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 #define DEBUG_TYPE "instcombine"
15 #include "InstCombine.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/Target/TargetLibraryInfo.h"
26 using namespace PatternMatch;
28 static ConstantInt *getOne(Constant *C) {
29 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
32 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
33 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
36 static bool HasAddOverflow(ConstantInt *Result,
37 ConstantInt *In1, ConstantInt *In2,
40 return Result->getValue().ult(In1->getValue());
42 if (In2->isNegative())
43 return Result->getValue().sgt(In1->getValue());
44 return Result->getValue().slt(In1->getValue());
47 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
48 /// overflowed for this type.
49 static bool AddWithOverflow(Constant *&Result, Constant *In1,
50 Constant *In2, bool IsSigned = false) {
51 Result = ConstantExpr::getAdd(In1, In2);
53 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
54 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
55 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
56 if (HasAddOverflow(ExtractElement(Result, Idx),
57 ExtractElement(In1, Idx),
58 ExtractElement(In2, Idx),
65 return HasAddOverflow(cast<ConstantInt>(Result),
66 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
70 static bool HasSubOverflow(ConstantInt *Result,
71 ConstantInt *In1, ConstantInt *In2,
74 return Result->getValue().ugt(In1->getValue());
76 if (In2->isNegative())
77 return Result->getValue().slt(In1->getValue());
79 return Result->getValue().sgt(In1->getValue());
82 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
83 /// overflowed for this type.
84 static bool SubWithOverflow(Constant *&Result, Constant *In1,
85 Constant *In2, bool IsSigned = false) {
86 Result = ConstantExpr::getSub(In1, In2);
88 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
89 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
90 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
91 if (HasSubOverflow(ExtractElement(Result, Idx),
92 ExtractElement(In1, Idx),
93 ExtractElement(In2, Idx),
100 return HasSubOverflow(cast<ConstantInt>(Result),
101 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
105 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
106 /// comparison only checks the sign bit. If it only checks the sign bit, set
107 /// TrueIfSigned if the result of the comparison is true when the input value is
109 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
110 bool &TrueIfSigned) {
112 case ICmpInst::ICMP_SLT: // True if LHS s< 0
114 return RHS->isZero();
115 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
117 return RHS->isAllOnesValue();
118 case ICmpInst::ICMP_SGT: // True if LHS s> -1
119 TrueIfSigned = false;
120 return RHS->isAllOnesValue();
121 case ICmpInst::ICMP_UGT:
122 // True if LHS u> RHS and RHS == high-bit-mask - 1
124 return RHS->isMaxValue(true);
125 case ICmpInst::ICMP_UGE:
126 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
128 return RHS->getValue().isSignBit();
134 /// Returns true if the exploded icmp can be expressed as a signed comparison
135 /// to zero and updates the predicate accordingly.
136 /// The signedness of the comparison is preserved.
137 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) {
138 if (!ICmpInst::isSigned(pred))
142 return ICmpInst::isRelational(pred);
145 if (pred == ICmpInst::ICMP_SLT) {
146 pred = ICmpInst::ICMP_SLE;
149 } else if (RHS->isAllOnesValue()) {
150 if (pred == ICmpInst::ICMP_SGT) {
151 pred = ICmpInst::ICMP_SGE;
159 // isHighOnes - Return true if the constant is of the form 1+0+.
160 // This is the same as lowones(~X).
161 static bool isHighOnes(const ConstantInt *CI) {
162 return (~CI->getValue() + 1).isPowerOf2();
165 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
166 /// set of known zero and one bits, compute the maximum and minimum values that
167 /// could have the specified known zero and known one bits, returning them in
169 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
170 const APInt& KnownOne,
171 APInt& Min, APInt& Max) {
172 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
173 KnownZero.getBitWidth() == Min.getBitWidth() &&
174 KnownZero.getBitWidth() == Max.getBitWidth() &&
175 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
176 APInt UnknownBits = ~(KnownZero|KnownOne);
178 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
179 // bit if it is unknown.
181 Max = KnownOne|UnknownBits;
183 if (UnknownBits.isNegative()) { // Sign bit is unknown
184 Min.setBit(Min.getBitWidth()-1);
185 Max.clearBit(Max.getBitWidth()-1);
189 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
190 // a set of known zero and one bits, compute the maximum and minimum values that
191 // could have the specified known zero and known one bits, returning them in
193 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
194 const APInt &KnownOne,
195 APInt &Min, APInt &Max) {
196 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
197 KnownZero.getBitWidth() == Min.getBitWidth() &&
198 KnownZero.getBitWidth() == Max.getBitWidth() &&
199 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
200 APInt UnknownBits = ~(KnownZero|KnownOne);
202 // The minimum value is when the unknown bits are all zeros.
204 // The maximum value is when the unknown bits are all ones.
205 Max = KnownOne|UnknownBits;
210 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
211 /// cmp pred (load (gep GV, ...)), cmpcst
212 /// where GV is a global variable with a constant initializer. Try to simplify
213 /// this into some simple computation that does not need the load. For example
214 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
216 /// If AndCst is non-null, then the loaded value is masked with that constant
217 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
218 Instruction *InstCombiner::
219 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
220 CmpInst &ICI, ConstantInt *AndCst) {
221 // We need TD information to know the pointer size unless this is inbounds.
222 if (!GEP->isInBounds() && DL == 0)
225 Constant *Init = GV->getInitializer();
226 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
229 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
230 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays.
232 // There are many forms of this optimization we can handle, for now, just do
233 // the simple index into a single-dimensional array.
235 // Require: GEP GV, 0, i {{, constant indices}}
236 if (GEP->getNumOperands() < 3 ||
237 !isa<ConstantInt>(GEP->getOperand(1)) ||
238 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
239 isa<Constant>(GEP->getOperand(2)))
242 // Check that indices after the variable are constants and in-range for the
243 // type they index. Collect the indices. This is typically for arrays of
245 SmallVector<unsigned, 4> LaterIndices;
247 Type *EltTy = Init->getType()->getArrayElementType();
248 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
249 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
250 if (Idx == 0) return 0; // Variable index.
252 uint64_t IdxVal = Idx->getZExtValue();
253 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
255 if (StructType *STy = dyn_cast<StructType>(EltTy))
256 EltTy = STy->getElementType(IdxVal);
257 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
258 if (IdxVal >= ATy->getNumElements()) return 0;
259 EltTy = ATy->getElementType();
261 return 0; // Unknown type.
264 LaterIndices.push_back(IdxVal);
267 enum { Overdefined = -3, Undefined = -2 };
269 // Variables for our state machines.
271 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
272 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
273 // and 87 is the second (and last) index. FirstTrueElement is -2 when
274 // undefined, otherwise set to the first true element. SecondTrueElement is
275 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
276 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
278 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
279 // form "i != 47 & i != 87". Same state transitions as for true elements.
280 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
282 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
283 /// define a state machine that triggers for ranges of values that the index
284 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
285 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
286 /// index in the range (inclusive). We use -2 for undefined here because we
287 /// use relative comparisons and don't want 0-1 to match -1.
288 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
290 // MagicBitvector - This is a magic bitvector where we set a bit if the
291 // comparison is true for element 'i'. If there are 64 elements or less in
292 // the array, this will fully represent all the comparison results.
293 uint64_t MagicBitvector = 0;
296 // Scan the array and see if one of our patterns matches.
297 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
298 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
299 Constant *Elt = Init->getAggregateElement(i);
300 if (Elt == 0) return 0;
302 // If this is indexing an array of structures, get the structure element.
303 if (!LaterIndices.empty())
304 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
306 // If the element is masked, handle it.
307 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
309 // Find out if the comparison would be true or false for the i'th element.
310 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
311 CompareRHS, DL, TLI);
312 // If the result is undef for this element, ignore it.
313 if (isa<UndefValue>(C)) {
314 // Extend range state machines to cover this element in case there is an
315 // undef in the middle of the range.
316 if (TrueRangeEnd == (int)i-1)
318 if (FalseRangeEnd == (int)i-1)
323 // If we can't compute the result for any of the elements, we have to give
324 // up evaluating the entire conditional.
325 if (!isa<ConstantInt>(C)) return 0;
327 // Otherwise, we know if the comparison is true or false for this element,
328 // update our state machines.
329 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
331 // State machine for single/double/range index comparison.
333 // Update the TrueElement state machine.
334 if (FirstTrueElement == Undefined)
335 FirstTrueElement = TrueRangeEnd = i; // First true element.
337 // Update double-compare state machine.
338 if (SecondTrueElement == Undefined)
339 SecondTrueElement = i;
341 SecondTrueElement = Overdefined;
343 // Update range state machine.
344 if (TrueRangeEnd == (int)i-1)
347 TrueRangeEnd = Overdefined;
350 // Update the FalseElement state machine.
351 if (FirstFalseElement == Undefined)
352 FirstFalseElement = FalseRangeEnd = i; // First false element.
354 // Update double-compare state machine.
355 if (SecondFalseElement == Undefined)
356 SecondFalseElement = i;
358 SecondFalseElement = Overdefined;
360 // Update range state machine.
361 if (FalseRangeEnd == (int)i-1)
364 FalseRangeEnd = Overdefined;
369 // If this element is in range, update our magic bitvector.
370 if (i < 64 && IsTrueForElt)
371 MagicBitvector |= 1ULL << i;
373 // If all of our states become overdefined, bail out early. Since the
374 // predicate is expensive, only check it every 8 elements. This is only
375 // really useful for really huge arrays.
376 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
377 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
378 FalseRangeEnd == Overdefined)
382 // Now that we've scanned the entire array, emit our new comparison(s). We
383 // order the state machines in complexity of the generated code.
384 Value *Idx = GEP->getOperand(2);
386 // If the index is larger than the pointer size of the target, truncate the
387 // index down like the GEP would do implicitly. We don't have to do this for
388 // an inbounds GEP because the index can't be out of range.
389 if (!GEP->isInBounds()) {
390 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
391 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
392 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
393 Idx = Builder->CreateTrunc(Idx, IntPtrTy);
396 // If the comparison is only true for one or two elements, emit direct
398 if (SecondTrueElement != Overdefined) {
399 // None true -> false.
400 if (FirstTrueElement == Undefined)
401 return ReplaceInstUsesWith(ICI, Builder->getFalse());
403 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
405 // True for one element -> 'i == 47'.
406 if (SecondTrueElement == Undefined)
407 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
409 // True for two elements -> 'i == 47 | i == 72'.
410 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
411 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
412 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
413 return BinaryOperator::CreateOr(C1, C2);
416 // If the comparison is only false for one or two elements, emit direct
418 if (SecondFalseElement != Overdefined) {
419 // None false -> true.
420 if (FirstFalseElement == Undefined)
421 return ReplaceInstUsesWith(ICI, Builder->getTrue());
423 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
425 // False for one element -> 'i != 47'.
426 if (SecondFalseElement == Undefined)
427 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
429 // False for two elements -> 'i != 47 & i != 72'.
430 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
431 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
432 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
433 return BinaryOperator::CreateAnd(C1, C2);
436 // If the comparison can be replaced with a range comparison for the elements
437 // where it is true, emit the range check.
438 if (TrueRangeEnd != Overdefined) {
439 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
441 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
442 if (FirstTrueElement) {
443 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
444 Idx = Builder->CreateAdd(Idx, Offs);
447 Value *End = ConstantInt::get(Idx->getType(),
448 TrueRangeEnd-FirstTrueElement+1);
449 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
452 // False range check.
453 if (FalseRangeEnd != Overdefined) {
454 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
455 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
456 if (FirstFalseElement) {
457 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
458 Idx = Builder->CreateAdd(Idx, Offs);
461 Value *End = ConstantInt::get(Idx->getType(),
462 FalseRangeEnd-FirstFalseElement);
463 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
467 // If a magic bitvector captures the entire comparison state
468 // of this load, replace it with computation that does:
469 // ((magic_cst >> i) & 1) != 0
473 // Look for an appropriate type:
474 // - The type of Idx if the magic fits
475 // - The smallest fitting legal type if we have a DataLayout
477 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
480 Ty = DL->getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
481 else if (ArrayElementCount <= 32)
482 Ty = Type::getInt32Ty(Init->getContext());
485 Value *V = Builder->CreateIntCast(Idx, Ty, false);
486 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
487 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
488 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
496 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
497 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
498 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
499 /// be complex, and scales are involved. The above expression would also be
500 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
501 /// This later form is less amenable to optimization though, and we are allowed
502 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
504 /// If we can't emit an optimized form for this expression, this returns null.
