1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
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 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "InstCombine.h"
39 #include "llvm/IntrinsicInst.h"
40 #include "llvm/LLVMContext.h"
41 #include "llvm/DerivedTypes.h"
42 #include "llvm/GlobalVariable.h"
43 #include "llvm/Operator.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/InstructionSimplify.h"
46 #include "llvm/Analysis/MemoryBuiltins.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Support/CallSite.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/GetElementPtrTypeIterator.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Support/PatternMatch.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/STLExtras.h"
62 using namespace llvm::PatternMatch;
64 STATISTIC(NumCombined , "Number of insts combined");
65 STATISTIC(NumConstProp, "Number of constant folds");
66 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
67 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
68 STATISTIC(NumSunkInst , "Number of instructions sunk");
71 char InstCombiner::ID = 0;
72 static RegisterPass<InstCombiner>
73 X("instcombine", "Combine redundant instructions");
75 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
76 AU.addPreservedID(LCSSAID);
81 // isOnlyUse - Return true if this instruction will be deleted if we stop using
83 static bool isOnlyUse(Value *V) {
84 return V->hasOneUse() || isa<Constant>(V);
87 // getPromotedType - Return the specified type promoted as it would be to pass
88 // though a va_arg area...
89 static const Type *getPromotedType(const Type *Ty) {
90 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
91 if (ITy->getBitWidth() < 32)
92 return Type::getInt32Ty(Ty->getContext());
97 /// ShouldChangeType - Return true if it is desirable to convert a computation
98 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
99 /// type for example, or from a smaller to a larger illegal type.
100 bool InstCombiner::ShouldChangeType(const Type *From, const Type *To) const {
101 assert(isa<IntegerType>(From) && isa<IntegerType>(To));
103 // If we don't have TD, we don't know if the source/dest are legal.
104 if (!TD) return false;
106 unsigned FromWidth = From->getPrimitiveSizeInBits();
107 unsigned ToWidth = To->getPrimitiveSizeInBits();
108 bool FromLegal = TD->isLegalInteger(FromWidth);
109 bool ToLegal = TD->isLegalInteger(ToWidth);
111 // If this is a legal integer from type, and the result would be an illegal
112 // type, don't do the transformation.
113 if (FromLegal && !ToLegal)
116 // Otherwise, if both are illegal, do not increase the size of the result. We
117 // do allow things like i160 -> i64, but not i64 -> i160.
118 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
124 /// getBitCastOperand - If the specified operand is a CastInst, a constant
125 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
126 /// operand value, otherwise return null.
127 static Value *getBitCastOperand(Value *V) {
128 if (Operator *O = dyn_cast<Operator>(V)) {
129 if (O->getOpcode() == Instruction::BitCast)
130 return O->getOperand(0);
131 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
132 if (GEP->hasAllZeroIndices())
133 return GEP->getPointerOperand();
140 // SimplifyCommutative - This performs a few simplifications for commutative
143 // 1. Order operands such that they are listed from right (least complex) to
144 // left (most complex). This puts constants before unary operators before
147 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
148 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
150 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
151 bool Changed = false;
152 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
153 Changed = !I.swapOperands();
155 if (!I.isAssociative()) return Changed;
156 Instruction::BinaryOps Opcode = I.getOpcode();
157 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
158 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
159 if (isa<Constant>(I.getOperand(1))) {
160 Constant *Folded = ConstantExpr::get(I.getOpcode(),
161 cast<Constant>(I.getOperand(1)),
162 cast<Constant>(Op->getOperand(1)));
163 I.setOperand(0, Op->getOperand(0));
164 I.setOperand(1, Folded);
166 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
167 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
168 isOnlyUse(Op) && isOnlyUse(Op1)) {
169 Constant *C1 = cast<Constant>(Op->getOperand(1));
170 Constant *C2 = cast<Constant>(Op1->getOperand(1));
172 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
173 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
174 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
178 I.setOperand(0, New);
179 I.setOperand(1, Folded);
186 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
187 // if the LHS is a constant zero (which is the 'negate' form).
189 Value *InstCombiner::dyn_castNegVal(Value *V) const {
190 if (BinaryOperator::isNeg(V))
191 return BinaryOperator::getNegArgument(V);
193 // Constants can be considered to be negated values if they can be folded.
194 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
195 return ConstantExpr::getNeg(C);
197 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
198 if (C->getType()->getElementType()->isInteger())
199 return ConstantExpr::getNeg(C);
204 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
205 // instruction if the LHS is a constant negative zero (which is the 'negate'
208 static inline Value *dyn_castFNegVal(Value *V) {
209 if (BinaryOperator::isFNeg(V))
210 return BinaryOperator::getFNegArgument(V);
212 // Constants can be considered to be negated values if they can be folded.
213 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
214 return ConstantExpr::getFNeg(C);
216 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
217 if (C->getType()->getElementType()->isFloatingPoint())
218 return ConstantExpr::getFNeg(C);
223 /// MatchSelectPattern - Pattern match integer [SU]MIN, [SU]MAX, and ABS idioms,
224 /// returning the kind and providing the out parameter results if we
225 /// successfully match.
226 static SelectPatternFlavor
227 MatchSelectPattern(Value *V, Value *&LHS, Value *&RHS) {
228 SelectInst *SI = dyn_cast<SelectInst>(V);
229 if (SI == 0) return SPF_UNKNOWN;
231 ICmpInst *ICI = dyn_cast<ICmpInst>(SI->getCondition());
232 if (ICI == 0) return SPF_UNKNOWN;
234 LHS = ICI->getOperand(0);
235 RHS = ICI->getOperand(1);
237 // (icmp X, Y) ? X : Y
238 if (SI->getTrueValue() == ICI->getOperand(0) &&
239 SI->getFalseValue() == ICI->getOperand(1)) {
240 switch (ICI->getPredicate()) {
241 default: return SPF_UNKNOWN; // Equality.
242 case ICmpInst::ICMP_UGT:
243 case ICmpInst::ICMP_UGE: return SPF_UMAX;
244 case ICmpInst::ICMP_SGT:
245 case ICmpInst::ICMP_SGE: return SPF_SMAX;
246 case ICmpInst::ICMP_ULT:
247 case ICmpInst::ICMP_ULE: return SPF_UMIN;
248 case ICmpInst::ICMP_SLT:
249 case ICmpInst::ICMP_SLE: return SPF_SMIN;
253 // (icmp X, Y) ? Y : X
254 if (SI->getTrueValue() == ICI->getOperand(1) &&
255 SI->getFalseValue() == ICI->getOperand(0)) {
256 switch (ICI->getPredicate()) {
257 default: return SPF_UNKNOWN; // Equality.
258 case ICmpInst::ICMP_UGT:
259 case ICmpInst::ICMP_UGE: return SPF_UMIN;
260 case ICmpInst::ICMP_SGT:
261 case ICmpInst::ICMP_SGE: return SPF_SMIN;
262 case ICmpInst::ICMP_ULT:
263 case ICmpInst::ICMP_ULE: return SPF_UMAX;
264 case ICmpInst::ICMP_SLT:
265 case ICmpInst::ICMP_SLE: return SPF_SMAX;
269 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
274 /// isFreeToInvert - Return true if the specified value is free to invert (apply
275 /// ~ to). This happens in cases where the ~ can be eliminated.
276 static inline bool isFreeToInvert(Value *V) {
278 if (BinaryOperator::isNot(V))
281 // Constants can be considered to be not'ed values.
282 if (isa<ConstantInt>(V))
285 // Compares can be inverted if they have a single use.
286 if (CmpInst *CI = dyn_cast<CmpInst>(V))
287 return CI->hasOneUse();
292 static inline Value *dyn_castNotVal(Value *V) {
293 // If this is not(not(x)) don't return that this is a not: we want the two
294 // not's to be folded first.
295 if (BinaryOperator::isNot(V)) {
296 Value *Operand = BinaryOperator::getNotArgument(V);
297 if (!isFreeToInvert(Operand))
301 // Constants can be considered to be not'ed values...
