CI2))
2727 if (match(Op0, m_Shl(m_ConstantInt(CI2), m_Value(A)))) {
2728 // (icmp eq/ne (shl const2, A), const1)
2729 if (Instruction *Inst = FoldICmpCstShlCst(I, Op0, A, CI, CI2))
2734 // If this comparison is a normal comparison, it demands all
2735 // bits, if it is a sign bit comparison, it only demands the sign bit.
2737 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2740 // See if we can fold the comparison based on range information we can get
2741 // by checking whether bits are known to be zero or one in the input.
2742 if (BitWidth != 0) {
2743 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2744 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2746 if (SimplifyDemandedBits(I.getOperandUse(0),
2747 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2748 Op0KnownZero, Op0KnownOne, 0))
2750 if (SimplifyDemandedBits(I.getOperandUse(1),
2751 APInt::getAllOnesValue(BitWidth),
2752 Op1KnownZero, Op1KnownOne, 0))
2755 // Given the known and unknown bits, compute a range that the LHS could be
2756 // in. Compute the Min, Max and RHS values based on the known bits. For the
2757 // EQ and NE we use unsigned values.
2758 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2759 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2761 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2763 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2766 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2768 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2772 // If Min and Max are known to be the same, then SimplifyDemandedBits
2773 // figured out that the LHS is a constant. Just constant fold this now so
2774 // that code below can assume that Min != Max.
2775 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2776 return new ICmpInst(I.getPredicate(),
2777 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2778 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2779 return new ICmpInst(I.getPredicate(), Op0,
2780 ConstantInt::get(Op1->getType(), Op1Min));
2782 // Based on the range information we know about the LHS, see if we can
2783 // simplify this comparison. For example, (x&4) < 8 is always true.
2784 switch (I.getPredicate()) {
2785 default: llvm_unreachable("Unknown icmp opcode!");
2786 case ICmpInst::ICMP_EQ: {
2787 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2788 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2790 // If all bits are known zero except for one, then we know at most one
2791 // bit is set. If the comparison is against zero, then this is a check
2792 // to see if *that* bit is set.
2793 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2794 if (~Op1KnownZero == 0) {
2795 // If the LHS is an AND with the same constant, look through it.
2796 Value *LHS = nullptr;
2797 ConstantInt *LHSC = nullptr;
2798 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2799 LHSC->getValue() != Op0KnownZeroInverted)
2802 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2803 // then turn "((1 << x)&8) == 0" into "x != 3".
2804 // or turn "((1 << x)&7) == 0" into "x > 2".
2806 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2807 APInt ValToCheck = Op0KnownZeroInverted;
2808 if (ValToCheck.isPowerOf2()) {
2809 unsigned CmpVal = ValToCheck.countTrailingZeros();
2810 return new ICmpInst(ICmpInst::ICMP_NE, X,
2811 ConstantInt::get(X->getType(), CmpVal));
2812 } else if ((++ValToCheck).isPowerOf2()) {
2813 unsigned CmpVal = ValToCheck.countTrailingZeros() - 1;
2814 return new ICmpInst(ICmpInst::ICMP_UGT, X,
2815 ConstantInt::get(X->getType(), CmpVal));
2819 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2820 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2822 if (Op0KnownZeroInverted == 1 &&
2823 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2824 return new ICmpInst(ICmpInst::ICMP_NE, X,
2825 ConstantInt::get(X->getType(),
2826 CI->countTrailingZeros()));
2831 case ICmpInst::ICMP_NE: {
2832 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2833 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2835 // If all bits are known zero except for one, then we know at most one
2836 // bit is set. If the comparison is against zero, then this is a check
2837 // to see if *that* bit is set.
2838 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2839 if (~Op1KnownZero == 0) {
2840 // If the LHS is an AND with the same constant, look through it.
2841 Value *LHS = nullptr;
2842 ConstantInt *LHSC = nullptr;
2843 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2844 LHSC->getValue() != Op0KnownZeroInverted)
2847 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2848 // then turn "((1 << x)&8) != 0" into "x == 3".
2849 // or turn "((1 << x)&7) != 0" into "x < 3".