506 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
507 const DataLayout &DL = *IC.getDataLayout();
508 gep_type_iterator GTI = gep_type_begin(GEP);
510 // Check to see if this gep only has a single variable index. If so, and if
511 // any constant indices are a multiple of its scale, then we can compute this
512 // in terms of the scale of the variable index. For example, if the GEP
513 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
514 // because the expression will cross zero at the same point.
515 unsigned i, e = GEP->getNumOperands();
517 for (i = 1; i != e; ++i, ++GTI) {
518 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
519 // Compute the aggregate offset of constant indices.
520 if (CI->isZero()) continue;
522 // Handle a struct index, which adds its field offset to the pointer.
523 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
524 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
526 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
527 Offset += Size*CI->getSExtValue();
530 // Found our variable index.
535 // If there are no variable indices, we must have a constant offset, just
536 // evaluate it the general way.
537 if (i == e) return 0;
539 Value *VariableIdx = GEP->getOperand(i);
540 // Determine the scale factor of the variable element. For example, this is
541 // 4 if the variable index is into an array of i32.
542 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
544 // Verify that there are no other variable indices. If so, emit the hard way.
545 for (++i, ++GTI; i != e; ++i, ++GTI) {
546 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
549 // Compute the aggregate offset of constant indices.
550 if (CI->isZero()) continue;
552 // Handle a struct index, which adds its field offset to the pointer.
553 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
554 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
556 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
557 Offset += Size*CI->getSExtValue();
563 // Okay, we know we have a single variable index, which must be a
564 // pointer/array/vector index. If there is no offset, life is simple, return
566 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
567 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
569 // Cast to intptrty in case a truncation occurs. If an extension is needed,
570 // we don't need to bother extending: the extension won't affect where the
571 // computation crosses zero.
572 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
573 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
578 // Otherwise, there is an index. The computation we will do will be modulo
579 // the pointer size, so get it.
580 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
582 Offset &= PtrSizeMask;
583 VariableScale &= PtrSizeMask;
585 // To do this transformation, any constant index must be a multiple of the
586 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
587 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
588 // multiple of the variable scale.
589 int64_t NewOffs = Offset / (int64_t)VariableScale;
590 if (Offset != NewOffs*(int64_t)VariableScale)
593 // Okay, we can do this evaluation. Start by converting the index to intptr.
594 if (VariableIdx->getType() != IntPtrTy)
595 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
597 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
598 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
601 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
602 /// else. At this point we know that the GEP is on the LHS of the comparison.
603 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
604 ICmpInst::Predicate Cond,
606 // Don't transform signed compares of GEPs into index compares. Even if the
607 // GEP is inbounds, the final add of the base pointer can have signed overflow
608 // and would change the result of the icmp.
609 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
610 // the maximum signed value for the pointer type.
611 if (ICmpInst::isSigned(Cond))
614 // Look through bitcasts.
615 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
616 RHS = BCI->getOperand(0);
618 Value *PtrBase = GEPLHS->getOperand(0);
619 if (DL && PtrBase == RHS && GEPLHS->isInBounds()) {
620 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
621 // This transformation (ignoring the base and scales) is valid because we
622 // know pointers can't overflow since the gep is inbounds. See if we can
623 // output an optimized form.
624 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
626 // If not, synthesize the offset the hard way.
628 Offset = EmitGEPOffset(GEPLHS);
629 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
630 Constant::getNullValue(Offset->getType()));
631 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
632 // If the base pointers are different, but the indices are the same, just
633 // compare the base pointer.
634 if (PtrBase != GEPRHS->getOperand(0)) {
635 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
636 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
637 GEPRHS->getOperand(0)->getType();
639 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
640 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
641 IndicesTheSame = false;
645 // If all indices are the same, just compare the base pointers.
647 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
649 // If we're comparing GEPs with two base pointers that only differ in type
650 // and both GEPs have only constant indices or just one use, then fold
651 // the compare with the adjusted indices.
652 if (DL && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
653 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
654 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
655 PtrBase->stripPointerCasts() ==
656 GEPRHS->getOperand(0)->stripPointerCasts()) {
657 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
658 EmitGEPOffset(GEPLHS),
659 EmitGEPOffset(GEPRHS));
660 return ReplaceInstUsesWith(I, Cmp);
663 // Otherwise, the base pointers are different and the indices are
664 // different, bail out.
668 // If one of the GEPs has all zero indices, recurse.
669 bool AllZeros = true;
670 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
671 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
672 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
677 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
678 ICmpInst::getSwappedPredicate(Cond), I);
680 // If the other GEP has all zero indices, recurse.
682 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
683 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
684 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
689 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
691 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
692 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
693 // If the GEPs only differ by one index, compare it.
694 unsigned NumDifferences = 0; // Keep track of # differences.
695 unsigned DiffOperand = 0; // The operand that differs.
696 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
697 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
698 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
699 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
700 // Irreconcilable differences.
704 if (NumDifferences++) break;
709 if (NumDifferences == 0) // SAME GEP?
710 return ReplaceInstUsesWith(I, // No comparison is needed here.
711 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond)));
713 else if (NumDifferences == 1 && GEPsInBounds) {
714 Value *LHSV = GEPLHS->getOperand(DiffOperand);
715 Value *RHSV = GEPRHS->getOperand(DiffOperand);
716 // Make sure we do a signed comparison here.
717 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
721 // Only lower this if the icmp is the only user of the GEP or if we expect
722 // the result to fold to a constant!
725 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
726 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
727 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
728 Value *L = EmitGEPOffset(GEPLHS);
729 Value *R = EmitGEPOffset(GEPRHS);
730 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
736 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
737 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI,
738 Value *X, ConstantInt *CI,
739 ICmpInst::Predicate Pred) {
740 // If we have X+0, exit early (simplifying logic below) and let it get folded
741 // elsewhere. icmp X+0, X -> icmp X, X
743 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
744 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
747 // (X+4) == X -> false.
748 if (Pred == ICmpInst::ICMP_EQ)
749 return ReplaceInstUsesWith(ICI, Builder->getFalse());
751 // (X+4) != X -> true.
752 if (Pred == ICmpInst::ICMP_NE)
753 return ReplaceInstUsesWith(ICI, Builder->getTrue());
755 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
756 // so the values can never be equal. Similarly for all other "or equals"
759 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
760 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
761 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
762 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
764 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
765 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
768 // (X+1) >u X --> X <u (0-1) --> X != 255
769 // (X+2) >u X --> X <u (0-2) --> X <u 254
770 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
771 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
772 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
774 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
775 ConstantInt *SMax = ConstantInt::get(X->getContext(),
776 APInt::getSignedMaxValue(BitWidth));
778 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
779 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
780 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
781 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
782 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
783 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
784 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
785 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
787 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
788 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
789 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
790 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
791 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
792 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
794 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
795 Constant *C = Builder->getInt(CI->getValue()-1);
796 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
799 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
800 /// and CmpRHS are both known to be integer constants.
801 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
802 ConstantInt *DivRHS) {
803 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
804 const APInt &CmpRHSV = CmpRHS->getValue();
806 // FIXME: If the operand types don't match the type of the divide
807 // then don't attempt this transform. The code below doesn't have the
808 // logic to deal with a signed divide and an unsigned compare (and
809 // vice versa). This is because (x /s C1) <s C2 produces different
810 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
811 // (x /u C1) <u C2. Simply casting the operands and result won't
812 // work. :( The if statement below tests that condition and bails
814 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
815 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
817 if (DivRHS->isZero())
818 return 0; // The ProdOV computation fails on divide by zero.
819 if (DivIsSigned && DivRHS->isAllOnesValue())
820 return 0; // The overflow computation also screws up here
821 if (DivRHS->isOne()) {
822 // This eliminates some funny cases with INT_MIN.
823 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
827 // Compute Prod = CI * DivRHS. We are essentially solving an equation
828 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
829 // C2 (CI). By solving for X we can turn this into a range check
830 // instead of computing a divide.
831 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
833 // Determine if the product overflows by seeing if the product is
834 // not equal to the divide. Make sure we do the same kind of divide
835 // as in the LHS instruction that we're folding.
836 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
837 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
839 // Get the ICmp opcode
840 ICmpInst::Predicate Pred = ICI.getPredicate();
842 /// If the division is known to be exact, then there is no remainder from the
843 /// divide, so the covered range size is unit, otherwise it is the divisor.
844 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
846 // Figure out the interval that is being checked. For example, a comparison
847 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
848 // Compute this interval based on the constants involved and the signedness of
849 // the compare/divide. This computes a half-open interval, keeping track of
850 // whether either value in the interval overflows. After analysis each
851 // overflow variable is set to 0 if it's corresponding bound variable is valid
852 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
853 int LoOverflow = 0, HiOverflow = 0;
854 Constant *LoBound = 0, *HiBound = 0;
856 if (!DivIsSigned) { // udiv
857 // e.g. X/5 op 3 --> [15, 20)
859 HiOverflow = LoOverflow = ProdOV;
861 // If this is not an exact divide, then many values in the range collapse
862 // to the same result value.
863 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
866 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
867 if (CmpRHSV == 0) { // (X / pos) op 0
868 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
869 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
871 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
872 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
873 HiOverflow = LoOverflow = ProdOV;
875 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
876 } else { // (X / pos) op neg
877 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
878 HiBound = AddOne(Prod);
879 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
881 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
882 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
885 } else if (DivRHS->isNegative()) { // Divisor is < 0.
887 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
888 if (CmpRHSV == 0) { // (X / neg) op 0
889 // e.g. X/-5 op 0 --> [-4, 5)
890 LoBound = AddOne(RangeSize);
891 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
892 if (HiBound == DivRHS) { // -INTMIN = INTMIN
893 HiOverflow = 1; // [INTMIN+1, overflow)
894 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
896 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
897 // e.g. X/-5 op 3 --> [-19, -14)
898 HiBound = AddOne(Prod);
899 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
901 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
902 } else { // (X / neg) op neg
903 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
904 LoOverflow = HiOverflow = ProdOV;
906 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
909 // Dividing by a negative swaps the condition. LT <-> GT
910 Pred = ICmpInst::getSwappedPredicate(Pred);
913 Value *X = DivI->getOperand(0);
915 default: llvm_unreachable("Unhandled icmp opcode!");
916 case ICmpInst::ICMP_EQ:
917 if (LoOverflow && HiOverflow)
918 return ReplaceInstUsesWith(ICI, Builder->getFalse());
920 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
921 ICmpInst::ICMP_UGE, X, LoBound);
923 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
924 ICmpInst::ICMP_ULT, X, HiBound);
925 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
927 case ICmpInst::ICMP_NE:
928 if (LoOverflow && HiOverflow)
929 return ReplaceInstUsesWith(ICI, Builder->getTrue());
931 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
932 ICmpInst::ICMP_ULT, X, LoBound);
934 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
935 ICmpInst::ICMP_UGE, X, HiBound);
936 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
937 DivIsSigned, false));
938 case ICmpInst::ICMP_ULT:
939 case ICmpInst::ICMP_SLT:
940 if (LoOverflow == +1) // Low bound is greater than input range.
941 return ReplaceInstUsesWith(ICI, Builder->getTrue());
942 if (LoOverflow == -1) // Low bound is less than input range.
943 return ReplaceInstUsesWith(ICI, Builder->getFalse());
944 return new ICmpInst(Pred, X, LoBound);
945 case ICmpInst::ICMP_UGT:
946 case ICmpInst::ICMP_SGT:
947 if (HiOverflow == +1) // High bound greater than input range.
948 return ReplaceInstUsesWith(ICI, Builder->getFalse());
949 if (HiOverflow == -1) // High bound less than input range.
950 return ReplaceInstUsesWith(ICI, Builder->getTrue());
951 if (Pred == ICmpInst::ICMP_UGT)
952 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
953 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
957 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
958 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
959 ConstantInt *ShAmt) {
960 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
962 // Check that the shift amount is in range. If not, don't perform
963 // undefined shifts. When the shift is visited it will be
965 uint32_t TypeBits = CmpRHSV.getBitWidth();
966 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
967 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
970 if (!ICI.isEquality()) {
971 // If we have an unsigned comparison and an ashr, we can't simplify this.