302 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
303 return ConstantInt::get(C->getType(), ~C->getValue());
309 // dyn_castFoldableMul - If this value is a multiply that can be folded into
310 // other computations (because it has a constant operand), return the
311 // non-constant operand of the multiply, and set CST to point to the multiplier.
312 // Otherwise, return null.
314 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
315 if (V->hasOneUse() && V->getType()->isInteger())
316 if (Instruction *I = dyn_cast<Instruction>(V)) {
317 if (I->getOpcode() == Instruction::Mul)
318 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
319 return I->getOperand(0);
320 if (I->getOpcode() == Instruction::Shl)
321 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
322 // The multiplier is really 1 << CST.
323 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
324 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
325 CST = ConstantInt::get(V->getType()->getContext(),
326 APInt(BitWidth, 1).shl(CSTVal));
327 return I->getOperand(0);
333 /// AddOne - Add one to a ConstantInt
334 static Constant *AddOne(Constant *C) {
335 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
337 /// SubOne - Subtract one from a ConstantInt
338 static Constant *SubOne(ConstantInt *C) {
339 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
341 /// MultiplyOverflows - True if the multiply can not be expressed in an int
343 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
344 uint32_t W = C1->getBitWidth();
345 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
354 APInt MulExt = LHSExt * RHSExt;
357 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
359 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
360 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
361 return MulExt.slt(Min) || MulExt.sgt(Max);
366 /// AssociativeOpt - Perform an optimization on an associative operator. This
367 /// function is designed to check a chain of associative operators for a
368 /// potential to apply a certain optimization. Since the optimization may be
369 /// applicable if the expression was reassociated, this checks the chain, then
370 /// reassociates the expression as necessary to expose the optimization
371 /// opportunity. This makes use of a special Functor, which must define
372 /// 'shouldApply' and 'apply' methods.
374 template<typename Functor>
375 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
376 unsigned Opcode = Root.getOpcode();
377 Value *LHS = Root.getOperand(0);
379 // Quick check, see if the immediate LHS matches...
380 if (F.shouldApply(LHS))
381 return F.apply(Root);
383 // Otherwise, if the LHS is not of the same opcode as the root, return.
384 Instruction *LHSI = dyn_cast<Instruction>(LHS);
385 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
386 // Should we apply this transform to the RHS?
387 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
389 // If not to the RHS, check to see if we should apply to the LHS...
390 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
391 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
395 // If the functor wants to apply the optimization to the RHS of LHSI,
396 // reassociate the expression from ((? op A) op B) to (? op (A op B))
398 // Now all of the instructions are in the current basic block, go ahead
399 // and perform the reassociation.
400 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
402 // First move the selected RHS to the LHS of the root...
403 Root.setOperand(0, LHSI->getOperand(1));
405 // Make what used to be the LHS of the root be the user of the root...
406 Value *ExtraOperand = TmpLHSI->getOperand(1);
407 if (&Root == TmpLHSI) {
408 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
411 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
412 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
413 BasicBlock::iterator ARI = &Root; ++ARI;
414 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
417 // Now propagate the ExtraOperand down the chain of instructions until we
419 while (TmpLHSI != LHSI) {
420 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
421 // Move the instruction to immediately before the chain we are
422 // constructing to avoid breaking dominance properties.
423 NextLHSI->moveBefore(ARI);
426 Value *NextOp = NextLHSI->getOperand(1);
427 NextLHSI->setOperand(1, ExtraOperand);
429 ExtraOperand = NextOp;
432 // Now that the instructions are reassociated, have the functor perform
433 // the transformation...
434 return F.apply(Root);
437 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
444 // AddRHS - Implements: X + X --> X << 1
447 explicit AddRHS(Value *rhs) : RHS(rhs) {}
448 bool shouldApply(Value *LHS) const { return LHS == RHS; }
449 Instruction *apply(BinaryOperator &Add) const {
450 return BinaryOperator::CreateShl(Add.getOperand(0),
451 ConstantInt::get(Add.getType(), 1));
455 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
457 struct AddMaskingAnd {
459 explicit AddMaskingAnd(Constant *c) : C2(c) {}
460 bool shouldApply(Value *LHS) const {
462 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
463 ConstantExpr::getAnd(C1, C2)->isNullValue();
465 Instruction *apply(BinaryOperator &Add) const {
466 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
472 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
474 if (CastInst *CI = dyn_cast<CastInst>(&I))
475 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
477 // Figure out if the constant is the left or the right argument.
478 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
479 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
481 if (Constant *SOC = dyn_cast<Constant>(SO)) {
483 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
484 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
487 Value *Op0 = SO, *Op1 = ConstOperand;
491 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
492 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
493 SO->getName()+".op");
494 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
495 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
496 SO->getName()+".cmp");
497 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
498 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
499 SO->getName()+".cmp");
500 llvm_unreachable("Unknown binary instruction type!");
503 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
504 // constant as the other operand, try to fold the binary operator into the
505 // select arguments. This also works for Cast instructions, which obviously do
506 // not have a second operand.
507 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
508 // Don't modify shared select instructions
509 if (!SI->hasOneUse()) return 0;
510 Value *TV = SI->getOperand(1);
511 Value *FV = SI->getOperand(2);
513 if (isa<Constant>(TV) || isa<Constant>(FV)) {
514 // Bool selects with constant operands can be folded to logical ops.
515 if (SI->getType() == Type::getInt1Ty(SI->getContext())) return 0;
517 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
518 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
520 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
527 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
528 /// has a PHI node as operand #0, see if we can fold the instruction into the
529 /// PHI (which is only possible if all operands to the PHI are constants).
531 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
532 /// that would normally be unprofitable because they strongly encourage jump
534 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
535 bool AllowAggressive) {
536 AllowAggressive = false;
537 PHINode *PN = cast<PHINode>(I.getOperand(0));
538 unsigned NumPHIValues = PN->getNumIncomingValues();
539 if (NumPHIValues == 0 ||
540 // We normally only transform phis with a single use, unless we're trying
541 // hard to make jump threading happen.
542 (!PN->hasOneUse() && !AllowAggressive))
546 // Check to see if all of the operands of the PHI are simple constants
547 // (constantint/constantfp/undef). If there is one non-constant value,
548 // remember the BB it is in. If there is more than one or if *it* is a PHI,
549 // bail out. We don't do arbitrary constant expressions here because moving
550 // their computation can be expensive without a cost model.
551 BasicBlock *NonConstBB = 0;
552 for (unsigned i = 0; i != NumPHIValues; ++i)
553 if (!isa<Constant>(PN->getIncomingValue(i)) ||
554 isa<ConstantExpr>(PN->getIncomingValue(i))) {
555 if (NonConstBB) return 0; // More than one non-const value.
556 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
557 NonConstBB = PN->getIncomingBlock(i);
559 // If the incoming non-constant value is in I's block, we have an infinite
561 if (NonConstBB == I.getParent())
565 // If there is exactly one non-constant value, we can insert a copy of the
566 // operation in that block. However, if this is a critical edge, we would be
567 // inserting the computation one some other paths (e.g. inside a loop). Only
568 // do this if the pred block is unconditionally branching into the phi block.
569 if (NonConstBB != 0 && !AllowAggressive) {
570 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
571 if (!BI || !BI->isUnconditional()) return 0;
574 // Okay, we can do the transformation: create the new PHI node.
575 PHINode *NewPN = PHINode::Create(I.getType(), "");
576 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
577 InsertNewInstBefore(NewPN, *PN);
580 // Next, add all of the operands to the PHI.
581 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
582 // We only currently try to fold the condition of a select when it is a phi,
583 // not the true/false values.