2851 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2852 APInt ValToCheck = Op0KnownZeroInverted;
2853 if (ValToCheck.isPowerOf2()) {
2854 unsigned CmpVal = ValToCheck.countTrailingZeros();
2855 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2856 ConstantInt::get(X->getType(), CmpVal));
2857 } else if ((++ValToCheck).isPowerOf2()) {
2858 unsigned CmpVal = ValToCheck.countTrailingZeros();
2859 return new ICmpInst(ICmpInst::ICMP_ULT, X,
2860 ConstantInt::get(X->getType(), CmpVal));
2864 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2865 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2867 if (Op0KnownZeroInverted == 1 &&
2868 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2869 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2870 ConstantInt::get(X->getType(),
2871 CI->countTrailingZeros()));
2876 case ICmpInst::ICMP_ULT:
2877 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2878 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2879 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2880 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2881 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2882 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2883 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2884 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2885 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2886 Builder->getInt(CI->getValue()-1));
2888 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2889 if (CI->isMinValue(true))
2890 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2891 Constant::getAllOnesValue(Op0->getType()));
2894 case ICmpInst::ICMP_UGT:
2895 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2896 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2897 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2898 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2900 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2901 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2902 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2903 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2904 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2905 Builder->getInt(CI->getValue()+1));
2907 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2908 if (CI->isMaxValue(true))
2909 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2910 Constant::getNullValue(Op0->getType()));
2913 case ICmpInst::ICMP_SLT:
2914 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2915 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2916 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2917 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2918 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2919 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2920 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2921 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2922 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2923 Builder->getInt(CI->getValue()-1));
2926 case ICmpInst::ICMP_SGT:
2927 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2928 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2929 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2930 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2932 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2933 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2934 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2935 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2936 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2937 Builder->getInt(CI->getValue()+1));
2940 case ICmpInst::ICMP_SGE:
2941 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2942 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2943 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2944 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2945 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2947 case ICmpInst::ICMP_SLE:
2948 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2949 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2950 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2951 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2952 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2954 case ICmpInst::ICMP_UGE:
2955 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2956 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2957 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2958 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2959 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2961 case ICmpInst::ICMP_ULE:
2962 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2963 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2964 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2965 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2966 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2970 // Turn a signed comparison into an unsigned one if both operands
2971 // are known to have the same sign.
2973 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2974 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2975 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2978 // Test if the ICmpInst instruction is used exclusively by a select as
2979 // part of a minimum or maximum operation. If so, refrain from doing
2980 // any other folding. This helps out other analyses which understand
2981 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2982 // and CodeGen. And in this case, at least one of the comparison
2983 // operands has at least one user besides the compare (the select),
2984 // which would often largely negate the benefit of folding anyway.
2986 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
2987 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2988 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2991 // See if we are doing a comparison between a constant and an instruction that
2992 // can be folded into the comparison.
2993 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2994 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2995 // instruction, see if that instruction also has constants so that the
2996 // instruction can be folded into the icmp
2997 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2998 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
3002 // Handle icmp with constant (but not simple integer constant) RHS
3003 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3004 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3005 switch (LHSI->getOpcode()) {
3006 case Instruction::GetElementPtr:
3007 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3008 if (RHSC->isNullValue() &&
3009 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3010 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
3011 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3013 case Instruction::PHI:
3014 // Only fold icmp into the PHI if the phi and icmp are in the same
3015 // block. If in the same block, we're encouraging jump threading. If
3016 // not, we are just pessimizing the code by making an i1 phi.
3017 if (LHSI->getParent() == I.getParent())
3018 if (Instruction *NV = FoldOpIntoPhi(I))
3021 case Instruction::Select: {
3022 // If either operand of the select is a constant, we can fold the
3023 // comparison into the select arms, which will cause one to be
3024 // constant folded and the select turned into a bitwise or.
3025 Value *Op1 = nullptr, *Op2 = nullptr;
3026 ConstantInt *CI = 0;
3027 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3028 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3029 CI = dyn_cast<ConstantInt>(Op1);
3031 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3032 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3033 CI = dyn_cast<ConstantInt>(Op2);
3036 // We only want to perform this transformation if it will not lead to
3037 // additional code. This is true if either both sides of the select
3038 // fold to a constant (in which case the icmp is replaced with a select
3039 // which will usually simplify) or this is the only user of the
3040 // select (in which case we are trading a select+icmp for a simpler
3041 // select+icmp) or all uses of the select can be replaced based on
3042 // dominance information ("Global cases").