972 // Similarly for signed comparisons with lshr.
973 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
976 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
977 // by a power of 2. Since we already have logic to simplify these,
978 // transform to div and then simplify the resultant comparison.
979 if (Shr->getOpcode() == Instruction::AShr &&
980 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
983 // Revisit the shift (to delete it).
987 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
990 Shr->getOpcode() == Instruction::AShr ?
991 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
992 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
994 ICI.setOperand(0, Tmp);
996 // If the builder folded the binop, just return it.
997 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
1001 // Otherwise, fold this div/compare.
1002 assert(TheDiv->getOpcode() == Instruction::SDiv ||
1003 TheDiv->getOpcode() == Instruction::UDiv);
1005 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
1006 assert(Res && "This div/cst should have folded!");
1011 // If we are comparing against bits always shifted out, the
1012 // comparison cannot succeed.
1013 APInt Comp = CmpRHSV << ShAmtVal;
1014 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
1015 if (Shr->getOpcode() == Instruction::LShr)
1016 Comp = Comp.lshr(ShAmtVal);
1018 Comp = Comp.ashr(ShAmtVal);
1020 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
1021 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1022 Constant *Cst = Builder->getInt1(IsICMP_NE);
1023 return ReplaceInstUsesWith(ICI, Cst);
1026 // Otherwise, check to see if the bits shifted out are known to be zero.
1027 // If so, we can compare against the unshifted value:
1028 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
1029 if (Shr->hasOneUse() && Shr->isExact())
1030 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
1032 if (Shr->hasOneUse()) {
1033 // Otherwise strength reduce the shift into an and.
1034 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
1035 Constant *Mask = Builder->getInt(Val);
1037 Value *And = Builder->CreateAnd(Shr->getOperand(0),
1038 Mask, Shr->getName()+".mask");
1039 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
1045 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
1047 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
1050 const APInt &RHSV = RHS->getValue();
1052 switch (LHSI->getOpcode()) {
1053 case Instruction::Trunc:
1054 if (ICI.isEquality() && LHSI->hasOneUse()) {
1055 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1056 // of the high bits truncated out of x are known.
1057 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1058 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1059 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1060 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
1062 // If all the high bits are known, we can do this xform.
1063 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1064 // Pull in the high bits from known-ones set.
1065 APInt NewRHS = RHS->getValue().zext(SrcBits);
1066 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
1067 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1068 Builder->getInt(NewRHS));
1073 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI)
1074 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1075 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1077 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1078 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1079 Value *CompareVal = LHSI->getOperand(0);
1081 // If the sign bit of the XorCst is not set, there is no change to
1082 // the operation, just stop using the Xor.
1083 if (!XorCst->isNegative()) {
1084 ICI.setOperand(0, CompareVal);
1089 // Was the old condition true if the operand is positive?
1090 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1092 // If so, the new one isn't.
1093 isTrueIfPositive ^= true;
1095 if (isTrueIfPositive)
1096 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1099 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1103 if (LHSI->hasOneUse()) {
1104 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1105 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) {
1106 const APInt &SignBit = XorCst->getValue();
1107 ICmpInst::Predicate Pred = ICI.isSigned()
1108 ? ICI.getUnsignedPredicate()
1109 : ICI.getSignedPredicate();
1110 return new ICmpInst(Pred, LHSI->getOperand(0),
1111 Builder->getInt(RHSV ^ SignBit));
1114 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1115 if (!ICI.isEquality() && XorCst->isMaxValue(true)) {
1116 const APInt &NotSignBit = XorCst->getValue();
1117 ICmpInst::Predicate Pred = ICI.isSigned()
1118 ? ICI.getUnsignedPredicate()
1119 : ICI.getSignedPredicate();
1120 Pred = ICI.getSwappedPredicate(Pred);
1121 return new ICmpInst(Pred, LHSI->getOperand(0),
1122 Builder->getInt(RHSV ^ NotSignBit));
1126 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
1127 // iff -C is a power of 2
1128 if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
1129 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2())
1130 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst);
1132 // (icmp ult (xor X, C), -C) -> (icmp uge X, C)
1133 // iff -C is a power of 2
1134 if (ICI.getPredicate() == ICmpInst::ICMP_ULT &&
1135 XorCst->getValue() == -RHSV && RHSV.isPowerOf2())
1136 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst);
1139 case Instruction::And: // (icmp pred (and X, AndCst), RHS)
1140 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1141 LHSI->getOperand(0)->hasOneUse()) {
1142 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1));
1144 // If the LHS is an AND of a truncating cast, we can widen the
1145 // and/compare to be the input width without changing the value
1146 // produced, eliminating a cast.
1147 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1148 // We can do this transformation if either the AND constant does not
1149 // have its sign bit set or if it is an equality comparison.
1150 // Extending a relational comparison when we're checking the sign
1151 // bit would not work.
1152 if (ICI.isEquality() ||
1153 (!AndCst->isNegative() && RHSV.isNonNegative())) {
1155 Builder->CreateAnd(Cast->getOperand(0),
1156 ConstantExpr::getZExt(AndCst, Cast->getSrcTy()));
1157 NewAnd->takeName(LHSI);
1158 return new ICmpInst(ICI.getPredicate(), NewAnd,
1159 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1163 // If the LHS is an AND of a zext, and we have an equality compare, we can
1164 // shrink the and/compare to the smaller type, eliminating the cast.
1165 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1166 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1167 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1168 // should fold the icmp to true/false in that case.
1169 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1171 Builder->CreateAnd(Cast->getOperand(0),
1172 ConstantExpr::getTrunc(AndCst, Ty));
1173 NewAnd->takeName(LHSI);
1174 return new ICmpInst(ICI.getPredicate(), NewAnd,
1175 ConstantExpr::getTrunc(RHS, Ty));
1179 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1180 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1181 // happens a LOT in code produced by the C front-end, for bitfield
1183 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1184 if (Shift && !Shift->isShift())
1188 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1190 // This seemingly simple opportunity to fold away a shift turns out to
1191 // be rather complicated. See PR17827
1192 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details.
1194 bool CanFold = false;
1195 unsigned ShiftOpcode = Shift->getOpcode();
1196 if (ShiftOpcode == Instruction::AShr) {
1197 // There may be some constraints that make this possible,
1198 // but nothing simple has been discovered yet.
1200 } else if (ShiftOpcode == Instruction::Shl) {
1201 // For a left shift, we can fold if the comparison is not signed.
1202 // We can also fold a signed comparison if the mask value and
1203 // comparison value are not negative. These constraints may not be
1204 // obvious, but we can prove that they are correct using an SMT
1206 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative()))
1208 } else if (ShiftOpcode == Instruction::LShr) {
1209 // For a logical right shift, we can fold if the comparison is not
1210 // signed. We can also fold a signed comparison if the shifted mask
1211 // value and the shifted comparison value are not negative.
1212 // These constraints may not be obvious, but we can prove that they
1213 // are correct using an SMT solver.
1214 if (!ICI.isSigned())
1217 ConstantInt *ShiftedAndCst =
1218 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt));
1219 ConstantInt *ShiftedRHSCst =
1220 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt));
1222 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative())
1229 if (ShiftOpcode == Instruction::Shl)
1230 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1232 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1234 // Check to see if we are shifting out any of the bits being
1236 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) {
1237 // If we shifted bits out, the fold is not going to work out.
1238 // As a special case, check to see if this means that the
1239 // result is always true or false now.
1240 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1241 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1242 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1243 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1245 ICI.setOperand(1, NewCst);
1246 Constant *NewAndCst;
1247 if (ShiftOpcode == Instruction::Shl)
1248 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt);
1250 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt);
1251 LHSI->setOperand(1, NewAndCst);
1252 LHSI->setOperand(0, Shift->getOperand(0));
1253 Worklist.Add(Shift); // Shift is dead.
1259 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1260 // preferable because it allows the C<<Y expression to be hoisted out
1261 // of a loop if Y is invariant and X is not.
1262 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1263 ICI.isEquality() && !Shift->isArithmeticShift() &&
1264 !isa<Constant>(Shift->getOperand(0))) {
1267 if (Shift->getOpcode() == Instruction::LShr) {
1268 NS = Builder->CreateShl(AndCst, Shift->getOperand(1));
1270 // Insert a logical shift.
1271 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1));
1274 // Compute X & (C << Y).
1276 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1278 ICI.setOperand(0, NewAnd);
1282 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any
1283 // bit set in (X & AndCst) will produce a result greater than RHSV.
1284 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
1285 unsigned NTZ = AndCst->getValue().countTrailingZeros();
1286 if ((NTZ < AndCst->getBitWidth()) &&
1287 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV))
1288 return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
1289 Constant::getNullValue(RHS->getType()));
1293 // Try to optimize things like "A[i]&42 == 0" to index computations.
1294 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1295 if (GetElementPtrInst *GEP =
1296 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1297 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1298 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1299 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1300 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1301 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1306 // X & -C == -C -> X > u ~C
1307 // X & -C != -C -> X <= u ~C
1308 // iff C is a power of 2
1309 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2())
1310 return new ICmpInst(
1311 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT
1312 : ICmpInst::ICMP_ULE,
1313 LHSI->getOperand(0), SubOne(RHS));
1316 case Instruction::Or: {
1317 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1320 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1321 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1322 // -> and (icmp eq P, null), (icmp eq Q, null).
1323 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1324 Constant::getNullValue(P->getType()));
1325 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1326 Constant::getNullValue(Q->getType()));
1328 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1329 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1331 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1337 case Instruction::Mul: { // (icmp pred (mul X, Val), CI)
1338 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1341 // If this is a signed comparison to 0 and the mul is sign preserving,
1342 // use the mul LHS operand instead.
1343 ICmpInst::Predicate pred = ICI.getPredicate();
1344 if (isSignTest(pred, RHS) && !Val->isZero() &&
1345 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1346 return new ICmpInst(Val->isNegative() ?
1347 ICmpInst::getSwappedPredicate(pred) : pred,
1348 LHSI->getOperand(0),
1349 Constant::getNullValue(RHS->getType()));
1354 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1355 uint32_t TypeBits = RHSV.getBitWidth();
1356 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1359 // (1 << X) pred P2 -> X pred Log2(P2)
1360 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) {
1361 bool RHSVIsPowerOf2 = RHSV.isPowerOf2();
1362 ICmpInst::Predicate Pred = ICI.getPredicate();
1363 if (ICI.isUnsigned()) {
1364 if (!RHSVIsPowerOf2) {
1365 // (1 << X) < 30 -> X <= 4
1366 // (1 << X) <= 30 -> X <= 4
1367 // (1 << X) >= 30 -> X > 4
1368 // (1 << X) > 30 -> X > 4
1369 if (Pred == ICmpInst::ICMP_ULT)
1370 Pred = ICmpInst::ICMP_ULE;
1371 else if (Pred == ICmpInst::ICMP_UGE)
1372 Pred = ICmpInst::ICMP_UGT;
1374 unsigned RHSLog2 = RHSV.logBase2();
1376 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31
1377 // (1 << X) > 2147483648 -> X > 31 -> false
1378 // (1 << X) <= 2147483648 -> X <= 31 -> true
1379 // (1 << X) < 2147483648 -> X < 31 -> X != 31
1380 if (RHSLog2 == TypeBits-1) {
1381 if (Pred == ICmpInst::ICMP_UGE)
1382 Pred = ICmpInst::ICMP_EQ;
1383 else if (Pred == ICmpInst::ICMP_UGT)
1384 return ReplaceInstUsesWith(ICI, Builder->getFalse());
1385 else if (Pred == ICmpInst::ICMP_ULE)
1386 return ReplaceInstUsesWith(ICI, Builder->getTrue());
1387 else if (Pred == ICmpInst::ICMP_ULT)
1388 Pred = ICmpInst::ICMP_NE;
1391 return new ICmpInst(Pred, X,
1392 ConstantInt::get(RHS->getType(), RHSLog2));
1393 } else if (ICI.isSigned()) {
1394 if (RHSV.isAllOnesValue()) {
1395 // (1 << X) <= -1 -> X == 31
1396 if (Pred == ICmpInst::ICMP_SLE)
1397 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1398 ConstantInt::get(RHS->getType(), TypeBits-1));
1400 // (1 << X) > -1 -> X != 31
1401 if (Pred == ICmpInst::ICMP_SGT)
1402 return new ICmpInst(ICmpInst::ICMP_NE, X,
1403 ConstantInt::get(RHS->getType(), TypeBits-1));
1405 // (1 << X) < 0 -> X == 31
1406 // (1 << X) <= 0 -> X == 31
1407 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1408 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1409 ConstantInt::get(RHS->getType(), TypeBits-1));
1411 // (1 << X) >= 0 -> X != 31
1412 // (1 << X) > 0 -> X != 31
1413 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1414 return new ICmpInst(ICmpInst::ICMP_NE, X,
1415 ConstantInt::get(RHS->getType(), TypeBits-1));
1417 } else if (ICI.isEquality()) {
1419 return new ICmpInst(
1420 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2()));
1422 return ReplaceInstUsesWith(
1423 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse()
1424 : Builder->getTrue());
1430 // Check that the shift amount is in range. If not, don't perform
1431 // undefined shifts. When the shift is visited it will be
1433 if (ShAmt->uge(TypeBits))
1436 if (ICI.isEquality()) {
1437 // If we are comparing against bits always shifted out, the
1438 // comparison cannot succeed.