584 Value *TrueV = SI->getTrueValue();
585 Value *FalseV = SI->getFalseValue();
586 BasicBlock *PhiTransBB = PN->getParent();
587 for (unsigned i = 0; i != NumPHIValues; ++i) {
588 BasicBlock *ThisBB = PN->getIncomingBlock(i);
589 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
590 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
592 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
593 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
595 assert(PN->getIncomingBlock(i) == NonConstBB);
596 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
598 "phitmp", NonConstBB->getTerminator());
599 Worklist.Add(cast<Instruction>(InV));
601 NewPN->addIncoming(InV, ThisBB);
603 } else if (I.getNumOperands() == 2) {
604 Constant *C = cast<Constant>(I.getOperand(1));
605 for (unsigned i = 0; i != NumPHIValues; ++i) {
607 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
608 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
609 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
611 InV = ConstantExpr::get(I.getOpcode(), InC, C);
613 assert(PN->getIncomingBlock(i) == NonConstBB);
614 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
615 InV = BinaryOperator::Create(BO->getOpcode(),
616 PN->getIncomingValue(i), C, "phitmp",
617 NonConstBB->getTerminator());
618 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
619 InV = CmpInst::Create(CI->getOpcode(),
621 PN->getIncomingValue(i), C, "phitmp",
622 NonConstBB->getTerminator());
624 llvm_unreachable("Unknown binop!");
626 Worklist.Add(cast<Instruction>(InV));
628 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
631 CastInst *CI = cast<CastInst>(&I);
632 const Type *RetTy = CI->getType();
633 for (unsigned i = 0; i != NumPHIValues; ++i) {
635 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
636 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
638 assert(PN->getIncomingBlock(i) == NonConstBB);
639 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
640 I.getType(), "phitmp",
641 NonConstBB->getTerminator());
642 Worklist.Add(cast<Instruction>(InV));
644 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
647 return ReplaceInstUsesWith(I, NewPN);
651 /// WillNotOverflowSignedAdd - Return true if we can prove that:
652 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
653 /// This basically requires proving that the add in the original type would not
654 /// overflow to change the sign bit or have a carry out.
655 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
656 // There are different heuristics we can use for this. Here are some simple
659 // Add has the property that adding any two 2's complement numbers can only
660 // have one carry bit which can change a sign. As such, if LHS and RHS each
661 // have at least two sign bits, we know that the addition of the two values
662 // will sign extend fine.
663 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
667 // If one of the operands only has one non-zero bit, and if the other operand
668 // has a known-zero bit in a more significant place than it (not including the
669 // sign bit) the ripple may go up to and fill the zero, but won't change the
670 // sign. For example, (X & ~4) + 1.
678 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
679 bool Changed = SimplifyCommutative(I);
680 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
682 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(),
683 I.hasNoUnsignedWrap(), TD))
684 return ReplaceInstUsesWith(I, V);
687 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
688 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
689 // X + (signbit) --> X ^ signbit
690 const APInt& Val = CI->getValue();
691 uint32_t BitWidth = Val.getBitWidth();
692 if (Val == APInt::getSignBit(BitWidth))
693 return BinaryOperator::CreateXor(LHS, RHS);
695 // See if SimplifyDemandedBits can simplify this. This handles stuff like
696 // (X & 254)+1 -> (X&254)|1
697 if (SimplifyDemandedInstructionBits(I))
700 // zext(bool) + C -> bool ? C + 1 : C
701 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
702 if (ZI->getSrcTy() == Type::getInt1Ty(I.getContext()))
703 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
706 if (isa<PHINode>(LHS))
707 if (Instruction *NV = FoldOpIntoPhi(I))
710 ConstantInt *XorRHS = 0;
712 if (isa<ConstantInt>(RHSC) &&
713 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
714 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
715 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
717 uint32_t Size = TySizeBits / 2;
718 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
719 APInt CFF80Val(-C0080Val);
721 if (TySizeBits > Size) {
722 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
723 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
724 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
725 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
726 // This is a sign extend if the top bits are known zero.
727 if (!MaskedValueIsZero(XorLHS,
728 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
729 Size = 0; // Not a sign ext, but can't be any others either.
734 C0080Val = APIntOps::lshr(C0080Val, Size);
735 CFF80Val = APIntOps::ashr(CFF80Val, Size);
738 // FIXME: This shouldn't be necessary. When the backends can handle types
739 // with funny bit widths then this switch statement should be removed. It
740 // is just here to get the size of the "middle" type back up to something
741 // that the back ends can handle.
742 const Type *MiddleType = 0;
747 case 8: MiddleType = IntegerType::get(I.getContext(), Size); break;
750 Value *NewTrunc = Builder->CreateTrunc(XorLHS, MiddleType, "sext");
751 return new SExtInst(NewTrunc, I.getType(), I.getName());
756 if (I.getType() == Type::getInt1Ty(I.getContext()))
757 return BinaryOperator::CreateXor(LHS, RHS);
760 if (I.getType()->isInteger()) {
761 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
764 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
765 if (RHSI->getOpcode() == Instruction::Sub)
766 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
767 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
769 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
770 if (LHSI->getOpcode() == Instruction::Sub)
771 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
772 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
777 // -A + -B --> -(A + B)
778 if (Value *LHSV = dyn_castNegVal(LHS)) {
779 if (LHS->getType()->isIntOrIntVector()) {
780 if (Value *RHSV = dyn_castNegVal(RHS)) {
781 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
782 return BinaryOperator::CreateNeg(NewAdd);
786 return BinaryOperator::CreateSub(RHS, LHSV);
790 if (!isa<Constant>(RHS))
791 if (Value *V = dyn_castNegVal(RHS))
792 return BinaryOperator::CreateSub(LHS, V);
796 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
797 if (X == RHS) // X*C + X --> X * (C+1)
798 return BinaryOperator::CreateMul(RHS, AddOne(C2));
800 // X*C1 + X*C2 --> X * (C1+C2)
802 if (X == dyn_castFoldableMul(RHS, C1))
803 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
806 // X + X*C --> X * (C+1)
807 if (dyn_castFoldableMul(RHS, C2) == LHS)
808 return BinaryOperator::CreateMul(LHS, AddOne(C2));
810 // X + ~X --> -1 since ~X = -X-1
811 if (dyn_castNotVal(LHS) == RHS ||
812 dyn_castNotVal(RHS) == LHS)
813 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
816 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
817 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
818 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
821 // A+B --> A|B iff A and B have no bits set in common.
822 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
823 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
824 APInt LHSKnownOne(IT->getBitWidth(), 0);
825 APInt LHSKnownZero(IT->getBitWidth(), 0);
826 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
827 if (LHSKnownZero != 0) {
828 APInt RHSKnownOne(IT->getBitWidth(), 0);
829 APInt RHSKnownZero(IT->getBitWidth(), 0);
830 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
832 // No bits in common -> bitwise or.
833 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
834 return BinaryOperator::CreateOr(LHS, RHS);
838 // W*X + Y*Z --> W * (X+Z) iff W == Y
839 if (I.getType()->isIntOrIntVector()) {
840 Value *W, *X, *Y, *Z;
841 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
842 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
855 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
856 return BinaryOperator::CreateMul(W, NewAdd);
861 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
863 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
864 return BinaryOperator::CreateSub(SubOne(CRHS), X);
866 // (X & FF00) + xx00 -> (X+xx00) & FF00
867 if (LHS->hasOneUse() &&
868 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
869 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
871 // See if all bits from the first bit set in the Add RHS up are included
872 // in the mask. First, get the rightmost bit.
873 const APInt& AddRHSV = CRHS->getValue();
875 // Form a mask of all bits from the lowest bit added through the top.
876 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
878 // See if the and mask includes all of these bits.
879 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
881 if (AddRHSHighBits == AddRHSHighBitsAnd) {
882 // Okay, the xform is safe. Insert the new add pronto.
883 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
884 return BinaryOperator::CreateAnd(NewAdd, C2);
889 // Try to fold constant add into select arguments.
890 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
891 if (Instruction *R = FoldOpIntoSelect(I, SI))
895 // add (select X 0 (sub n A)) A --> select X A n
897 SelectInst *SI = dyn_cast<SelectInst>(LHS);
900 SI = dyn_cast<SelectInst>(RHS);
903 if (SI && SI->hasOneUse()) {
904 Value *TV = SI->getTrueValue();
905 Value *FV = SI->getFalseValue();
908 // Can we fold the add into the argument of the select?
909 // We check both true and false select arguments for a matching subtract.
910 if (match(FV, m_Zero()) &&
911 match(TV, m_Sub(m_Value(N), m_Specific(A))))
912 // Fold the add into the true select value.