3043 bool Transform = false;
3046 else if (Op1 || Op2) {
3048 if (LHSI->hasOneUse())
3051 else if (CI && !CI->isZero())
3052 // When Op1 is constant try replacing select with second operand.
3053 // Otherwise Op2 is constant and try replacing select with first
3055 Transform = replacedSelectWithOperand(cast<SelectInst>(LHSI), &I,
3060 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
3063 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
3065 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3069 case Instruction::IntToPtr:
3070 // icmp pred inttoptr(X), null -> icmp pred X, 0
3071 if (RHSC->isNullValue() && DL &&
3072 DL->getIntPtrType(RHSC->getType()) ==
3073 LHSI->getOperand(0)->getType())
3074 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
3075 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3078 case Instruction::Load:
3079 // Try to optimize things like "A[i] > 4" to index computations.
3080 if (GetElementPtrInst *GEP =
3081 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3082 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3083 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3084 !cast<LoadInst>(LHSI)->isVolatile())
3085 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3092 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
3093 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
3094 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
3096 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
3097 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
3098 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
3101 // Test to see if the operands of the icmp are casted versions of other
3102 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
3104 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
3105 if (Op0->getType()->isPointerTy() &&
3106 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
3107 // We keep moving the cast from the left operand over to the right
3108 // operand, where it can often be eliminated completely.
3109 Op0 = CI->getOperand(0);
3111 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
3112 // so eliminate it as well.
3113 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
3114 Op1 = CI2->getOperand(0);
3116 // If Op1 is a constant, we can fold the cast into the constant.
3117 if (Op0->getType() != Op1->getType()) {
3118 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3119 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
3121 // Otherwise, cast the RHS right before the icmp
3122 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
3125 return new ICmpInst(I.getPredicate(), Op0, Op1);
3129 if (isa<CastInst>(Op0)) {
3130 // Handle the special case of: icmp (cast bool to X), <cst>
3131 // This comes up when you have code like
3134 // For generality, we handle any zero-extension of any operand comparison
3135 // with a constant or another cast from the same type.
3136 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
3137 if (Instruction *R = visitICmpInstWithCastAndCast(I))
3141 // Special logic for binary operators.
3142 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3143 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3145 CmpInst::Predicate Pred = I.getPredicate();
3146 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3147 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3148 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
3149 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3150 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3151 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3152 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
3153 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3154 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3156 // Analyze the case when either Op0 or Op1 is an add instruction.
3157 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3158 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3159 if (BO0 && BO0->getOpcode() == Instruction::Add)
3160 A = BO0->getOperand(0), B = BO0->getOperand(1);
3161 if (BO1 && BO1->getOpcode() == Instruction::Add)
3162 C = BO1->getOperand(0), D = BO1->getOperand(1);
3164 // icmp (X+cst) < 0 --> X < -cst
3165 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred) && match(Op1, m_Zero()))
3166 if (ConstantInt *RHSC = dyn_cast_or_null<ConstantInt>(B))
3167 if (!RHSC->isMinValue(/*isSigned=*/true))
3168 return new ICmpInst(Pred, A, ConstantExpr::getNeg(RHSC));
3170 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3171 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3172 return new ICmpInst(Pred, A == Op1 ? B : A,
3173 Constant::getNullValue(Op1->getType()));
3175 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3176 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3177 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3180 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3181 if (A && C && (A == C || A == D || B == C || B == D) &&
3182 NoOp0WrapProblem && NoOp1WrapProblem &&
3183 // Try not to increase register pressure.