1440 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1442 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1443 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1444 Constant *Cst = Builder->getInt1(IsICMP_NE);
1445 return ReplaceInstUsesWith(ICI, Cst);
1448 // If the shift is NUW, then it is just shifting out zeros, no need for an
1450 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1451 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1452 ConstantExpr::getLShr(RHS, ShAmt));
1454 // If the shift is NSW and we compare to 0, then it is just shifting out
1455 // sign bits, no need for an AND either.
1456 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0)
1457 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1458 ConstantExpr::getLShr(RHS, ShAmt));
1460 if (LHSI->hasOneUse()) {
1461 // Otherwise strength reduce the shift into an and.
1462 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1463 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits,
1464 TypeBits - ShAmtVal));
1467 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1468 return new ICmpInst(ICI.getPredicate(), And,
1469 ConstantExpr::getLShr(RHS, ShAmt));
1473 // If this is a signed comparison to 0 and the shift is sign preserving,
1474 // use the shift LHS operand instead.
1475 ICmpInst::Predicate pred = ICI.getPredicate();
1476 if (isSignTest(pred, RHS) &&
1477 cast<BinaryOperator>(LHSI)->hasNoSignedWrap())
1478 return new ICmpInst(pred,
1479 LHSI->getOperand(0),
1480 Constant::getNullValue(RHS->getType()));
1482 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1483 bool TrueIfSigned = false;
1484 if (LHSI->hasOneUse() &&
1485 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1486 // (X << 31) <s 0 --> (X&1) != 0
1487 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1488 APInt::getOneBitSet(TypeBits,
1489 TypeBits-ShAmt->getZExtValue()-1));
1491 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1492 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1493 And, Constant::getNullValue(And->getType()));
1496 // Transform (icmp pred iM (shl iM %v, N), CI)
1497 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N))
1498 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N.
1499 // This enables to get rid of the shift in favor of a trunc which can be
1500 // free on the target. It has the additional benefit of comparing to a
1501 // smaller constant, which will be target friendly.
1502 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
1503 if (LHSI->hasOneUse() &&
1504 Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
1505 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
1506 Constant *NCI = ConstantExpr::getTrunc(
1507 ConstantExpr::getAShr(RHS,
1508 ConstantInt::get(RHS->getType(), Amt)),
1510 return new ICmpInst(ICI.getPredicate(),
1511 Builder->CreateTrunc(LHSI->getOperand(0), NTy),
1518 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1519 case Instruction::AShr: {
1520 // Handle equality comparisons of shift-by-constant.
1521 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1522 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1523 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1527 // Handle exact shr's.
1528 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1529 if (RHSV.isMinValue())
1530 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1535 case Instruction::SDiv:
1536 case Instruction::UDiv:
1537 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1538 // Fold this div into the comparison, producing a range check.
1539 // Determine, based on the divide type, what the range is being
1540 // checked. If there is an overflow on the low or high side, remember
1541 // it, otherwise compute the range [low, hi) bounding the new value.
1542 // See: InsertRangeTest above for the kinds of replacements possible.
1543 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1544 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1549 case Instruction::Sub: {
1550 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0));
1552 const APInt &LHSV = LHSC->getValue();
1554 // C1-X <u C2 -> (X|(C2-1)) == C1
1555 // iff C1 & (C2-1) == C2-1
1556 // C2 is a power of 2
1557 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1558 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1))
1559 return new ICmpInst(ICmpInst::ICMP_EQ,
1560 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1),
1563 // C1-X >u C2 -> (X|C2) != C1
1564 // iff C1 & C2 == C2
1565 // C2+1 is a power of 2
1566 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1567 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV)
1568 return new ICmpInst(ICmpInst::ICMP_NE,
1569 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC);
1573 case Instruction::Add:
1574 // Fold: icmp pred (add X, C1), C2
1575 if (!ICI.isEquality()) {
1576 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1578 const APInt &LHSV = LHSC->getValue();
1580 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1583 if (ICI.isSigned()) {
1584 if (CR.getLower().isSignBit()) {
1585 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1586 Builder->getInt(CR.getUpper()));
1587 } else if (CR.getUpper().isSignBit()) {
1588 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1589 Builder->getInt(CR.getLower()));
1592 if (CR.getLower().isMinValue()) {
1593 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1594 Builder->getInt(CR.getUpper()));
1595 } else if (CR.getUpper().isMinValue()) {
1596 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1597 Builder->getInt(CR.getLower()));
1601 // X-C1 <u C2 -> (X & -C2) == C1
1602 // iff C1 & (C2-1) == 0
1603 // C2 is a power of 2
1604 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() &&
1605 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0)
1606 return new ICmpInst(ICmpInst::ICMP_EQ,
1607 Builder->CreateAnd(LHSI->getOperand(0), -RHSV),
1608 ConstantExpr::getNeg(LHSC));
1610 // X-C1 >u C2 -> (X & ~C2) != C1
1612 // C2+1 is a power of 2
1613 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() &&
1614 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0)
1615 return new ICmpInst(ICmpInst::ICMP_NE,
1616 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV),
1617 ConstantExpr::getNeg(LHSC));
1622 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1623 if (ICI.isEquality()) {
1624 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1626 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1627 // the second operand is a constant, simplify a bit.
1628 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1629 switch (BO->getOpcode()) {
1630 case Instruction::SRem:
1631 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1632 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1633 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1634 if (V.sgt(1) && V.isPowerOf2()) {
1636 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1638 return new ICmpInst(ICI.getPredicate(), NewRem,
1639 Constant::getNullValue(BO->getType()));
1643 case Instruction::Add:
1644 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1645 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1646 if (BO->hasOneUse())
1647 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1648 ConstantExpr::getSub(RHS, BOp1C));
1649 } else if (RHSV == 0) {
1650 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1651 // efficiently invertible, or if the add has just this one use.
1652 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1654 if (Value *NegVal = dyn_castNegVal(BOp1))
1655 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1656 if (Value *NegVal = dyn_castNegVal(BOp0))
1657 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1658 if (BO->hasOneUse()) {
1659 Value *Neg = Builder->CreateNeg(BOp1);
1661 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1665 case Instruction::Xor:
1666 // For the xor case, we can xor two constants together, eliminating
1667 // the explicit xor.
1668 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1669 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1670 ConstantExpr::getXor(RHS, BOC));
1671 } else if (RHSV == 0) {
1672 // Replace ((xor A, B) != 0) with (A != B)
1673 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1677 case Instruction::Sub:
1678 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1679 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1680 if (BO->hasOneUse())
1681 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1682 ConstantExpr::getSub(BOp0C, RHS));
1683 } else if (RHSV == 0) {
1684 // Replace ((sub A, B) != 0) with (A != B)
1685 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1689 case Instruction::Or:
1690 // If bits are being or'd in that are not present in the constant we
1691 // are comparing against, then the comparison could never succeed!
1692 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1693 Constant *NotCI = ConstantExpr::getNot(RHS);
1694 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1695 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1699 case Instruction::And:
1700 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1701 // If bits are being compared against that are and'd out, then the
1702 // comparison can never succeed!
1703 if ((RHSV & ~BOC->getValue()) != 0)
1704 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE));
1706 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1707 if (RHS == BOC && RHSV.isPowerOf2())
1708 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1709 ICmpInst::ICMP_NE, LHSI,
1710 Constant::getNullValue(RHS->getType()));
1712 // Don't perform the following transforms if the AND has multiple uses
1713 if (!BO->hasOneUse())
1716 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1717 if (BOC->getValue().isSignBit()) {
1718 Value *X = BO->getOperand(0);
1719 Constant *Zero = Constant::getNullValue(X->getType());
1720 ICmpInst::Predicate pred = isICMP_NE ?
1721 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1722 return new ICmpInst(pred, X, Zero);
1725 // ((X & ~7) == 0) --> X < 8
1726 if (RHSV == 0 && isHighOnes(BOC)) {
1727 Value *X = BO->getOperand(0);
1728 Constant *NegX = ConstantExpr::getNeg(BOC);
1729 ICmpInst::Predicate pred = isICMP_NE ?
1730 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1731 return new ICmpInst(pred, X, NegX);
1735 case Instruction::Mul:
1736 if (RHSV == 0 && BO->hasNoSignedWrap()) {
1737 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1738 // The trivial case (mul X, 0) is handled by InstSimplify
1739 // General case : (mul X, C) != 0 iff X != 0
1740 // (mul X, C) == 0 iff X == 0
1742 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1743 Constant::getNullValue(RHS->getType()));
1749 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1750 // Handle icmp {eq|ne} <intrinsic>, intcst.
1751 switch (II->getIntrinsicID()) {
1752 case Intrinsic::bswap:
1754 ICI.setOperand(0, II->getArgOperand(0));
1755 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap()));
1757 case Intrinsic::ctlz:
1758 case Intrinsic::cttz:
1759 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1760 if (RHSV == RHS->getType()->getBitWidth()) {
1762 ICI.setOperand(0, II->getArgOperand(0));
1763 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1767 case Intrinsic::ctpop:
1768 // popcount(A) == 0 -> A == 0 and likewise for !=
1769 if (RHS->isZero()) {
1771 ICI.setOperand(0, II->getArgOperand(0));
1772 ICI.setOperand(1, RHS);
1784 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1785 /// We only handle extending casts so far.
1787 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1788 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1789 Value *LHSCIOp = LHSCI->getOperand(0);
1790 Type *SrcTy = LHSCIOp->getType();
1791 Type *DestTy = LHSCI->getType();
1794 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1795 // integer type is the same size as the pointer type.
1796 if (DL && LHSCI->getOpcode() == Instruction::PtrToInt &&
1797 DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) {
1799 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1800 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1801 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1802 RHSOp = RHSC->getOperand(0);
1803 // If the pointer types don't match, insert a bitcast.
1804 if (LHSCIOp->getType() != RHSOp->getType())
1805 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1809 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1812 // The code below only handles extension cast instructions, so far.
1814 if (LHSCI->getOpcode() != Instruction::ZExt &&
1815 LHSCI->getOpcode() != Instruction::SExt)
1818 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1819 bool isSignedCmp = ICI.isSigned();
1821 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1822 // Not an extension from the same type?
1823 RHSCIOp = CI->getOperand(0);
1824 if (RHSCIOp->getType() != LHSCIOp->getType())
1827 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1828 // and the other is a zext), then we can't handle this.
1829 if (CI->getOpcode() != LHSCI->getOpcode())
1832 // Deal with equality cases early.
1833 if (ICI.isEquality())
1834 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1836 // A signed comparison of sign extended values simplifies into a
1837 // signed comparison.
1838 if (isSignedCmp && isSignedExt)
1839 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1841 // The other three cases all fold into an unsigned comparison.
1842 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1845 // If we aren't dealing with a constant on the RHS, exit early
1846 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1850 // Compute the constant that would happen if we truncated to SrcTy then
1851 // reextended to DestTy.
1852 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1853 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1856 // If the re-extended constant didn't change...
1858 // Deal with equality cases early.
1859 if (ICI.isEquality())
1860 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1862 // A signed comparison of sign extended values simplifies into a
1863 // signed comparison.
1864 if (isSignedExt && isSignedCmp)
1865 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1867 // The other three cases all fold into an unsigned comparison.
1868 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1871 // The re-extended constant changed so the constant cannot be represented
1872 // in the shorter type. Consequently, we cannot emit a simple comparison.