913 return SelectInst::Create(SI->getCondition(), N, A);
914 if (match(TV, m_Zero()) &&
915 match(FV, m_Sub(m_Value(N), m_Specific(A))))
916 // Fold the add into the false select value.
917 return SelectInst::Create(SI->getCondition(), A, N);
921 // Check for (add (sext x), y), see if we can merge this into an
922 // integer add followed by a sext.
923 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
924 // (add (sext x), cst) --> (sext (add x, cst'))
925 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
927 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
928 if (LHSConv->hasOneUse() &&
929 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
930 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
931 // Insert the new, smaller add.
932 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
934 return new SExtInst(NewAdd, I.getType());
938 // (add (sext x), (sext y)) --> (sext (add int x, y))
939 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
940 // Only do this if x/y have the same type, if at last one of them has a
941 // single use (so we don't increase the number of sexts), and if the
942 // integer add will not overflow.
943 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
944 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
945 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
946 RHSConv->getOperand(0))) {
947 // Insert the new integer add.
948 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
949 RHSConv->getOperand(0), "addconv");
950 return new SExtInst(NewAdd, I.getType());
955 return Changed ? &I : 0;
958 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
959 bool Changed = SimplifyCommutative(I);
960 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
962 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
964 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
965 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
966 (I.getType())->getValueAPF()))
967 return ReplaceInstUsesWith(I, LHS);
970 if (isa<PHINode>(LHS))
971 if (Instruction *NV = FoldOpIntoPhi(I))
976 // -A + -B --> -(A + B)
977 if (Value *LHSV = dyn_castFNegVal(LHS))
978 return BinaryOperator::CreateFSub(RHS, LHSV);
981 if (!isa<Constant>(RHS))
982 if (Value *V = dyn_castFNegVal(RHS))
983 return BinaryOperator::CreateFSub(LHS, V);
985 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
986 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
987 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
988 return ReplaceInstUsesWith(I, LHS);
990 // Check for (add double (sitofp x), y), see if we can merge this into an
991 // integer add followed by a promotion.
992 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
993 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
994 // ... if the constant fits in the integer value. This is useful for things
995 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
996 // requires a constant pool load, and generally allows the add to be better
998 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
1000 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
1001 if (LHSConv->hasOneUse() &&
1002 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
1003 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
1004 // Insert the new integer add.
1005 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1007 return new SIToFPInst(NewAdd, I.getType());
1011 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
1012 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
1013 // Only do this if x/y have the same type, if at last one of them has a
1014 // single use (so we don't increase the number of int->fp conversions),
1015 // and if the integer add will not overflow.
1016 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
1017 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
1018 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
1019 RHSConv->getOperand(0))) {
1020 // Insert the new integer add.
1021 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
1022 RHSConv->getOperand(0),"addconv");
1023 return new SIToFPInst(NewAdd, I.getType());
1028 return Changed ? &I : 0;
1032 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
1033 /// code necessary to compute the offset from the base pointer (without adding
1034 /// in the base pointer). Return the result as a signed integer of intptr size.
1035 Value *InstCombiner::EmitGEPOffset(User *GEP) {
1036 TargetData &TD = *getTargetData();
1037 gep_type_iterator GTI = gep_type_begin(GEP);
1038 const Type *IntPtrTy = TD.getIntPtrType(GEP->getContext());
1039 Value *Result = Constant::getNullValue(IntPtrTy);
1041 // Build a mask for high order bits.
1042 unsigned IntPtrWidth = TD.getPointerSizeInBits();
1043 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
1045 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
1048 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
1049 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
1050 if (OpC->isZero()) continue;
1052 // Handle a struct index, which adds its field offset to the pointer.
1053 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
1054 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
1056 Result = Builder->CreateAdd(Result,
1057 ConstantInt::get(IntPtrTy, Size),
1058 GEP->getName()+".offs");
1062 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
1064 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
1065 Scale = ConstantExpr::getMul(OC, Scale);
1066 // Emit an add instruction.
1067 Result = Builder->CreateAdd(Result, Scale, GEP->getName()+".offs");
1070 // Convert to correct type.
1071 if (Op->getType() != IntPtrTy)
1072 Op = Builder->CreateIntCast(Op, IntPtrTy, true, Op->getName()+".c");
1074 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
1075 // We'll let instcombine(mul) convert this to a shl if possible.
1076 Op = Builder->CreateMul(Op, Scale, GEP->getName()+".idx");
1079 // Emit an add instruction.
1080 Result = Builder->CreateAdd(Op, Result, GEP->getName()+".offs");
1088 /// Optimize pointer differences into the same array into a size. Consider:
1089 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
1090 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1092 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
1094 assert(TD && "Must have target data info for this");
1096 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
1098 bool Swapped = false;
1099 GetElementPtrInst *GEP = 0;
1100 ConstantExpr *CstGEP = 0;
1102 // TODO: Could also optimize &A[i] - &A[j] -> "i-j", and "&A.foo[i] - &A.foo".
1103 // For now we require one side to be the base pointer "A" or a constant
1104 // expression derived from it.
1105 if (GetElementPtrInst *LHSGEP = dyn_cast<GetElementPtrInst>(LHS)) {
1107 if (LHSGEP->getOperand(0) == RHS) {
1110 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(RHS)) {
1111 // (gep X, ...) - (ce_gep X, ...)
1112 if (CE->getOpcode() == Instruction::GetElementPtr &&
1113 LHSGEP->getOperand(0) == CE->getOperand(0)) {
1121 if (GetElementPtrInst *RHSGEP = dyn_cast<GetElementPtrInst>(RHS)) {
1123 if (RHSGEP->getOperand(0) == LHS) {
1126 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(LHS)) {
1127 // (ce_gep X, ...) - (gep X, ...)
1128 if (CE->getOpcode() == Instruction::GetElementPtr &&
1129 RHSGEP->getOperand(0) == CE->getOperand(0)) {
1140 // Emit the offset of the GEP and an intptr_t.
1141 Value *Result = EmitGEPOffset(GEP);
1143 // If we had a constant expression GEP on the other side offsetting the
1144 // pointer, subtract it from the offset we have.
1146 Value *CstOffset = EmitGEPOffset(CstGEP);
1147 Result = Builder->CreateSub(Result, CstOffset);
1151 // If we have p - gep(p, ...) then we have to negate the result.
1153 Result = Builder->CreateNeg(Result, "diff.neg");
1155 return Builder->CreateIntCast(Result, Ty, true);
1159 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1160 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1162 if (Op0 == Op1) // sub X, X -> 0
1163 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1165 // If this is a 'B = x-(-A)', change to B = x+A. This preserves NSW/NUW.
1166 if (Value *V = dyn_castNegVal(Op1)) {
1167 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
1168 Res->setHasNoSignedWrap(I.hasNoSignedWrap());
1169 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
1173 if (isa<UndefValue>(Op0))
1174 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1175 if (isa<UndefValue>(Op1))
1176 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1177 if (I.getType() == Type::getInt1Ty(I.getContext()))
1178 return BinaryOperator::CreateXor(Op0, Op1);
1180 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1181 // Replace (-1 - A) with (~A).
1182 if (C->isAllOnesValue())
1183 return BinaryOperator::CreateNot(Op1);
1185 // C - ~X == X + (1+C)
1187 if (match(Op1, m_Not(m_Value(X))))
1188 return BinaryOperator::CreateAdd(X, AddOne(C));
1190 // -(X >>u 31) -> (X >>s 31)
1191 // -(X >>s 31) -> (X >>u 31)
1193 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
1194 if (SI->getOpcode() == Instruction::LShr) {
1195 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1196 // Check to see if we are shifting out everything but the sign bit.
1197 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
1198 SI->getType()->getPrimitiveSizeInBits()-1) {
1199 // Ok, the transformation is safe. Insert AShr.
1200 return BinaryOperator::Create(Instruction::AShr,
1201 SI->getOperand(0), CU, SI->getName());
1204 } else if (SI->getOpcode() == Instruction::AShr) {
1205 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1206 // Check to see if we are shifting out everything but the sign bit.