3184 BO0->hasOneUse() && BO1->hasOneUse()) {
3185 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3188 // C + B == C + D -> B == D
3191 } else if (A == D) {
3192 // D + B == C + D -> B == C
3195 } else if (B == C) {
3196 // A + C == C + D -> A == D
3201 // A + D == C + D -> A == C
3205 return new ICmpInst(Pred, Y, Z);
3208 // icmp slt (X + -1), Y -> icmp sle X, Y
3209 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3210 match(B, m_AllOnes()))
3211 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3213 // icmp sge (X + -1), Y -> icmp sgt X, Y
3214 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3215 match(B, m_AllOnes()))
3216 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3218 // icmp sle (X + 1), Y -> icmp slt X, Y
3219 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
3221 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3223 // icmp sgt (X + 1), Y -> icmp sge X, Y
3224 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
3226 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3228 // if C1 has greater magnitude than C2:
3229 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3230 // s.t. C3 = C1 - C2
3232 // if C2 has greater magnitude than C1:
3233 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3234 // s.t. C3 = C2 - C1
3235 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3236 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3237 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3238 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3239 const APInt &AP1 = C1->getValue();
3240 const APInt &AP2 = C2->getValue();
3241 if (AP1.isNegative() == AP2.isNegative()) {
3242 APInt AP1Abs = C1->getValue().abs();
3243 APInt AP2Abs = C2->getValue().abs();
3244 if (AP1Abs.uge(AP2Abs)) {
3245 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
3246 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
3247 return new ICmpInst(Pred, NewAdd, C);
3249 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
3250 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
3251 return new ICmpInst(Pred, A, NewAdd);
3257 // Analyze the case when either Op0 or Op1 is a sub instruction.
3258 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3259 A = nullptr; B = nullptr; C = nullptr; D = nullptr;
3260 if (BO0 && BO0->getOpcode() == Instruction::Sub)
3261 A = BO0->getOperand(0), B = BO0->getOperand(1);
3262 if (BO1 && BO1->getOpcode() == Instruction::Sub)
3263 C = BO1->getOperand(0), D = BO1->getOperand(1);
3265 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3266 if (A == Op1 && NoOp0WrapProblem)
3267 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3269 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3270 if (C == Op0 && NoOp1WrapProblem)
3271 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3273 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3274 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3275 // Try not to increase register pressure.
3276 BO0->hasOneUse() && BO1->hasOneUse())
3277 return new ICmpInst(Pred, A, C);
3279 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3280 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3281 // Try not to increase register pressure.
3282 BO0->hasOneUse() && BO1->hasOneUse())
3283 return new ICmpInst(Pred, D, B);
3285 // icmp (0-X) < cst --> x > -cst
3286 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3288 if (match(BO0, m_Neg(m_Value(X))))
3289 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3290 if (!RHSC->isMinValue(/*isSigned=*/true))
3291 return new ICmpInst(I.getSwappedPredicate(), X,
3292 ConstantExpr::getNeg(RHSC));
3295 BinaryOperator *SRem = nullptr;
3296 // icmp (srem X, Y), Y
3297 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
3298 Op1 == BO0->getOperand(1))
3300 // icmp Y, (srem X, Y)
3301 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3302 Op0 == BO1->getOperand(1))
3305 // We don't check hasOneUse to avoid increasing register pressure because
3306 // the value we use is the same value this instruction was already using.
3307 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3309 case ICmpInst::ICMP_EQ:
3310 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3311 case ICmpInst::ICMP_NE:
3312 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3313 case ICmpInst::ICMP_SGT:
3314 case ICmpInst::ICMP_SGE:
3315 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3316 Constant::getAllOnesValue(SRem->getType()));
3317 case ICmpInst::ICMP_SLT:
3318 case ICmpInst::ICMP_SLE:
3319 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3320 Constant::getNullValue(SRem->getType()));
3324 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
3325 BO0->hasOneUse() && BO1->hasOneUse() &&
3326 BO0->getOperand(1) == BO1->getOperand(1)) {
3327 switch (BO0->getOpcode()) {
3329 case Instruction::Add:
3330 case Instruction::Sub:
3331 case Instruction::Xor:
3332 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3333 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3334 BO1->getOperand(0));
3335 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
3336 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3337 if (CI->getValue().isSignBit()) {
3338 ICmpInst::Predicate Pred = I.isSigned()
3339 ? I.getUnsignedPredicate()
3340 : I.getSignedPredicate();
3341 return new ICmpInst(Pred, BO0->getOperand(0),
3342 BO1->getOperand(0));
3345 if (CI->isMaxValue(true)) {
3346 ICmpInst::Predicate Pred = I.isSigned()
3347 ? I.getUnsignedPredicate()
3348 : I.getSignedPredicate();
3349 Pred = I.getSwappedPredicate(Pred);
3350 return new ICmpInst(Pred, BO0->getOperand(0),
3351 BO1->getOperand(0));
3355 case Instruction::Mul:
3356 if (!I.isEquality())
3359 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3360 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
3361 // Mask = -1 >> count-trailing-zeros(Cst).