1873 // All the cases that fold to true or false will have already been handled
1874 // by SimplifyICmpInst, so only deal with the tricky case.
1876 if (isSignedCmp || !isSignedExt)
1879 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1880 // should have been folded away previously and not enter in here.
1882 // We're performing an unsigned comp with a sign extended value.
1883 // This is true if the input is >= 0. [aka >s -1]
1884 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1885 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1887 // Finally, return the value computed.
1888 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1889 return ReplaceInstUsesWith(ICI, Result);
1891 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1892 return BinaryOperator::CreateNot(Result);
1895 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1896 /// I = icmp ugt (add (add A, B), CI2), CI1
1897 /// If this is of the form:
1899 /// if (sum+128 >u 255)
1900 /// Then replace it with llvm.sadd.with.overflow.i8.
1902 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1903 ConstantInt *CI2, ConstantInt *CI1,
1905 // The transformation we're trying to do here is to transform this into an
1906 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1907 // with a narrower add, and discard the add-with-constant that is part of the
1908 // range check (if we can't eliminate it, this isn't profitable).
1910 // In order to eliminate the add-with-constant, the compare can be its only
1912 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1913 if (!AddWithCst->hasOneUse()) return 0;
1915 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1916 if (!CI2->getValue().isPowerOf2()) return 0;
1917 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1918 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1920 // The width of the new add formed is 1 more than the bias.
1923 // Check to see that CI1 is an all-ones value with NewWidth bits.
1924 if (CI1->getBitWidth() == NewWidth ||
1925 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1928 // This is only really a signed overflow check if the inputs have been
1929 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1930 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1931 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1932 if (IC.ComputeNumSignBits(A) < NeededSignBits ||
1933 IC.ComputeNumSignBits(B) < NeededSignBits)
1936 // In order to replace the original add with a narrower
1937 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1938 // and truncates that discard the high bits of the add. Verify that this is
1940 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1941 for (User *U : OrigAdd->users()) {
1942 if (U == AddWithCst) continue;
1944 // Only accept truncates for now. We would really like a nice recursive
1945 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1946 // chain to see which bits of a value are actually demanded. If the
1947 // original add had another add which was then immediately truncated, we
1948 // could still do the transformation.
1949 TruncInst *TI = dyn_cast<TruncInst>(U);
1951 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1954 // If the pattern matches, truncate the inputs to the narrower type and
1955 // use the sadd_with_overflow intrinsic to efficiently compute both the
1956 // result and the overflow bit.
1957 Module *M = I.getParent()->getParent()->getParent();
1959 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1960 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1963 InstCombiner::BuilderTy *Builder = IC.Builder;
1965 // Put the new code above the original add, in case there are any uses of the
1966 // add between the add and the compare.
1967 Builder->SetInsertPoint(OrigAdd);
1969 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1970 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1971 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1972 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1973 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1975 // The inner add was the result of the narrow add, zero extended to the
1976 // wider type. Replace it with the result computed by the intrinsic.
1977 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1979 // The original icmp gets replaced with the overflow value.
1980 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1983 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1985 // Don't bother doing this transformation for pointers, don't do it for
1987 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1989 // If the add is a constant expr, then we don't bother transforming it.
1990 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1991 if (OrigAdd == 0) return 0;
1993 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1995 // Put the new code above the original add, in case there are any uses of the
1996 // add between the add and the compare.
1997 InstCombiner::BuilderTy *Builder = IC.Builder;
1998 Builder->SetInsertPoint(OrigAdd);
2000 Module *M = I.getParent()->getParent()->getParent();
2001 Type *Ty = LHS->getType();
2002 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
2003 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
2004 Value *Add = Builder->CreateExtractValue(Call, 0);
2006 IC.ReplaceInstUsesWith(*OrigAdd, Add);
2008 // The original icmp gets replaced with the overflow value.
2009 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
2012 /// \brief Recognize and process idiom involving test for multiplication
2015 /// The caller has matched a pattern of the form:
2016 /// I = cmp u (mul(zext A, zext B), V
2017 /// The function checks if this is a test for overflow and if so replaces
2018 /// multiplication with call to 'mul.with.overflow' intrinsic.
2020 /// \param I Compare instruction.
2021 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
2022 /// the compare instruction. Must be of integer type.
2023 /// \param OtherVal The other argument of compare instruction.
2024 /// \returns Instruction which must replace the compare instruction, NULL if no
2025 /// replacement required.
2026 static Instruction *ProcessUMulZExtIdiom(ICmpInst &I, Value *MulVal,
2027 Value *OtherVal, InstCombiner &IC) {
2028 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
2029 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
2030 assert(isa<IntegerType>(MulVal->getType()));
2031 Instruction *MulInstr = cast<Instruction>(MulVal);
2032 assert(MulInstr->getOpcode() == Instruction::Mul);
2034 Instruction *LHS = cast<Instruction>(MulInstr->getOperand(0)),
2035 *RHS = cast<Instruction>(MulInstr->getOperand(1));
2036 assert(LHS->getOpcode() == Instruction::ZExt);
2037 assert(RHS->getOpcode() == Instruction::ZExt);
2038 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
2040 // Calculate type and width of the result produced by mul.with.overflow.
2041 Type *TyA = A->getType(), *TyB = B->getType();
2042 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
2043 WidthB = TyB->getPrimitiveSizeInBits();
2046 if (WidthB > WidthA) {
2054 // In order to replace the original mul with a narrower mul.with.overflow,
2055 // all uses must ignore upper bits of the product. The number of used low
2056 // bits must be not greater than the width of mul.with.overflow.
2057 if (MulVal->hasNUsesOrMore(2))
2058 for (User *U : MulVal->users()) {
2061 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2062 // Check if truncation ignores bits above MulWidth.
2063 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
2064 if (TruncWidth > MulWidth)
2066 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2067 // Check if AND ignores bits above MulWidth.
2068 if (BO->getOpcode() != Instruction::And)
2070 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
2071 const APInt &CVal = CI->getValue();
2072 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
2076 // Other uses prohibit this transformation.
2081 // Recognize patterns
2082 switch (I.getPredicate()) {
2083 case ICmpInst::ICMP_EQ:
2084 case ICmpInst::ICMP_NE:
2085 // Recognize pattern:
2086 // mulval = mul(zext A, zext B)
2087 // cmp eq/neq mulval, zext trunc mulval
2088 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
2089 if (Zext->hasOneUse()) {
2090 Value *ZextArg = Zext->getOperand(0);
2091 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
2092 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
2096 // Recognize pattern:
2097 // mulval = mul(zext A, zext B)
2098 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
2101 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
2102 if (ValToMask != MulVal)
2104 const APInt &CVal = CI->getValue() + 1;
2105 if (CVal.isPowerOf2()) {
2106 unsigned MaskWidth = CVal.logBase2();
2107 if (MaskWidth == MulWidth)
2108 break; // Recognized
2113 case ICmpInst::ICMP_UGT:
2114 // Recognize pattern:
2115 // mulval = mul(zext A, zext B)
2116 // cmp ugt mulval, max
2117 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2118 APInt MaxVal = APInt::getMaxValue(MulWidth);
2119 MaxVal = MaxVal.zext(CI->getBitWidth());
2120 if (MaxVal.eq(CI->getValue()))
2121 break; // Recognized
2125 case ICmpInst::ICMP_UGE:
2126 // Recognize pattern:
2127 // mulval = mul(zext A, zext B)
2128 // cmp uge mulval, max+1
2129 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2130 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2131 if (MaxVal.eq(CI->getValue()))
2132 break; // Recognized
2136 case ICmpInst::ICMP_ULE:
2137 // Recognize pattern:
2138 // mulval = mul(zext A, zext B)
2139 // cmp ule mulval, max
2140 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2141 APInt MaxVal = APInt::getMaxValue(MulWidth);
2142 MaxVal = MaxVal.zext(CI->getBitWidth());
2143 if (MaxVal.eq(CI->getValue()))
2144 break; // Recognized
2148 case ICmpInst::ICMP_ULT:
2149 // Recognize pattern:
2150 // mulval = mul(zext A, zext B)
2151 // cmp ule mulval, max + 1
2152 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
2153 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
2154 if (MaxVal.eq(CI->getValue()))
2155 break; // Recognized
2163 InstCombiner::BuilderTy *Builder = IC.Builder;
2164 Builder->SetInsertPoint(MulInstr);
2165 Module *M = I.getParent()->getParent()->getParent();
2167 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
2168 Value *MulA = A, *MulB = B;
2169 if (WidthA < MulWidth)
2170 MulA = Builder->CreateZExt(A, MulType);
2171 if (WidthB < MulWidth)
2172 MulB = Builder->CreateZExt(B, MulType);
2174 Intrinsic::getDeclaration(M, Intrinsic::umul_with_overflow, MulType);
2175 CallInst *Call = Builder->CreateCall2(F, MulA, MulB, "umul");
2176 IC.Worklist.Add(MulInstr);
2178 // If there are uses of mul result other than the comparison, we know that
2179 // they are truncation or binary AND. Change them to use result of
2180 // mul.with.overflow and adjust properly mask/size.
2181 if (MulVal->hasNUsesOrMore(2)) {
2182 Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value");
2183 for (User *U : MulVal->users()) {
2184 if (U == &I || U == OtherVal)
2186 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
2187 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
2188 IC.ReplaceInstUsesWith(*TI, Mul);
2190 TI->setOperand(0, Mul);
2191 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
2192 assert(BO->getOpcode() == Instruction::And);
2193 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
2194 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
2195 APInt ShortMask = CI->getValue().trunc(MulWidth);
2196 Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask);
2198 cast<Instruction>(Builder->CreateZExt(ShortAnd, BO->getType()));
2199 IC.Worklist.Add(Zext);
2200 IC.ReplaceInstUsesWith(*BO, Zext);
2202 llvm_unreachable("Unexpected Binary operation");
2204 IC.Worklist.Add(cast<Instruction>(U));
2207 if (isa<Instruction>(OtherVal))
2208 IC.Worklist.Add(cast<Instruction>(OtherVal));
2210 // The original icmp gets replaced with the overflow value, maybe inverted
2211 // depending on predicate.
2212 bool Inverse = false;
2213 switch (I.getPredicate()) {
2214 case ICmpInst::ICMP_NE:
2216 case ICmpInst::ICMP_EQ:
2219 case ICmpInst::ICMP_UGT:
2220 case ICmpInst::ICMP_UGE:
2221 if (I.getOperand(0) == MulVal)
2225 case ICmpInst::ICMP_ULT:
2226 case ICmpInst::ICMP_ULE:
2227 if (I.getOperand(1) == MulVal)
2232 llvm_unreachable("Unexpected predicate");
2235 Value *Res = Builder->CreateExtractValue(Call, 1);
2236 return BinaryOperator::CreateNot(Res);
2239 return ExtractValueInst::Create(Call, 1);
2242 // DemandedBitsLHSMask - When performing a comparison against a constant,
2243 // it is possible that not all the bits in the LHS are demanded. This helper
2244 // method computes the mask that IS demanded.
2245 static APInt DemandedBitsLHSMask(ICmpInst &I,
2246 unsigned BitWidth, bool isSignCheck) {
2248 return APInt::getSignBit(BitWidth);
2250 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
2251 if (!CI) return APInt::getAllOnesValue(BitWidth);
2252 const APInt &RHS = CI->getValue();
2254 switch (I.getPredicate()) {
2255 // For a UGT comparison, we don't care about any bits that
2256 // correspond to the trailing ones of the comparand. The value of these
2257 // bits doesn't impact the outcome of the comparison, because any value
2258 // greater than the RHS must differ in a bit higher than these due to carry.
2259 case ICmpInst::ICMP_UGT: {
2260 unsigned trailingOnes = RHS.countTrailingOnes();
2261 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
2265 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
2266 // Any value less than the RHS must differ in a higher bit because of carries.
2267 case ICmpInst::ICMP_ULT: {
2268 unsigned trailingZeros = RHS.countTrailingZeros();
2269 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
2274 return APInt::getAllOnesValue(BitWidth);
2279 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst
2280 /// should be swapped.
2281 /// The decision is based on how many times these two operands are reused
2282 /// as subtract operands and their positions in those instructions.
2283 /// The rational is that several architectures use the same instruction for
2284 /// both subtract and cmp, thus it is better if the order of those operands
2286 /// \return true if Op0 and Op1 should be swapped.