1207 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
1208 SI->getType()->getPrimitiveSizeInBits()-1) {
1209 // Ok, the transformation is safe. Insert LShr.
1210 return BinaryOperator::CreateLShr(
1211 SI->getOperand(0), CU, SI->getName());
1218 // Try to fold constant sub into select arguments.
1219 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1220 if (Instruction *R = FoldOpIntoSelect(I, SI))
1223 // C - zext(bool) -> bool ? C - 1 : C
1224 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
1225 if (ZI->getSrcTy() == Type::getInt1Ty(I.getContext()))
1226 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
1229 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1230 if (Op1I->getOpcode() == Instruction::Add) {
1231 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1232 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
1234 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1235 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
1237 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
1238 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
1239 // C1-(X+C2) --> (C1-C2)-X
1240 return BinaryOperator::CreateSub(
1241 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
1245 if (Op1I->hasOneUse()) {
1246 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
1247 // is not used by anyone else...
1249 if (Op1I->getOpcode() == Instruction::Sub) {
1250 // Swap the two operands of the subexpr...
1251 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
1252 Op1I->setOperand(0, IIOp1);
1253 Op1I->setOperand(1, IIOp0);
1255 // Create the new top level add instruction...
1256 return BinaryOperator::CreateAdd(Op0, Op1);
1259 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
1261 if (Op1I->getOpcode() == Instruction::And &&
1262 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
1263 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
1265 Value *NewNot = Builder->CreateNot(OtherOp, "B.not");
1266 return BinaryOperator::CreateAnd(Op0, NewNot);
1269 // 0 - (X sdiv C) -> (X sdiv -C)
1270 if (Op1I->getOpcode() == Instruction::SDiv)
1271 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
1273 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
1274 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
1275 ConstantExpr::getNeg(DivRHS));
1277 // X - X*C --> X * (1-C)
1278 ConstantInt *C2 = 0;
1279 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
1281 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
1283 return BinaryOperator::CreateMul(Op0, CP1);
1288 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1289 if (Op0I->getOpcode() == Instruction::Add) {
1290 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
1291 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
1292 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
1293 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
1294 } else if (Op0I->getOpcode() == Instruction::Sub) {
1295 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
1296 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
1302 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
1303 if (X == Op1) // X*C - X --> X * (C-1)
1304 return BinaryOperator::CreateMul(Op1, SubOne(C1));
1306 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
1307 if (X == dyn_castFoldableMul(Op1, C2))
1308 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
1311 // Optimize pointer differences into the same array into a size. Consider:
1312 // &A[10] - &A[0]: we should compile this to "10".
1314 Value *LHSOp, *RHSOp;
1315 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
1316 match(Op1, m_PtrToInt(m_Value(RHSOp))))
1317 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1318 return ReplaceInstUsesWith(I, Res);
1320 // trunc(p)-trunc(q) -> trunc(p-q)
1321 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
1322 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
1323 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
1324 return ReplaceInstUsesWith(I, Res);
1330 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
1331 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1333 // If this is a 'B = x-(-A)', change to B = x+A...
1334 if (Value *V = dyn_castFNegVal(Op1))
1335 return BinaryOperator::CreateFAdd(Op0, V);
1337 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1338 if (Op1I->getOpcode() == Instruction::FAdd) {
1339 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1340 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
1342 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1343 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
1351 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
1352 bool Changed = SimplifyCommutative(I);
1353 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1355 if (isa<UndefValue>(Op1)) // undef * X -> 0
1356 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1358 // Simplify mul instructions with a constant RHS.
1359 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1360 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1C)) {
1362 // ((X << C1)*C2) == (X * (C2 << C1))
1363 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
1364 if (SI->getOpcode() == Instruction::Shl)
1365 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
1366 return BinaryOperator::CreateMul(SI->getOperand(0),
1367 ConstantExpr::getShl(CI, ShOp));
1370 return ReplaceInstUsesWith(I, Op1C); // X * 0 == 0
1371 if (CI->equalsInt(1)) // X * 1 == X
1372 return ReplaceInstUsesWith(I, Op0);
1373 if (CI->isAllOnesValue()) // X * -1 == 0 - X
1374 return BinaryOperator::CreateNeg(Op0, I.getName());
1376 const APInt& Val = cast<ConstantInt>(CI)->getValue();
1377 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
1378 return BinaryOperator::CreateShl(Op0,
1379 ConstantInt::get(Op0->getType(), Val.logBase2()));
1381 } else if (isa<VectorType>(Op1C->getType())) {
1382 if (Op1C->isNullValue())
1383 return ReplaceInstUsesWith(I, Op1C);
1385 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
1386 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
1387 return BinaryOperator::CreateNeg(Op0, I.getName());
1389 // As above, vector X*splat(1.0) -> X in all defined cases.
1390 if (Constant *Splat = Op1V->getSplatValue()) {
1391 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
1392 if (CI->equalsInt(1))
1393 return ReplaceInstUsesWith(I, Op0);
1398 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
1399 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
1400 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1C)) {
1401 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
1402 Value *Add = Builder->CreateMul(Op0I->getOperand(0), Op1C, "tmp");
1403 Value *C1C2 = Builder->CreateMul(Op1C, Op0I->getOperand(1));
1404 return BinaryOperator::CreateAdd(Add, C1C2);
1408 // Try to fold constant mul into select arguments.
1409 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1410 if (Instruction *R = FoldOpIntoSelect(I, SI))
1413 if (isa<PHINode>(Op0))
1414 if (Instruction *NV = FoldOpIntoPhi(I))
1418 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
1419 if (Value *Op1v = dyn_castNegVal(Op1))
1420 return BinaryOperator::CreateMul(Op0v, Op1v);
1422 // (X / Y) * Y = X - (X % Y)
1423 // (X / Y) * -Y = (X % Y) - X
1426 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
1428 (BO->getOpcode() != Instruction::UDiv &&
1429 BO->getOpcode() != Instruction::SDiv)) {
1431 BO = dyn_cast<BinaryOperator>(Op1);
1433 Value *Neg = dyn_castNegVal(Op1C);
1434 if (BO && BO->hasOneUse() &&
1435 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) &&
1436 (BO->getOpcode() == Instruction::UDiv ||
1437 BO->getOpcode() == Instruction::SDiv)) {
1438 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
1440 // If the division is exact, X % Y is zero.
1441 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
1442 if (SDiv->isExact()) {
1444 return ReplaceInstUsesWith(I, Op0BO);
1445 return BinaryOperator::CreateNeg(Op0BO);
1449 if (BO->getOpcode() == Instruction::UDiv)
1450 Rem = Builder->CreateURem(Op0BO, Op1BO);
1452 Rem = Builder->CreateSRem(Op0BO, Op1BO);
1456 return BinaryOperator::CreateSub(Op0BO, Rem);
1457 return BinaryOperator::CreateSub(Rem, Op0BO);
1461 /// i1 mul -> i1 and.
1462 if (I.getType() == Type::getInt1Ty(I.getContext()))
1463 return BinaryOperator::CreateAnd(Op0, Op1);
1465 // X*(1 << Y) --> X << Y
1466 // (1 << Y)*X --> X << Y
1469 if (match(Op0, m_Shl(m_One(), m_Value(Y))))
1470 return BinaryOperator::CreateShl(Op1, Y);
1471 if (match(Op1, m_Shl(m_One(), m_Value(Y))))
1472 return BinaryOperator::CreateShl(Op0, Y);
1475 // If one of the operands of the multiply is a cast from a boolean value, then
1476 // we know the bool is either zero or one, so this is a 'masking' multiply.
1477 // X * Y (where Y is 0 or 1) -> X & (0-Y)
1478 if (!isa<VectorType>(I.getType())) {
1479 // -2 is "-1 << 1" so it is all bits set except the low one.