3362 if (!CI->isZero() && !CI->isOne()) {
3363 const APInt &AP = CI->getValue();
3364 ConstantInt *Mask = ConstantInt::get(I.getContext(),
3365 APInt::getLowBitsSet(AP.getBitWidth(),
3367 AP.countTrailingZeros()));
3368 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
3369 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
3370 return new ICmpInst(I.getPredicate(), And1, And2);
3374 case Instruction::UDiv:
3375 case Instruction::LShr:
3379 case Instruction::SDiv:
3380 case Instruction::AShr:
3381 if (!BO0->isExact() || !BO1->isExact())
3383 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3384 BO1->getOperand(0));
3385 case Instruction::Shl: {
3386 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3387 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3390 if (!NSW && I.isSigned())
3392 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3393 BO1->getOperand(0));
3400 // Transform (A & ~B) == 0 --> (A & B) != 0
3401 // and (A & ~B) != 0 --> (A & B) == 0
3402 // if A is a power of 2.
3403 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
3404 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A, false,
3407 return new ICmpInst(I.getInversePredicate(),
3408 Builder->CreateAnd(A, B),
3411 // ~x < ~y --> y < x
3412 // ~x < cst --> ~cst < x
3413 if (match(Op0, m_Not(m_Value(A)))) {
3414 if (match(Op1, m_Not(m_Value(B))))
3415 return new ICmpInst(I.getPredicate(), B, A);
3416 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3417 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
3420 // (a+b) <u a --> llvm.uadd.with.overflow.
3421 // (a+b) <u b --> llvm.uadd.with.overflow.
3422 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
3423 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
3424 (Op1 == A || Op1 == B))
3425 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
3428 // a >u (a+b) --> llvm.uadd.with.overflow.
3429 // b >u (a+b) --> llvm.uadd.with.overflow.
3430 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
3431 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
3432 (Op0 == A || Op0 == B))
3433 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
3436 // (zext a) * (zext b) --> llvm.umul.with.overflow.
3437 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3438 if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
3441 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3442 if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
3447 if (I.isEquality()) {
3448 Value *A, *B, *C, *D;
3450 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3451 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3452 Value *OtherVal = A == Op1 ? B : A;
3453 return new ICmpInst(I.getPredicate(), OtherVal,
3454 Constant::getNullValue(A->getType()));
3457 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3458 // A^c1 == C^c2 --> A == C^(c1^c2)
3459 ConstantInt *C1, *C2;
3460 if (match(B, m_ConstantInt(C1)) &&
3461 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
3462 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3463 Value *Xor = Builder->CreateXor(C, NC);
3464 return new ICmpInst(I.getPredicate(), A, Xor);
3467 // A^B == A^D -> B == D
3468 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
3469 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
3470 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
3471 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
3475 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
3476 (A == Op0 || B == Op0)) {
3477 // A == (A^B) -> B == 0
3478 Value *OtherVal = A == Op0 ? B : A;
3479 return new ICmpInst(I.getPredicate(), OtherVal,
3480 Constant::getNullValue(A->getType()));
3483 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3484 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3485 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3486 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3489 X = B; Y = D; Z = A;
3490 } else if (A == D) {
3491 X = B; Y = C; Z = A;
3492 } else if (B == C) {
3493 X = A; Y = D; Z = B;
3494 } else if (B == D) {
3495 X = A; Y = C; Z = B;
3498 if (X) { // Build (X^Y) & Z
3499 Op1 = Builder->CreateXor(X, Y);
3500 Op1 = Builder->CreateAnd(Op1, Z);
3501 I.setOperand(0, Op1);
3502 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3507 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3508 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3510 if ((Op0->hasOneUse() &&
3511 match(Op0, m_ZExt(m_Value(A))) &&
3512 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3513 (Op1->hasOneUse() &&
3514 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3515 match(Op1, m_ZExt(m_Value(A))))) {
3516 APInt Pow2 = Cst1->getValue() + 1;
3517 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3518 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3519 return new ICmpInst(I.getPredicate(), A,
3520 Builder->CreateTrunc(B, A->getType()));
3523 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3524 // For lshr and ashr pairs.