2287 static bool swapMayExposeCSEOpportunities(const Value * Op0,
2288 const Value * Op1) {
2289 // Filter out pointer value as those cannot appears directly in subtract.
2290 // FIXME: we may want to go through inttoptrs or bitcasts.
2291 if (Op0->getType()->isPointerTy())
2293 // Count every uses of both Op0 and Op1 in a subtract.
2294 // Each time Op0 is the first operand, count -1: swapping is bad, the
2295 // subtract has already the same layout as the compare.
2296 // Each time Op0 is the second operand, count +1: swapping is good, the
2297 // subtract has a different layout as the compare.
2298 // At the end, if the benefit is greater than 0, Op0 should come second to
2299 // expose more CSE opportunities.
2300 int GlobalSwapBenefits = 0;
2301 for (const User *U : Op0->users()) {
2302 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U);
2303 if (!BinOp || BinOp->getOpcode() != Instruction::Sub)
2305 // If Op0 is the first argument, this is not beneficial to swap the
2307 int LocalSwapBenefits = -1;
2308 unsigned Op1Idx = 1;
2309 if (BinOp->getOperand(Op1Idx) == Op0) {
2311 LocalSwapBenefits = 1;
2313 if (BinOp->getOperand(Op1Idx) != Op1)
2315 GlobalSwapBenefits += LocalSwapBenefits;
2317 return GlobalSwapBenefits > 0;
2320 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
2321 bool Changed = false;
2322 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2323 unsigned Op0Cplxity = getComplexity(Op0);
2324 unsigned Op1Cplxity = getComplexity(Op1);
2326 /// Orders the operands of the compare so that they are listed from most
2327 /// complex to least complex. This puts constants before unary operators,
2328 /// before binary operators.
2329 if (Op0Cplxity < Op1Cplxity ||
2330 (Op0Cplxity == Op1Cplxity &&
2331 swapMayExposeCSEOpportunities(Op0, Op1))) {
2333 std::swap(Op0, Op1);
2337 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL))
2338 return ReplaceInstUsesWith(I, V);
2340 // comparing -val or val with non-zero is the same as just comparing val
2341 // ie, abs(val) != 0 -> val != 0
2342 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
2344 Value *Cond, *SelectTrue, *SelectFalse;
2345 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
2346 m_Value(SelectFalse)))) {
2347 if (Value *V = dyn_castNegVal(SelectTrue)) {
2348 if (V == SelectFalse)
2349 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2351 else if (Value *V = dyn_castNegVal(SelectFalse)) {
2352 if (V == SelectTrue)
2353 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
2358 Type *Ty = Op0->getType();
2360 // icmp's with boolean values can always be turned into bitwise operations
2361 if (Ty->isIntegerTy(1)) {
2362 switch (I.getPredicate()) {
2363 default: llvm_unreachable("Invalid icmp instruction!");
2364 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
2365 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
2366 return BinaryOperator::CreateNot(Xor);
2368 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
2369 return BinaryOperator::CreateXor(Op0, Op1);
2371 case ICmpInst::ICMP_UGT:
2372 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
2374 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
2375 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2376 return BinaryOperator::CreateAnd(Not, Op1);
2378 case ICmpInst::ICMP_SGT:
2379 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
2381 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
2382 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2383 return BinaryOperator::CreateAnd(Not, Op0);
2385 case ICmpInst::ICMP_UGE:
2386 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
2388 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
2389 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
2390 return BinaryOperator::CreateOr(Not, Op1);
2392 case ICmpInst::ICMP_SGE:
2393 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
2395 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
2396 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
2397 return BinaryOperator::CreateOr(Not, Op0);
2402 unsigned BitWidth = 0;
2403 if (Ty->isIntOrIntVectorTy())
2404 BitWidth = Ty->getScalarSizeInBits();
2405 else if (DL) // Pointers require DL info to get their size.
2406 BitWidth = DL->getTypeSizeInBits(Ty->getScalarType());
2408 bool isSignBit = false;
2410 // See if we are doing a comparison with a constant.
2411 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2412 Value *A = 0, *B = 0;
2414 // Match the following pattern, which is a common idiom when writing
2415 // overflow-safe integer arithmetic function. The source performs an
2416 // addition in wider type, and explicitly checks for overflow using
2417 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
2418 // sadd_with_overflow intrinsic.
2420 // TODO: This could probably be generalized to handle other overflow-safe
2421 // operations if we worked out the formulas to compute the appropriate
2425 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
2427 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
2428 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2429 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
2430 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
2434 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
2435 if (I.isEquality() && CI->isZero() &&
2436 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
2437 // (icmp cond A B) if cond is equality
2438 return new ICmpInst(I.getPredicate(), A, B);
2441 // If we have an icmp le or icmp ge instruction, turn it into the
2442 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
2443 // them being folded in the code below. The SimplifyICmpInst code has
2444 // already handled the edge cases for us, so we just assert on them.
2445 switch (I.getPredicate()) {
2447 case ICmpInst::ICMP_ULE:
2448 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
2449 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
2450 Builder->getInt(CI->getValue()+1));
2451 case ICmpInst::ICMP_SLE:
2452 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
2453 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2454 Builder->getInt(CI->getValue()+1));
2455 case ICmpInst::ICMP_UGE:
2456 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
2457 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
2458 Builder->getInt(CI->getValue()-1));
2459 case ICmpInst::ICMP_SGE:
2460 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
2461 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2462 Builder->getInt(CI->getValue()-1));
2465 // If this comparison is a normal comparison, it demands all
2466 // bits, if it is a sign bit comparison, it only demands the sign bit.
2468 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
2471 // See if we can fold the comparison based on range information we can get
2472 // by checking whether bits are known to be zero or one in the input.
2473 if (BitWidth != 0) {
2474 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
2475 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
2477 if (SimplifyDemandedBits(I.getOperandUse(0),
2478 DemandedBitsLHSMask(I, BitWidth, isSignBit),
2479 Op0KnownZero, Op0KnownOne, 0))
2481 if (SimplifyDemandedBits(I.getOperandUse(1),
2482 APInt::getAllOnesValue(BitWidth),
2483 Op1KnownZero, Op1KnownOne, 0))
2486 // Given the known and unknown bits, compute a range that the LHS could be
2487 // in. Compute the Min, Max and RHS values based on the known bits. For the
2488 // EQ and NE we use unsigned values.
2489 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
2490 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
2492 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2494 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2497 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
2499 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
2503 // If Min and Max are known to be the same, then SimplifyDemandedBits
2504 // figured out that the LHS is a constant. Just constant fold this now so
2505 // that code below can assume that Min != Max.
2506 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
2507 return new ICmpInst(I.getPredicate(),
2508 ConstantInt::get(Op0->getType(), Op0Min), Op1);
2509 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
2510 return new ICmpInst(I.getPredicate(), Op0,
2511 ConstantInt::get(Op1->getType(), Op1Min));
2513 // Based on the range information we know about the LHS, see if we can
2514 // simplify this comparison. For example, (x&4) < 8 is always true.
2515 switch (I.getPredicate()) {
2516 default: llvm_unreachable("Unknown icmp opcode!");
2517 case ICmpInst::ICMP_EQ: {
2518 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2519 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2521 // If all bits are known zero except for one, then we know at most one
2522 // bit is set. If the comparison is against zero, then this is a check
2523 // to see if *that* bit is set.
2524 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2525 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2526 // If the LHS is an AND with the same constant, look through it.
2528 ConstantInt *LHSC = 0;
2529 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2530 LHSC->getValue() != Op0KnownZeroInverted)
2533 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2534 // then turn "((1 << x)&8) == 0" into "x != 3".
2536 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2537 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2538 return new ICmpInst(ICmpInst::ICMP_NE, X,
2539 ConstantInt::get(X->getType(), CmpVal));
2542 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2543 // then turn "((8 >>u x)&1) == 0" into "x != 3".
2545 if (Op0KnownZeroInverted == 1 &&
2546 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2547 return new ICmpInst(ICmpInst::ICMP_NE, X,
2548 ConstantInt::get(X->getType(),
2549 CI->countTrailingZeros()));
2554 case ICmpInst::ICMP_NE: {
2555 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
2556 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2558 // If all bits are known zero except for one, then we know at most one
2559 // bit is set. If the comparison is against zero, then this is a check
2560 // to see if *that* bit is set.
2561 APInt Op0KnownZeroInverted = ~Op0KnownZero;
2562 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
2563 // If the LHS is an AND with the same constant, look through it.
2565 ConstantInt *LHSC = 0;
2566 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
2567 LHSC->getValue() != Op0KnownZeroInverted)
2570 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2571 // then turn "((1 << x)&8) != 0" into "x == 3".
2573 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2574 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2575 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2576 ConstantInt::get(X->getType(), CmpVal));
2579 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2580 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2582 if (Op0KnownZeroInverted == 1 &&
2583 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2584 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2585 ConstantInt::get(X->getType(),
2586 CI->countTrailingZeros()));
2591 case ICmpInst::ICMP_ULT:
2592 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2593 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2594 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2595 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2596 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2597 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2598 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2599 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2600 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2601 Builder->getInt(CI->getValue()-1));
2603 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2604 if (CI->isMinValue(true))
2605 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2606 Constant::getAllOnesValue(Op0->getType()));
2609 case ICmpInst::ICMP_UGT:
2610 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2611 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2612 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2613 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2615 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2616 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2617 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2618 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2619 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2620 Builder->getInt(CI->getValue()+1));
2622 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2623 if (CI->isMaxValue(true))
2624 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2625 Constant::getNullValue(Op0->getType()));
2628 case ICmpInst::ICMP_SLT:
2629 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2630 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2631 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2632 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2633 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2634 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2635 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2636 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2637 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2638 Builder->getInt(CI->getValue()-1));
2641 case ICmpInst::ICMP_SGT:
2642 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2643 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2644 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2645 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2647 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2648 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2649 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2650 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2651 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2652 Builder->getInt(CI->getValue()+1));
2655 case ICmpInst::ICMP_SGE:
2656 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2657 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2658 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2659 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2660 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2662 case ICmpInst::ICMP_SLE:
2663 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2664 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2665 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2666 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2667 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2669 case ICmpInst::ICMP_UGE:
2670 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2671 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2672 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2673 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2674 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2676 case ICmpInst::ICMP_ULE:
2677 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2678 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2679 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2680 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2681 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2685 // Turn a signed comparison into an unsigned one if both operands
2686 // are known to have the same sign.
2688 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2689 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2690 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2693 // Test if the ICmpInst instruction is used exclusively by a select as
2694 // part of a minimum or maximum operation. If so, refrain from doing
2695 // any other folding. This helps out other analyses which understand
2696 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2697 // and CodeGen. And in this case, at least one of the comparison
2698 // operands has at least one user besides the compare (the select),
2699 // which would often largely negate the benefit of folding anyway.
2701 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
2702 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2703 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2706 // See if we are doing a comparison between a constant and an instruction that
2707 // can be folded into the comparison.
2708 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2709 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2710 // instruction, see if that instruction also has constants so that the
2711 // instruction can be folded into the icmp
2712 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2713 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2717 // Handle icmp with constant (but not simple integer constant) RHS
2718 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2719 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2720 switch (LHSI->getOpcode()) {
2721 case Instruction::GetElementPtr:
2722 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2723 if (RHSC->isNullValue() &&
2724 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2725 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2726 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2728 case Instruction::PHI:
2729 // Only fold icmp into the PHI if the phi and icmp are in the same
2730 // block. If in the same block, we're encouraging jump threading. If
2731 // not, we are just pessimizing the code by making an i1 phi.
2732 if (LHSI->getParent() == I.getParent())
2733 if (Instruction *NV = FoldOpIntoPhi(I))
2736 case Instruction::Select: {
2737 // If either operand of the select is a constant, we can fold the
2738 // comparison into the select arms, which will cause one to be
2739 // constant folded and the select turned into a bitwise or.