1480 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
1482 Value *BoolCast = 0, *OtherOp = 0;
1483 if (MaskedValueIsZero(Op0, Negative2))
1484 BoolCast = Op0, OtherOp = Op1;
1485 else if (MaskedValueIsZero(Op1, Negative2))
1486 BoolCast = Op1, OtherOp = Op0;
1489 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
1491 return BinaryOperator::CreateAnd(V, OtherOp);
1495 return Changed ? &I : 0;
1498 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
1499 bool Changed = SimplifyCommutative(I);
1500 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1502 // Simplify mul instructions with a constant RHS...
1503 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1504 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1C)) {
1505 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
1506 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
1507 if (Op1F->isExactlyValue(1.0))
1508 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
1509 } else if (isa<VectorType>(Op1C->getType())) {
1510 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
1511 // As above, vector X*splat(1.0) -> X in all defined cases.
1512 if (Constant *Splat = Op1V->getSplatValue()) {
1513 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
1514 if (F->isExactlyValue(1.0))
1515 return ReplaceInstUsesWith(I, Op0);
1520 // Try to fold constant mul into select arguments.
1521 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1522 if (Instruction *R = FoldOpIntoSelect(I, SI))
1525 if (isa<PHINode>(Op0))
1526 if (Instruction *NV = FoldOpIntoPhi(I))
1530 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
1531 if (Value *Op1v = dyn_castFNegVal(Op1))
1532 return BinaryOperator::CreateFMul(Op0v, Op1v);
1534 return Changed ? &I : 0;
1537 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
1539 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
1540 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
1542 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
1543 int NonNullOperand = -1;
1544 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
1545 if (ST->isNullValue())
1547 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
1548 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
1549 if (ST->isNullValue())
1552 if (NonNullOperand == -1)
1555 Value *SelectCond = SI->getOperand(0);
1557 // Change the div/rem to use 'Y' instead of the select.
1558 I.setOperand(1, SI->getOperand(NonNullOperand));
1560 // Okay, we know we replace the operand of the div/rem with 'Y' with no
1561 // problem. However, the select, or the condition of the select may have
1562 // multiple uses. Based on our knowledge that the operand must be non-zero,
1563 // propagate the known value for the select into other uses of it, and
1564 // propagate a known value of the condition into its other users.
1566 // If the select and condition only have a single use, don't bother with this,
1568 if (SI->use_empty() && SelectCond->hasOneUse())
1571 // Scan the current block backward, looking for other uses of SI.
1572 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
1574 while (BBI != BBFront) {
1576 // If we found a call to a function, we can't assume it will return, so
1577 // information from below it cannot be propagated above it.
1578 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
1581 // Replace uses of the select or its condition with the known values.
1582 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
1585 *I = SI->getOperand(NonNullOperand);
1587 } else if (*I == SelectCond) {
1588 *I = NonNullOperand == 1 ? ConstantInt::getTrue(BBI->getContext()) :
1589 ConstantInt::getFalse(BBI->getContext());
1594 // If we past the instruction, quit looking for it.
1597 if (&*BBI == SelectCond)
1600 // If we ran out of things to eliminate, break out of the loop.
1601 if (SelectCond == 0 && SI == 0)
1609 /// This function implements the transforms on div instructions that work
1610 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
1611 /// used by the visitors to those instructions.
1612 /// @brief Transforms common to all three div instructions
1613 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
1614 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1616 // undef / X -> 0 for integer.
1617 // undef / X -> undef for FP (the undef could be a snan).
1618 if (isa<UndefValue>(Op0)) {
1619 if (Op0->getType()->isFPOrFPVector())
1620 return ReplaceInstUsesWith(I, Op0);
1621 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1624 // X / undef -> undef
1625 if (isa<UndefValue>(Op1))
1626 return ReplaceInstUsesWith(I, Op1);
1631 /// This function implements the transforms common to both integer division
1632 /// instructions (udiv and sdiv). It is called by the visitors to those integer
1633 /// division instructions.
1634 /// @brief Common integer divide transforms
1635 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
1636 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1638 // (sdiv X, X) --> 1 (udiv X, X) --> 1
1640 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
1641 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
1642 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
1643 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
1646 Constant *CI = ConstantInt::get(I.getType(), 1);
1647 return ReplaceInstUsesWith(I, CI);
1650 if (Instruction *Common = commonDivTransforms(I))
1653 // Handle cases involving: [su]div X, (select Cond, Y, Z)
1654 // This does not apply for fdiv.
1655 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
1658 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1660 if (RHS->equalsInt(1))
1661 return ReplaceInstUsesWith(I, Op0);
1663 // (X / C1) / C2 -> X / (C1*C2)
1664 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
1665 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
1666 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
1667 if (MultiplyOverflows(RHS, LHSRHS,
1668 I.getOpcode()==Instruction::SDiv))
1669 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1671 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
1672 ConstantExpr::getMul(RHS, LHSRHS));
1675 if (!RHS->isZero()) { // avoid X udiv 0
1676 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1677 if (Instruction *R = FoldOpIntoSelect(I, SI))
1679 if (isa<PHINode>(Op0))
1680 if (Instruction *NV = FoldOpIntoPhi(I))
1685 // 0 / X == 0, we don't need to preserve faults!
1686 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
1687 if (LHS->equalsInt(0))
1688 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1690 // It can't be division by zero, hence it must be division by one.
1691 if (I.getType() == Type::getInt1Ty(I.getContext()))
1692 return ReplaceInstUsesWith(I, Op0);
1694 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
1695 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
1698 return ReplaceInstUsesWith(I, Op0);
1704 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
1705 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1707 // Handle the integer div common cases
1708 if (Instruction *Common = commonIDivTransforms(I))
1711 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
1712 // X udiv C^2 -> X >> C
1713 // Check to see if this is an unsigned division with an exact power of 2,
1714 // if so, convert to a right shift.
1715 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
1716 return BinaryOperator::CreateLShr(Op0,
1717 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
1719 // X udiv C, where C >= signbit
1720 if (C->getValue().isNegative()) {
1721 Value *IC = Builder->CreateICmpULT( Op0, C);
1722 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
1723 ConstantInt::get(I.getType(), 1));
1727 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
1728 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
1729 if (RHSI->getOpcode() == Instruction::Shl &&
1730 isa<ConstantInt>(RHSI->getOperand(0))) {
1731 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
1732 if (C1.isPowerOf2()) {
1733 Value *N = RHSI->getOperand(1);
1734 const Type *NTy = N->getType();
1735 if (uint32_t C2 = C1.logBase2())
1736 N = Builder->CreateAdd(N, ConstantInt::get(NTy, C2), "tmp");
1737 return BinaryOperator::CreateLShr(Op0, N);
1742 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
1743 // where C1&C2 are powers of two.
1744 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1745 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
1746 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
1747 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
1748 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
1749 // Compute the shift amounts
1750 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
1751 // Construct the "on true" case of the select
1752 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
1753 Value *TSI = Builder->CreateLShr(Op0, TC, SI->getName()+".t");
1755 // Construct the "on false" case of the select
1756 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
1757 Value *FSI = Builder->CreateLShr(Op0, FC, SI->getName()+".f");
1759 // construct the select instruction and return it.
1760 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
1766 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
1767 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1769 // Handle the integer div common cases
1770 if (Instruction *Common = commonIDivTransforms(I))
1773 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1775 if (RHS->isAllOnesValue())
1776 return BinaryOperator::CreateNeg(Op0);
1778 // sdiv X, C --> ashr X, log2(C)
1779 if (cast<SDivOperator>(&I)->isExact() &&
1780 RHS->getValue().isNonNegative() &&
1781 RHS->getValue().isPowerOf2()) {
1782 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
1783 RHS->getValue().exactLogBase2());
1784 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
1787 // -X/C --> X/-C provided the negation doesn't overflow.
1788 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
1789 if (isa<Constant>(Sub->getOperand(0)) &&
1790 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
1791 Sub->hasNoSignedWrap())
1792 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
1793 ConstantExpr::getNeg(RHS));
1796 // If the sign bits of both operands are zero (i.e. we can prove they are
1797 // unsigned inputs), turn this into a udiv.