3525 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3526 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3527 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3528 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3529 unsigned TypeBits = Cst1->getBitWidth();
3530 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3531 if (ShAmt < TypeBits && ShAmt != 0) {
3532 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
3533 ? ICmpInst::ICMP_UGE
3534 : ICmpInst::ICMP_ULT;
3535 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3536 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3537 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
3541 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3542 // "icmp (and X, mask), cst"
3544 if (Op0->hasOneUse() &&
3545 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
3546 m_ConstantInt(ShAmt))))) &&
3547 match(Op1, m_ConstantInt(Cst1)) &&
3548 // Only do this when A has multiple uses. This is most important to do
3549 // when it exposes other optimizations.
3551 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3553 if (ShAmt < ASize) {
3555 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3558 APInt CmpV = Cst1->getValue().zext(ASize);
3561 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3562 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3567 // The 'cmpxchg' instruction returns an aggregate containing the old value and
3568 // an i1 which indicates whether or not we successfully did the swap.
3570 // Replace comparisons between the old value and the expected value with the
3571 // indicator that 'cmpxchg' returns.
3573 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
3574 // spuriously fail. In those cases, the old value may equal the expected
3575 // value but it is possible for the swap to not occur.
3576 if (I.getPredicate() == ICmpInst::ICMP_EQ)
3577 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
3578 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
3579 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
3581 return ExtractValueInst::Create(ACXI, 1);
3584 Value *X; ConstantInt *Cst;
3586 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3587 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3590 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3591 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3593 return Changed ? &I : nullptr;
3596 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3598 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3601 if (!isa<ConstantFP>(RHSC)) return nullptr;
3602 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3604 // Get the width of the mantissa. We don't want to hack on conversions that
3605 // might lose information from the integer, e.g. "i64 -> float"
3606 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3607 if (MantissaWidth == -1) return nullptr; // Unknown.
3609 // Check to see that the input is converted from an integer type that is small
3610 // enough that preserves all bits. TODO: check here for "known" sign bits.
3611 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3612 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
3614 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3615 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3619 // If the conversion would lose info, don't hack on this.
3620 if ((int)InputSize > MantissaWidth)
3623 // Otherwise, we can potentially simplify the comparison. We know that it
3624 // will always come through as an integer value and we know the constant is
3625 // not a NAN (it would have been previously simplified).
3626 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3628 ICmpInst::Predicate Pred;
3629 switch (I.getPredicate()) {
3630 default: llvm_unreachable("Unexpected predicate!");
3631 case FCmpInst::FCMP_UEQ:
3632 case FCmpInst::FCMP_OEQ:
3633 Pred = ICmpInst::ICMP_EQ;
3635 case FCmpInst::FCMP_UGT:
3636 case FCmpInst::FCMP_OGT:
3637 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3639 case FCmpInst::FCMP_UGE:
3640 case FCmpInst::FCMP_OGE:
3641 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3643 case FCmpInst::FCMP_ULT:
3644 case FCmpInst::FCMP_OLT:
3645 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3647 case FCmpInst::FCMP_ULE:
3648 case FCmpInst::FCMP_OLE:
3649 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3651 case FCmpInst::FCMP_UNE:
3652 case FCmpInst::FCMP_ONE:
3653 Pred = ICmpInst::ICMP_NE;
3655 case FCmpInst::FCMP_ORD:
3656 return ReplaceInstUsesWith(I, Builder->getTrue());
3657 case FCmpInst::FCMP_UNO:
3658 return ReplaceInstUsesWith(I, Builder->getFalse());
3661 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3663 // Now we know that the APFloat is a normal number, zero or inf.
3665 // See if the FP constant is too large for the integer. For example,
3666 // comparing an i8 to 300.0.
3667 unsigned IntWidth = IntTy->getScalarSizeInBits();
3670 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3671 // and large values.
3672 APFloat SMax(RHS.getSemantics());
3673 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3674 APFloat::rmNearestTiesToEven);
3675 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3676 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3677 Pred == ICmpInst::ICMP_SLE)
3678 return ReplaceInstUsesWith(I, Builder->getTrue());
3679 return ReplaceInstUsesWith(I, Builder->getFalse());
3682 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3683 // +INF and large values.