2740 Value *Op1 = 0, *Op2 = 0;
2741 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2742 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2743 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2744 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2746 // We only want to perform this transformation if it will not lead to
2747 // additional code. This is true if either both sides of the select
2748 // fold to a constant (in which case the icmp is replaced with a select
2749 // which will usually simplify) or this is the only user of the
2750 // select (in which case we are trading a select+icmp for a simpler
2752 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2754 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2757 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2759 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2763 case Instruction::IntToPtr:
2764 // icmp pred inttoptr(X), null -> icmp pred X, 0
2765 if (RHSC->isNullValue() && DL &&
2766 DL->getIntPtrType(RHSC->getType()) ==
2767 LHSI->getOperand(0)->getType())
2768 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2769 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2772 case Instruction::Load:
2773 // Try to optimize things like "A[i] > 4" to index computations.
2774 if (GetElementPtrInst *GEP =
2775 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2776 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2777 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2778 !cast<LoadInst>(LHSI)->isVolatile())
2779 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2786 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2787 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2788 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2790 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2791 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2792 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2795 // Test to see if the operands of the icmp are casted versions of other
2796 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2798 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2799 if (Op0->getType()->isPointerTy() &&
2800 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2801 // We keep moving the cast from the left operand over to the right
2802 // operand, where it can often be eliminated completely.
2803 Op0 = CI->getOperand(0);
2805 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2806 // so eliminate it as well.
2807 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2808 Op1 = CI2->getOperand(0);
2810 // If Op1 is a constant, we can fold the cast into the constant.
2811 if (Op0->getType() != Op1->getType()) {
2812 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2813 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2815 // Otherwise, cast the RHS right before the icmp
2816 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2819 return new ICmpInst(I.getPredicate(), Op0, Op1);
2823 if (isa<CastInst>(Op0)) {
2824 // Handle the special case of: icmp (cast bool to X), <cst>
2825 // This comes up when you have code like
2828 // For generality, we handle any zero-extension of any operand comparison
2829 // with a constant or another cast from the same type.
2830 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2831 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2835 // Special logic for binary operators.
2836 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2837 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2839 CmpInst::Predicate Pred = I.getPredicate();
2840 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2841 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2842 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2843 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2844 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2845 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2846 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2847 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2848 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2850 // Analyze the case when either Op0 or Op1 is an add instruction.
2851 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2852 Value *A = 0, *B = 0, *C = 0, *D = 0;
2853 if (BO0 && BO0->getOpcode() == Instruction::Add)
2854 A = BO0->getOperand(0), B = BO0->getOperand(1);
2855 if (BO1 && BO1->getOpcode() == Instruction::Add)
2856 C = BO1->getOperand(0), D = BO1->getOperand(1);
2858 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2859 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2860 return new ICmpInst(Pred, A == Op1 ? B : A,
2861 Constant::getNullValue(Op1->getType()));
2863 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2864 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2865 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2868 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2869 if (A && C && (A == C || A == D || B == C || B == D) &&
2870 NoOp0WrapProblem && NoOp1WrapProblem &&
2871 // Try not to increase register pressure.
2872 BO0->hasOneUse() && BO1->hasOneUse()) {
2873 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2876 // C + B == C + D -> B == D
2879 } else if (A == D) {
2880 // D + B == C + D -> B == C
2883 } else if (B == C) {
2884 // A + C == C + D -> A == D
2889 // A + D == C + D -> A == C
2893 return new ICmpInst(Pred, Y, Z);
2896 // icmp slt (X + -1), Y -> icmp sle X, Y
2897 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
2898 match(B, m_AllOnes()))
2899 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
2901 // icmp sge (X + -1), Y -> icmp sgt X, Y
2902 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
2903 match(B, m_AllOnes()))
2904 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
2906 // icmp sle (X + 1), Y -> icmp slt X, Y
2907 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE &&
2909 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
2911 // icmp sgt (X + 1), Y -> icmp sge X, Y
2912 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT &&
2914 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
2916 // if C1 has greater magnitude than C2:
2917 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
2918 // s.t. C3 = C1 - C2
2920 // if C2 has greater magnitude than C1:
2921 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
2922 // s.t. C3 = C2 - C1
2923 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
2924 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
2925 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
2926 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
2927 const APInt &AP1 = C1->getValue();
2928 const APInt &AP2 = C2->getValue();
2929 if (AP1.isNegative() == AP2.isNegative()) {
2930 APInt AP1Abs = C1->getValue().abs();
2931 APInt AP2Abs = C2->getValue().abs();
2932 if (AP1Abs.uge(AP2Abs)) {
2933 ConstantInt *C3 = Builder->getInt(AP1 - AP2);
2934 Value *NewAdd = Builder->CreateNSWAdd(A, C3);
2935 return new ICmpInst(Pred, NewAdd, C);
2937 ConstantInt *C3 = Builder->getInt(AP2 - AP1);
2938 Value *NewAdd = Builder->CreateNSWAdd(C, C3);
2939 return new ICmpInst(Pred, A, NewAdd);
2945 // Analyze the case when either Op0 or Op1 is a sub instruction.
2946 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2947 A = 0; B = 0; C = 0; D = 0;
2948 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2949 A = BO0->getOperand(0), B = BO0->getOperand(1);
2950 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2951 C = BO1->getOperand(0), D = BO1->getOperand(1);
2953 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2954 if (A == Op1 && NoOp0WrapProblem)
2955 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2957 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2958 if (C == Op0 && NoOp1WrapProblem)
2959 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2961 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2962 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2963 // Try not to increase register pressure.
2964 BO0->hasOneUse() && BO1->hasOneUse())
2965 return new ICmpInst(Pred, A, C);
2967 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2968 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2969 // Try not to increase register pressure.
2970 BO0->hasOneUse() && BO1->hasOneUse())
2971 return new ICmpInst(Pred, D, B);
2973 BinaryOperator *SRem = NULL;
2974 // icmp (srem X, Y), Y
2975 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2976 Op1 == BO0->getOperand(1))
2978 // icmp Y, (srem X, Y)
2979 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2980 Op0 == BO1->getOperand(1))
2983 // We don't check hasOneUse to avoid increasing register pressure because
2984 // the value we use is the same value this instruction was already using.
2985 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2987 case ICmpInst::ICMP_EQ:
2988 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2989 case ICmpInst::ICMP_NE:
2990 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2991 case ICmpInst::ICMP_SGT:
2992 case ICmpInst::ICMP_SGE:
2993 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2994 Constant::getAllOnesValue(SRem->getType()));
2995 case ICmpInst::ICMP_SLT:
2996 case ICmpInst::ICMP_SLE:
2997 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2998 Constant::getNullValue(SRem->getType()));
3002 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
3003 BO0->hasOneUse() && BO1->hasOneUse() &&
3004 BO0->getOperand(1) == BO1->getOperand(1)) {
3005 switch (BO0->getOpcode()) {
3007 case Instruction::Add:
3008 case Instruction::Sub:
3009 case Instruction::Xor:
3010 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3011 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3012 BO1->getOperand(0));
3013 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
3014 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3015 if (CI->getValue().isSignBit()) {
3016 ICmpInst::Predicate Pred = I.isSigned()
3017 ? I.getUnsignedPredicate()
3018 : I.getSignedPredicate();
3019 return new ICmpInst(Pred, BO0->getOperand(0),
3020 BO1->getOperand(0));
3023 if (CI->isMaxValue(true)) {
3024 ICmpInst::Predicate Pred = I.isSigned()
3025 ? I.getUnsignedPredicate()
3026 : I.getSignedPredicate();
3027 Pred = I.getSwappedPredicate(Pred);
3028 return new ICmpInst(Pred, BO0->getOperand(0),
3029 BO1->getOperand(0));
3033 case Instruction::Mul:
3034 if (!I.isEquality())
3037 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
3038 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
3039 // Mask = -1 >> count-trailing-zeros(Cst).
3040 if (!CI->isZero() && !CI->isOne()) {
3041 const APInt &AP = CI->getValue();
3042 ConstantInt *Mask = ConstantInt::get(I.getContext(),
3043 APInt::getLowBitsSet(AP.getBitWidth(),
3045 AP.countTrailingZeros()));
3046 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
3047 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
3048 return new ICmpInst(I.getPredicate(), And1, And2);
3052 case Instruction::UDiv:
3053 case Instruction::LShr:
3057 case Instruction::SDiv:
3058 case Instruction::AShr:
3059 if (!BO0->isExact() || !BO1->isExact())
3061 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3062 BO1->getOperand(0));
3063 case Instruction::Shl: {
3064 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3065 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3068 if (!NSW && I.isSigned())
3070 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
3071 BO1->getOperand(0));
3078 // Transform (A & ~B) == 0 --> (A & B) != 0
3079 // and (A & ~B) != 0 --> (A & B) == 0
3080 // if A is a power of 2.
3081 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
3082 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality())
3083 return new ICmpInst(I.getInversePredicate(),
3084 Builder->CreateAnd(A, B),
3087 // ~x < ~y --> y < x
3088 // ~x < cst --> ~cst < x
3089 if (match(Op0, m_Not(m_Value(A)))) {
3090 if (match(Op1, m_Not(m_Value(B))))
3091 return new ICmpInst(I.getPredicate(), B, A);
3092 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
3093 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
3096 // (a+b) <u a --> llvm.uadd.with.overflow.
3097 // (a+b) <u b --> llvm.uadd.with.overflow.
3098 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
3099 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
3100 (Op1 == A || Op1 == B))
3101 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
3104 // a >u (a+b) --> llvm.uadd.with.overflow.
3105 // b >u (a+b) --> llvm.uadd.with.overflow.
3106 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
3107 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
3108 (Op0 == A || Op0 == B))
3109 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
3112 // (zext a) * (zext b) --> llvm.umul.with.overflow.
3113 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3114 if (Instruction *R = ProcessUMulZExtIdiom(I, Op0, Op1, *this))
3117 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
3118 if (Instruction *R = ProcessUMulZExtIdiom(I, Op1, Op0, *this))
3123 if (I.isEquality()) {
3124 Value *A, *B, *C, *D;
3126 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3127 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3128 Value *OtherVal = A == Op1 ? B : A;
3129 return new ICmpInst(I.getPredicate(), OtherVal,
3130 Constant::getNullValue(A->getType()));
3133 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3134 // A^c1 == C^c2 --> A == C^(c1^c2)
3135 ConstantInt *C1, *C2;
3136 if (match(B, m_ConstantInt(C1)) &&
3137 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
3138 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue());
3139 Value *Xor = Builder->CreateXor(C, NC);
3140 return new ICmpInst(I.getPredicate(), A, Xor);
3143 // A^B == A^D -> B == D
3144 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
3145 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
3146 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
3147 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
3151 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
3152 (A == Op0 || B == Op0)) {
3153 // A == (A^B) -> B == 0
3154 Value *OtherVal = A == Op0 ? B : A;
3155 return new ICmpInst(I.getPredicate(), OtherVal,
3156 Constant::getNullValue(A->getType()));
3159 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3160 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3161 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3162 Value *X = 0, *Y = 0, *Z = 0;
3165 X = B; Y = D; Z = A;
3166 } else if (A == D) {
3167 X = B; Y = C; Z = A;
3168 } else if (B == C) {
3169 X = A; Y = D; Z = B;
3170 } else if (B == D) {
3171 X = A; Y = C; Z = B;
3174 if (X) { // Build (X^Y) & Z
3175 Op1 = Builder->CreateXor(X, Y);
3176 Op1 = Builder->CreateAnd(Op1, Z);
3177 I.setOperand(0, Op1);
3178 I.setOperand(1, Constant::getNullValue(Op1->getType()));
3183 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3184 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3186 if ((Op0->hasOneUse() &&
3187 match(Op0, m_ZExt(m_Value(A))) &&
3188 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3189 (Op1->hasOneUse() &&
3190 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3191 match(Op1, m_ZExt(m_Value(A))))) {
3192 APInt Pow2 = Cst1->getValue() + 1;
3193 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3194 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3195 return new ICmpInst(I.getPredicate(), A,
3196 Builder->CreateTrunc(B, A->getType()));
3199 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3200 // For lshr and ashr pairs.
3201 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3202 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3203 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3204 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3205 unsigned TypeBits = Cst1->getBitWidth();
3206 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3207 if (ShAmt < TypeBits && ShAmt != 0) {
3208 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE
3209 ? ICmpInst::ICMP_UGE
3210 : ICmpInst::ICMP_ULT;
3211 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted");
3212 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3213 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal));
3217 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3218 // "icmp (and X, mask), cst"
3220 if (Op0->hasOneUse() &&
3221 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
3222 m_ConstantInt(ShAmt))))) &&
3223 match(Op1, m_ConstantInt(Cst1)) &&
3224 // Only do this when A has multiple uses. This is most important to do
3225 // when it exposes other optimizations.