1798 if (I.getType()->isInteger()) {
1799 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
1800 if (MaskedValueIsZero(Op0, Mask)) {
1801 if (MaskedValueIsZero(Op1, Mask)) {
1802 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
1803 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1805 ConstantInt *ShiftedInt;
1806 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
1807 ShiftedInt->getValue().isPowerOf2()) {
1808 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
1809 // Safe because the only negative value (1 << Y) can take on is
1810 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
1811 // the sign bit set.
1812 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1820 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
1821 return commonDivTransforms(I);
1824 /// This function implements the transforms on rem instructions that work
1825 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
1826 /// is used by the visitors to those instructions.
1827 /// @brief Transforms common to all three rem instructions
1828 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
1829 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1831 if (isa<UndefValue>(Op0)) { // undef % X -> 0
1832 if (I.getType()->isFPOrFPVector())
1833 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
1834 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1836 if (isa<UndefValue>(Op1))
1837 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
1839 // Handle cases involving: rem X, (select Cond, Y, Z)
1840 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
1846 /// This function implements the transforms common to both integer remainder
1847 /// instructions (urem and srem). It is called by the visitors to those integer
1848 /// remainder instructions.
1849 /// @brief Common integer remainder transforms
1850 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
1851 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1853 if (Instruction *common = commonRemTransforms(I))
1856 // 0 % X == 0 for integer, we don't need to preserve faults!
1857 if (Constant *LHS = dyn_cast<Constant>(Op0))
1858 if (LHS->isNullValue())
1859 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1861 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1862 // X % 0 == undef, we don't need to preserve faults!
1863 if (RHS->equalsInt(0))
1864 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
1866 if (RHS->equalsInt(1)) // X % 1 == 0
1867 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1869 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
1870 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
1871 if (Instruction *R = FoldOpIntoSelect(I, SI))
1873 } else if (isa<PHINode>(Op0I)) {
1874 if (Instruction *NV = FoldOpIntoPhi(I))
1878 // See if we can fold away this rem instruction.
1879 if (SimplifyDemandedInstructionBits(I))
1887 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
1888 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1890 if (Instruction *common = commonIRemTransforms(I))
1893 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1894 // X urem C^2 -> X and C
1895 // Check to see if this is an unsigned remainder with an exact power of 2,
1896 // if so, convert to a bitwise and.
1897 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
1898 if (C->getValue().isPowerOf2())
1899 return BinaryOperator::CreateAnd(Op0, SubOne(C));
1902 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
1903 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
1904 if (RHSI->getOpcode() == Instruction::Shl &&
1905 isa<ConstantInt>(RHSI->getOperand(0))) {
1906 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
1907 Constant *N1 = Constant::getAllOnesValue(I.getType());
1908 Value *Add = Builder->CreateAdd(RHSI, N1, "tmp");
1909 return BinaryOperator::CreateAnd(Op0, Add);
1914 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
1915 // where C1&C2 are powers of two.
1916 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
1917 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
1918 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
1919 // STO == 0 and SFO == 0 handled above.
1920 if ((STO->getValue().isPowerOf2()) &&
1921 (SFO->getValue().isPowerOf2())) {
1922 Value *TrueAnd = Builder->CreateAnd(Op0, SubOne(STO),
1923 SI->getName()+".t");
1924 Value *FalseAnd = Builder->CreateAnd(Op0, SubOne(SFO),
1925 SI->getName()+".f");
1926 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
1934 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
1935 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1937 // Handle the integer rem common cases
1938 if (Instruction *Common = commonIRemTransforms(I))
1941 if (Value *RHSNeg = dyn_castNegVal(Op1))
1942 if (!isa<Constant>(RHSNeg) ||
1943 (isa<ConstantInt>(RHSNeg) &&
1944 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
1946 Worklist.AddValue(I.getOperand(1));
1947 I.setOperand(1, RHSNeg);
1951 // If the sign bits of both operands are zero (i.e. we can prove they are
1952 // unsigned inputs), turn this into a urem.
1953 if (I.getType()->isInteger()) {
1954 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
1955 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
1956 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
1957 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
1961 // If it's a constant vector, flip any negative values positive.
1962 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
1963 unsigned VWidth = RHSV->getNumOperands();
1965 bool hasNegative = false;
1966 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
1967 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
1968 if (RHS->getValue().isNegative())
1972 std::vector<Constant *> Elts(VWidth);
1973 for (unsigned i = 0; i != VWidth; ++i) {
1974 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
1975 if (RHS->getValue().isNegative())
1976 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
1982 Constant *NewRHSV = ConstantVector::get(Elts);
1983 if (NewRHSV != RHSV) {
1984 Worklist.AddValue(I.getOperand(1));
1985 I.setOperand(1, NewRHSV);
1994 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
1995 return commonRemTransforms(I);
1998 // isOneBitSet - Return true if there is exactly one bit set in the specified
2000 static bool isOneBitSet(const ConstantInt *CI) {
2001 return CI->getValue().isPowerOf2();
2004 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2005 /// are carefully arranged to allow folding of expressions such as:
2007 /// (A < B) | (A > B) --> (A != B)
2009 /// Note that this is only valid if the first and second predicates have the
2010 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2012 /// Three bits are used to represent the condition, as follows:
2017 /// <=> Value Definition
2018 /// 000 0 Always false
2025 /// 111 7 Always true
2027 static unsigned getICmpCode(const ICmpInst *ICI) {
2028 switch (ICI->getPredicate()) {
2030 case ICmpInst::ICMP_UGT: return 1; // 001
2031 case ICmpInst::ICMP_SGT: return 1; // 001
2032 case ICmpInst::ICMP_EQ: return 2; // 010
2033 case ICmpInst::ICMP_UGE: return 3; // 011
2034 case ICmpInst::ICMP_SGE: return 3; // 011
2035 case ICmpInst::ICMP_ULT: return 4; // 100
2036 case ICmpInst::ICMP_SLT: return 4; // 100
2037 case ICmpInst::ICMP_NE: return 5; // 101
2038 case ICmpInst::ICMP_ULE: return 6; // 110
2039 case ICmpInst::ICMP_SLE: return 6; // 110
2042 llvm_unreachable("Invalid ICmp predicate!");
2047 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
2048 /// predicate into a three bit mask. It also returns whether it is an ordered
2049 /// predicate by reference.
2050 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
2053 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
2054 case FCmpInst::FCMP_UNO: return 0; // 000
2055 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
2056 case FCmpInst::FCMP_UGT: return 1; // 001
2057 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
2058 case FCmpInst::FCMP_UEQ: return 2; // 010
2059 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
2060 case FCmpInst::FCMP_UGE: return 3; // 011
2061 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
2062 case FCmpInst::FCMP_ULT: return 4; // 100
2063 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
2064 case FCmpInst::FCMP_UNE: return 5; // 101
2065 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
2066 case FCmpInst::FCMP_ULE: return 6; // 110
2069 // Not expecting FCMP_FALSE and FCMP_TRUE;
2070 llvm_unreachable("Unexpected FCmp predicate!");
2075 /// getICmpValue - This is the complement of getICmpCode, which turns an
2076 /// opcode and two operands into either a constant true or false, or a brand
2077 /// new ICmp instruction. The sign is passed in to determine which kind
2078 /// of predicate to use in the new icmp instruction.
2079 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2081 default: llvm_unreachable("Illegal ICmp code!");
2082 case 0: return ConstantInt::getFalse(LHS->getContext());
2085 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2087 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2088 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2091 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2093 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2096 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2098 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2099 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2102 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2104 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2105 case 7: return ConstantInt::getTrue(LHS->getContext());
2109 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
2110 /// opcode and two operands into either a FCmp instruction. isordered is passed
2111 /// in to determine which kind of predicate to use in the new fcmp instruction.
2112 static Value *getFCmpValue(bool isordered, unsigned code,
2113 Value *LHS, Value *RHS) {
2115 default: llvm_unreachable("Illegal FCmp code!");
2118 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
2120 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
2123 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
2125 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
2128 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
2130 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
2133 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
2135 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
2138 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
2140 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
2143 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
2145 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
2148 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
2150 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
2151 case 7: return ConstantInt::getTrue(LHS->getContext());
2155 /// PredicatesFoldable - Return true if both predicates match sign or if at
2156 /// least one of them is an equality comparison (which is signless).