3684 APFloat UMax(RHS.getSemantics());
3685 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3686 APFloat::rmNearestTiesToEven);
3687 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3688 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3689 Pred == ICmpInst::ICMP_ULE)
3690 return ReplaceInstUsesWith(I, Builder->getTrue());
3691 return ReplaceInstUsesWith(I, Builder->getFalse());
3696 // See if the RHS value is < SignedMin.
3697 APFloat SMin(RHS.getSemantics());
3698 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3699 APFloat::rmNearestTiesToEven);
3700 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3701 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3702 Pred == ICmpInst::ICMP_SGE)
3703 return ReplaceInstUsesWith(I, Builder->getTrue());
3704 return ReplaceInstUsesWith(I, Builder->getFalse());
3707 // See if the RHS value is < UnsignedMin.
3708 APFloat SMin(RHS.getSemantics());
3709 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3710 APFloat::rmNearestTiesToEven);
3711 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3712 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3713 Pred == ICmpInst::ICMP_UGE)
3714 return ReplaceInstUsesWith(I, Builder->getTrue());
3715 return ReplaceInstUsesWith(I, Builder->getFalse());
3719 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3720 // [0, UMAX], but it may still be fractional. See if it is fractional by
3721 // casting the FP value to the integer value and back, checking for equality.
3722 // Don't do this for zero, because -0.0 is not fractional.
3723 Constant *RHSInt = LHSUnsigned
3724 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3725 : ConstantExpr::getFPToSI(RHSC, IntTy);
3726 if (!RHS.isZero()) {
3727 bool Equal = LHSUnsigned
3728 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3729 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3731 // If we had a comparison against a fractional value, we have to adjust
3732 // the compare predicate and sometimes the value. RHSC is rounded towards
3733 // zero at this point.
3735 default: llvm_unreachable("Unexpected integer comparison!");
3736 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3737 return ReplaceInstUsesWith(I, Builder->getTrue());
3738 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3739 return ReplaceInstUsesWith(I, Builder->getFalse());
3740 case ICmpInst::ICMP_ULE:
3741 // (float)int <= 4.4 --> int <= 4
3742 // (float)int <= -4.4 --> false
3743 if (RHS.isNegative())
3744 return ReplaceInstUsesWith(I, Builder->getFalse());
3746 case ICmpInst::ICMP_SLE:
3747 // (float)int <= 4.4 --> int <= 4
3748 // (float)int <= -4.4 --> int < -4
3749 if (RHS.isNegative())
3750 Pred = ICmpInst::ICMP_SLT;
3752 case ICmpInst::ICMP_ULT:
3753 // (float)int < -4.4 --> false
3754 // (float)int < 4.4 --> int <= 4
3755 if (RHS.isNegative())
3756 return ReplaceInstUsesWith(I, Builder->getFalse());
3757 Pred = ICmpInst::ICMP_ULE;
3759 case ICmpInst::ICMP_SLT:
3760 // (float)int < -4.4 --> int < -4
3761 // (float)int < 4.4 --> int <= 4
3762 if (!RHS.isNegative())
3763 Pred = ICmpInst::ICMP_SLE;
3765 case ICmpInst::ICMP_UGT:
3766 // (float)int > 4.4 --> int > 4
3767 // (float)int > -4.4 --> true
3768 if (RHS.isNegative())
3769 return ReplaceInstUsesWith(I, Builder->getTrue());
3771 case ICmpInst::ICMP_SGT:
3772 // (float)int > 4.4 --> int > 4
3773 // (float)int > -4.4 --> int >= -4
3774 if (RHS.isNegative())
3775 Pred = ICmpInst::ICMP_SGE;
3777 case ICmpInst::ICMP_UGE:
3778 // (float)int >= -4.4 --> true
3779 // (float)int >= 4.4 --> int > 4
3780 if (RHS.isNegative())
3781 return ReplaceInstUsesWith(I, Builder->getTrue());
3782 Pred = ICmpInst::ICMP_UGT;
3784 case ICmpInst::ICMP_SGE:
3785 // (float)int >= -4.4 --> int >= -4
3786 // (float)int >= 4.4 --> int > 4
3787 if (!RHS.isNegative())
3788 Pred = ICmpInst::ICMP_SGT;
3794 // Lower this FP comparison into an appropriate integer version of the
3796 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3799 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3800 bool Changed = false;
3802 /// Orders the operands of the compare so that they are listed from most
3803 /// complex to least complex. This puts constants before unary operators,
3804 /// before binary operators.