3227 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3229 if (ShAmt < ASize) {
3231 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
3234 APInt CmpV = Cst1->getValue().zext(ASize);
3237 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
3238 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
3244 Value *X; ConstantInt *Cst;
3246 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
3247 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate());
3250 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
3251 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
3253 return Changed ? &I : 0;
3256 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
3258 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
3261 if (!isa<ConstantFP>(RHSC)) return 0;
3262 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
3264 // Get the width of the mantissa. We don't want to hack on conversions that
3265 // might lose information from the integer, e.g. "i64 -> float"
3266 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
3267 if (MantissaWidth == -1) return 0; // Unknown.
3269 // Check to see that the input is converted from an integer type that is small
3270 // enough that preserves all bits. TODO: check here for "known" sign bits.
3271 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
3272 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
3274 // If this is a uitofp instruction, we need an extra bit to hold the sign.
3275 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
3279 // If the conversion would lose info, don't hack on this.
3280 if ((int)InputSize > MantissaWidth)
3283 // Otherwise, we can potentially simplify the comparison. We know that it
3284 // will always come through as an integer value and we know the constant is
3285 // not a NAN (it would have been previously simplified).
3286 assert(!RHS.isNaN() && "NaN comparison not already folded!");
3288 ICmpInst::Predicate Pred;
3289 switch (I.getPredicate()) {
3290 default: llvm_unreachable("Unexpected predicate!");
3291 case FCmpInst::FCMP_UEQ:
3292 case FCmpInst::FCMP_OEQ:
3293 Pred = ICmpInst::ICMP_EQ;
3295 case FCmpInst::FCMP_UGT:
3296 case FCmpInst::FCMP_OGT:
3297 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
3299 case FCmpInst::FCMP_UGE:
3300 case FCmpInst::FCMP_OGE:
3301 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
3303 case FCmpInst::FCMP_ULT:
3304 case FCmpInst::FCMP_OLT:
3305 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
3307 case FCmpInst::FCMP_ULE:
3308 case FCmpInst::FCMP_OLE:
3309 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
3311 case FCmpInst::FCMP_UNE:
3312 case FCmpInst::FCMP_ONE:
3313 Pred = ICmpInst::ICMP_NE;
3315 case FCmpInst::FCMP_ORD:
3316 return ReplaceInstUsesWith(I, Builder->getTrue());
3317 case FCmpInst::FCMP_UNO:
3318 return ReplaceInstUsesWith(I, Builder->getFalse());
3321 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
3323 // Now we know that the APFloat is a normal number, zero or inf.
3325 // See if the FP constant is too large for the integer. For example,
3326 // comparing an i8 to 300.0.
3327 unsigned IntWidth = IntTy->getScalarSizeInBits();
3330 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
3331 // and large values.
3332 APFloat SMax(RHS.getSemantics());
3333 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
3334 APFloat::rmNearestTiesToEven);
3335 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
3336 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
3337 Pred == ICmpInst::ICMP_SLE)
3338 return ReplaceInstUsesWith(I, Builder->getTrue());
3339 return ReplaceInstUsesWith(I, Builder->getFalse());
3342 // If the RHS value is > UnsignedMax, fold the comparison. This handles
3343 // +INF and large values.
3344 APFloat UMax(RHS.getSemantics());
3345 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
3346 APFloat::rmNearestTiesToEven);
3347 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
3348 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
3349 Pred == ICmpInst::ICMP_ULE)
3350 return ReplaceInstUsesWith(I, Builder->getTrue());
3351 return ReplaceInstUsesWith(I, Builder->getFalse());
3356 // See if the RHS value is < SignedMin.
3357 APFloat SMin(RHS.getSemantics());
3358 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
3359 APFloat::rmNearestTiesToEven);
3360 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
3361 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
3362 Pred == ICmpInst::ICMP_SGE)
3363 return ReplaceInstUsesWith(I, Builder->getTrue());
3364 return ReplaceInstUsesWith(I, Builder->getFalse());
3367 // See if the RHS value is < UnsignedMin.
3368 APFloat SMin(RHS.getSemantics());
3369 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
3370 APFloat::rmNearestTiesToEven);
3371 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
3372 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
3373 Pred == ICmpInst::ICMP_UGE)
3374 return ReplaceInstUsesWith(I, Builder->getTrue());
3375 return ReplaceInstUsesWith(I, Builder->getFalse());
3379 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
3380 // [0, UMAX], but it may still be fractional. See if it is fractional by
3381 // casting the FP value to the integer value and back, checking for equality.
3382 // Don't do this for zero, because -0.0 is not fractional.
3383 Constant *RHSInt = LHSUnsigned
3384 ? ConstantExpr::getFPToUI(RHSC, IntTy)
3385 : ConstantExpr::getFPToSI(RHSC, IntTy);
3386 if (!RHS.isZero()) {
3387 bool Equal = LHSUnsigned
3388 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
3389 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
3391 // If we had a comparison against a fractional value, we have to adjust
3392 // the compare predicate and sometimes the value. RHSC is rounded towards
3393 // zero at this point.
3395 default: llvm_unreachable("Unexpected integer comparison!");
3396 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
3397 return ReplaceInstUsesWith(I, Builder->getTrue());
3398 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
3399 return ReplaceInstUsesWith(I, Builder->getFalse());
3400 case ICmpInst::ICMP_ULE:
3401 // (float)int <= 4.4 --> int <= 4
3402 // (float)int <= -4.4 --> false
3403 if (RHS.isNegative())
3404 return ReplaceInstUsesWith(I, Builder->getFalse());
3406 case ICmpInst::ICMP_SLE:
3407 // (float)int <= 4.4 --> int <= 4
3408 // (float)int <= -4.4 --> int < -4
3409 if (RHS.isNegative())
3410 Pred = ICmpInst::ICMP_SLT;
3412 case ICmpInst::ICMP_ULT:
3413 // (float)int < -4.4 --> false
3414 // (float)int < 4.4 --> int <= 4
3415 if (RHS.isNegative())
3416 return ReplaceInstUsesWith(I, Builder->getFalse());
3417 Pred = ICmpInst::ICMP_ULE;
3419 case ICmpInst::ICMP_SLT:
3420 // (float)int < -4.4 --> int < -4
3421 // (float)int < 4.4 --> int <= 4
3422 if (!RHS.isNegative())
3423 Pred = ICmpInst::ICMP_SLE;
3425 case ICmpInst::ICMP_UGT:
3426 // (float)int > 4.4 --> int > 4
3427 // (float)int > -4.4 --> true
3428 if (RHS.isNegative())
3429 return ReplaceInstUsesWith(I, Builder->getTrue());
3431 case ICmpInst::ICMP_SGT:
3432 // (float)int > 4.4 --> int > 4
3433 // (float)int > -4.4 --> int >= -4
3434 if (RHS.isNegative())
3435 Pred = ICmpInst::ICMP_SGE;
3437 case ICmpInst::ICMP_UGE:
3438 // (float)int >= -4.4 --> true
3439 // (float)int >= 4.4 --> int > 4
3440 if (RHS.isNegative())
3441 return ReplaceInstUsesWith(I, Builder->getTrue());
3442 Pred = ICmpInst::ICMP_UGT;
3444 case ICmpInst::ICMP_SGE:
3445 // (float)int >= -4.4 --> int >= -4
3446 // (float)int >= 4.4 --> int > 4
3447 if (!RHS.isNegative())
3448 Pred = ICmpInst::ICMP_SGT;
3454 // Lower this FP comparison into an appropriate integer version of the
3456 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
3459 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
3460 bool Changed = false;
3462 /// Orders the operands of the compare so that they are listed from most
3463 /// complex to least complex. This puts constants before unary operators,
3464 /// before binary operators.
3465 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
3470 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3472 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL))
3473 return ReplaceInstUsesWith(I, V);
3475 // Simplify 'fcmp pred X, X'
3477 switch (I.getPredicate()) {
3478 default: llvm_unreachable("Unknown predicate!");
3479 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
3480 case FCmpInst::FCMP_ULT: // True if unordered or less than
3481 case FCmpInst::FCMP_UGT: // True if unordered or greater than
3482 case FCmpInst::FCMP_UNE: // True if unordered or not equal
3483 // Canonicalize these to be 'fcmp uno %X, 0.0'.
3484 I.setPredicate(FCmpInst::FCMP_UNO);
3485 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3488 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
3489 case FCmpInst::FCMP_OEQ: // True if ordered and equal
3490 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
3491 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
3492 // Canonicalize these to be 'fcmp ord %X, 0.0'.
3493 I.setPredicate(FCmpInst::FCMP_ORD);
3494 I.setOperand(1, Constant::getNullValue(Op0->getType()));
3499 // Handle fcmp with constant RHS
3500 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
3501 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
3502 switch (LHSI->getOpcode()) {
3503 case Instruction::FPExt: {
3504 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
3505 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
3506 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
3510 const fltSemantics *Sem;
3511 // FIXME: This shouldn't be here.
3512 if (LHSExt->getSrcTy()->isHalfTy())
3513 Sem = &APFloat::IEEEhalf;
3514 else if (LHSExt->getSrcTy()->isFloatTy())
3515 Sem = &APFloat::IEEEsingle;
3516 else if (LHSExt->getSrcTy()->isDoubleTy())
3517 Sem = &APFloat::IEEEdouble;
3518 else if (LHSExt->getSrcTy()->isFP128Ty())
3519 Sem = &APFloat::IEEEquad;
3520 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
3521 Sem = &APFloat::x87DoubleExtended;
3522 else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
3523 Sem = &APFloat::PPCDoubleDouble;
3528 APFloat F = RHSF->getValueAPF();
3529 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
3531 // Avoid lossy conversions and denormals. Zero is a special case
3532 // that's OK to convert.
3536 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
3537 APFloat::cmpLessThan) || Fabs.isZero()))
3539 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3540 ConstantFP::get(RHSC->getContext(), F));
3543 case Instruction::PHI:
3544 // Only fold fcmp into the PHI if the phi and fcmp are in the same
3545 // block. If in the same block, we're encouraging jump threading. If
3546 // not, we are just pessimizing the code by making an i1 phi.
3547 if (LHSI->getParent() == I.getParent())
3548 if (Instruction *NV = FoldOpIntoPhi(I))
3551 case Instruction::SIToFP:
3552 case Instruction::UIToFP:
3553 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
3556 case Instruction::FSub: {
3557 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
3559 if (match(LHSI, m_FNeg(m_Value(Op))))
3560 return new FCmpInst(I.getSwappedPredicate(), Op,
3561 ConstantExpr::getFNeg(RHSC));
3564 case Instruction::Load:
3565 if (GetElementPtrInst *GEP =
3566 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3567 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3568 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3569 !cast<LoadInst>(LHSI)->isVolatile())
3570 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
3574 case Instruction::Call: {
3575 CallInst *CI = cast<CallInst>(LHSI);
3577 // Various optimization for fabs compared with zero.
3578 if (RHSC->isNullValue() && CI->getCalledFunction() &&
3579 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
3581 if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
3582 Func == LibFunc::fabsl) {
3583 switch (I.getPredicate()) {
3585 // fabs(x) < 0 --> false
3586 case FCmpInst::FCMP_OLT:
3587 return ReplaceInstUsesWith(I, Builder->getFalse());
3588 // fabs(x) > 0 --> x != 0
3589 case FCmpInst::FCMP_OGT:
3590 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
3592 // fabs(x) <= 0 --> x == 0
3593 case FCmpInst::FCMP_OLE:
3594 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
3596 // fabs(x) >= 0 --> !isnan(x)
3597 case FCmpInst::FCMP_OGE:
3598 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
3600 // fabs(x) == 0 --> x == 0
3601 // fabs(x) != 0 --> x != 0
3602 case FCmpInst::FCMP_OEQ:
3603 case FCmpInst::FCMP_UEQ:
3604 case FCmpInst::FCMP_ONE:
3605 case FCmpInst::FCMP_UNE:
3606 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
3615 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
3617 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
3618 return new FCmpInst(I.getSwappedPredicate(), X, Y);
3620 // fcmp (fpext x), (fpext y) -> fcmp x, y
3621 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
3622 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
3623 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
3624 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
3625 RHSExt->getOperand(0));
3627 return Changed ? &I : 0;