3805 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3810 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3812 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL, TLI, DT, AT))
3813 return ReplaceInstUsesWith(I, V);
3815 // Simplify 'fcmp pred X, X'
3817 switch (I.getPredicate()) {
3818 default: llvm_unreachable("Unknown predicate!");
3819 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3820 case FCmpInst::FCMP_ULT: // True if unordered or less than
3821 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3822 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3823 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3824 I.setPredicate(FCmpInst::FCMP_UNO);
3825 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3828 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3829 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3830 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3831 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3832 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3833 I.setPredicate(FCmpInst::FCMP_ORD);
3834 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3839 // Handle fcmp with constant RHS
3840 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3841 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3842 switch (LHSI->getOpcode()) {
3843 case Instruction::FPExt: {
3844 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3845 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3846 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3850 const fltSemantics *Sem;
3851 // FIXME: This shouldn't be here.
3852 if (LHSExt->getSrcTy()->isHalfTy())
3853 Sem = &APFloat::IEEEhalf;
3854 else if (LHSExt->getSrcTy()->isFloatTy())
3855 Sem = &APFloat::IEEEsingle;
3856 else if (LHSExt->getSrcTy()->isDoubleTy())
3857 Sem = &APFloat::IEEEdouble;
3858 else if (LHSExt->getSrcTy()->isFP128Ty())
3859 Sem = &APFloat::IEEEquad;
3860 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3861 Sem = &APFloat::x87DoubleExtended;
3862 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3863 Sem = &APFloat::PPCDoubleDouble;
3868 APFloat F = RHSF->getValueAPF();
3869 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3871 // Avoid lossy conversions and denormals. Zero is a special case
3872 // that's OK to convert.
3876 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3877 APFloat::cmpLessThan) || Fabs.isZero()))
3879 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3880 ConstantFP::get(RHSC->getContext(), F));
3883 case Instruction::PHI:
3884 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3885 // block. If in the same block, we're encouraging jump threading. If
3886 // not, we are just pessimizing the code by making an i1 phi.
3887 if (LHSI->getParent() == I.getParent())
3888 if (Instruction *NV = FoldOpIntoPhi(I))
3891 case Instruction::SIToFP:
3892 case Instruction::UIToFP:
3893 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3896 case Instruction::FSub: {
3897 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3899 if (match(LHSI, m_FNeg(m_Value(Op))))
3900 return new FCmpInst(I.getSwappedPredicate(), Op,
3901 ConstantExpr::getFNeg(RHSC));
3904 case Instruction::Load:
3905 if (GetElementPtrInst *GEP =
3906 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3907 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3908 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3909 !cast<LoadInst>(LHSI)->isVolatile())
3910 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3914 case Instruction::Call: {
3915 CallInst *CI = cast<CallInst>(LHSI);
3917 // Various optimization for fabs compared with zero.
3918 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3919 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3921 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3922 Func == LibFunc::fabsl) {
3923 switch (I.getPredicate()) {
3925 // fabs(x) < 0 --> false
3926 case FCmpInst::FCMP_OLT:
3927 return ReplaceInstUsesWith(I, Builder->getFalse());
3928 // fabs(x) > 0 --> x != 0
3929 case FCmpInst::FCMP_OGT:
3930 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3932 // fabs(x) <= 0 --> x == 0
3933 case FCmpInst::FCMP_OLE:
3934 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3936 // fabs(x) >= 0 --> !isnan(x)
3937 case FCmpInst::FCMP_OGE:
3938 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3940 // fabs(x) == 0 --> x == 0
3941 // fabs(x) != 0 --> x != 0
3942 case FCmpInst::FCMP_OEQ:
3943 case FCmpInst::FCMP_UEQ:
3944 case FCmpInst::FCMP_ONE:
3945 case FCmpInst::FCMP_UNE:
3946 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3955 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3957 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3958 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3960 // fcmp (fpext x), (fpext y) -> fcmp x, y
3961 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3962 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3963 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3964 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3965 RHSExt->getOperand(0));
3967 return Changed ? &I : nullptr;