2157 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2158 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
2159 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
2160 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
2164 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2165 struct FoldICmpLogical {
2168 ICmpInst::Predicate pred;
2169 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2170 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2171 pred(ICI->getPredicate()) {}
2172 bool shouldApply(Value *V) const {
2173 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2174 if (PredicatesFoldable(pred, ICI->getPredicate()))
2175 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
2176 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
2179 Instruction *apply(Instruction &Log) const {
2180 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2181 if (ICI->getOperand(0) != LHS) {
2182 assert(ICI->getOperand(1) == LHS);
2183 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2186 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2187 unsigned LHSCode = getICmpCode(ICI);
2188 unsigned RHSCode = getICmpCode(RHSICI);
2190 switch (Log.getOpcode()) {
2191 case Instruction::And: Code = LHSCode & RHSCode; break;
2192 case Instruction::Or: Code = LHSCode | RHSCode; break;
2193 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2194 default: llvm_unreachable("Illegal logical opcode!"); return 0;
2197 bool isSigned = RHSICI->isSigned() || ICI->isSigned();
2198 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2199 if (Instruction *I = dyn_cast<Instruction>(RV))
2201 // Otherwise, it's a constant boolean value...
2202 return IC.ReplaceInstUsesWith(Log, RV);
2205 } // end anonymous namespace
2207 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2208 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2209 // guaranteed to be a binary operator.
2210 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2212 ConstantInt *AndRHS,
2213 BinaryOperator &TheAnd) {
2214 Value *X = Op->getOperand(0);
2215 Constant *Together = 0;
2217 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2219 switch (Op->getOpcode()) {
2220 case Instruction::Xor:
2221 if (Op->hasOneUse()) {
2222 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2223 Value *And = Builder->CreateAnd(X, AndRHS);
2225 return BinaryOperator::CreateXor(And, Together);
2228 case Instruction::Or:
2229 if (Together == AndRHS) // (X | C) & C --> C
2230 return ReplaceInstUsesWith(TheAnd, AndRHS);
2232 if (Op->hasOneUse() && Together != OpRHS) {
2233 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2234 Value *Or = Builder->CreateOr(X, Together);
2236 return BinaryOperator::CreateAnd(Or, AndRHS);
2239 case Instruction::Add:
2240 if (Op->hasOneUse()) {
2241 // Adding a one to a single bit bit-field should be turned into an XOR
2242 // of the bit. First thing to check is to see if this AND is with a
2243 // single bit constant.
2244 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
2246 // If there is only one bit set...
2247 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2248 // Ok, at this point, we know that we are masking the result of the
2249 // ADD down to exactly one bit. If the constant we are adding has
2250 // no bits set below this bit, then we can eliminate the ADD.
2251 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
2253 // Check to see if any bits below the one bit set in AndRHSV are set.
2254 if ((AddRHS & (AndRHSV-1)) == 0) {
2255 // If not, the only thing that can effect the output of the AND is
2256 // the bit specified by AndRHSV. If that bit is set, the effect of
2257 // the XOR is to toggle the bit. If it is clear, then the ADD has
2259 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2260 TheAnd.setOperand(0, X);
2263 // Pull the XOR out of the AND.
2264 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
2265 NewAnd->takeName(Op);
2266 return BinaryOperator::CreateXor(NewAnd, AndRHS);
2273 case Instruction::Shl: {
2274 // We know that the AND will not produce any of the bits shifted in, so if
2275 // the anded constant includes them, clear them now!
2277 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
2278 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
2279 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
2280 ConstantInt *CI = ConstantInt::get(AndRHS->getContext(),
2281 AndRHS->getValue() & ShlMask);
2283 if (CI->getValue() == ShlMask) {
2284 // Masking out bits that the shift already masks
2285 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2286 } else if (CI != AndRHS) { // Reducing bits set in and.
2287 TheAnd.setOperand(1, CI);
2292 case Instruction::LShr: {
2293 // We know that the AND will not produce any of the bits shifted in, so if
2294 // the anded constant includes them, clear them now! This only applies to
2295 // unsigned shifts, because a signed shr may bring in set bits!
2297 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
2298 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
2299 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
2300 ConstantInt *CI = ConstantInt::get(Op->getContext(),
2301 AndRHS->getValue() & ShrMask);
2303 if (CI->getValue() == ShrMask) {
2304 // Masking out bits that the shift already masks.
2305 return ReplaceInstUsesWith(TheAnd, Op);
2306 } else if (CI != AndRHS) {
2307 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2312 case Instruction::AShr:
2314 // See if this is shifting in some sign extension, then masking it out
2316 if (Op->hasOneUse()) {
2317 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
2318 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
2319 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
2320 Constant *C = ConstantInt::get(Op->getContext(),
2321 AndRHS->getValue() & ShrMask);
2322 if (C == AndRHS) { // Masking out bits shifted in.
2323 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2324 // Make the argument unsigned.
2325 Value *ShVal = Op->getOperand(0);
2326 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
2327 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
2336 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2337 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2338 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2339 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2340 /// insert new instructions.
2341 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2342 bool isSigned, bool Inside,
2344 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
2345 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
2346 "Lo is not <= Hi in range emission code!");
2349 if (Lo == Hi) // Trivially false.
2350 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
2352 // V >= Min && V < Hi --> V < Hi
2353 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2354 ICmpInst::Predicate pred = (isSigned ?
2355 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
2356 return new ICmpInst(pred, V, Hi);
2359 // Emit V-Lo <u Hi-Lo
2360 Constant *NegLo = ConstantExpr::getNeg(Lo);
2361 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
2362 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
2363 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
2366 if (Lo == Hi) // Trivially true.
2367 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
2369 // V < Min || V >= Hi -> V > Hi-1
2370 Hi = SubOne(cast<ConstantInt>(Hi));
2371 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2372 ICmpInst::Predicate pred = (isSigned ?
2373 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
2374 return new ICmpInst(pred, V, Hi);
2377 // Emit V-Lo >u Hi-1-Lo
2378 // Note that Hi has already had one subtracted from it, above.
2379 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
2380 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
2381 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
2382 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
2385 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2386 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2387 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2388 // not, since all 1s are not contiguous.
2389 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
2390 const APInt& V = Val->getValue();
2391 uint32_t BitWidth = Val->getType()->getBitWidth();
2392 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
2394 // look for the first zero bit after the run of ones
2395 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
2396 // look for the first non-zero bit
2397 ME = V.getActiveBits();
2401 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2402 /// where isSub determines whether the operator is a sub. If we can fold one of
2403 /// the following xforms:
2405 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
2406 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2407 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2409 /// return (A +/- B).
2411 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
2412 ConstantInt *Mask, bool isSub,
2414 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2415 if (!LHSI || LHSI->getNumOperands() != 2 ||
2416 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
2418 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
2420 switch (LHSI->getOpcode()) {
2422 case Instruction::And:
2423 if (ConstantExpr::getAnd(N, Mask) == Mask) {
2424 // If the AndRHS is a power of two minus one (0+1+), this is simple.
2425 if ((Mask->getValue().countLeadingZeros() +
2426 Mask->getValue().countPopulation()) ==
2427 Mask->getValue().getBitWidth())
2430 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
2431 // part, we don't need any explicit masks to take them out of A. If that
2432 // is all N is, ignore it.
2433 uint32_t MB = 0, ME = 0;
2434 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
2435 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
2436 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
2437 if (MaskedValueIsZero(RHS, Mask))
2442 case Instruction::Or:
2443 case Instruction::Xor:
2444 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
2445 if ((Mask->getValue().countLeadingZeros() +
2446 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
2447 && ConstantExpr::getAnd(N, Mask)->isNullValue())
2453 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
2454 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
2457 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
2458 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
2459 ICmpInst *LHS, ICmpInst *RHS) {
2461 ConstantInt *LHSCst, *RHSCst;
2462 ICmpInst::Predicate LHSCC, RHSCC;
2464 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
2465 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
2466 m_ConstantInt(LHSCst))) ||
2467 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
2468 m_ConstantInt(RHSCst))))
2471 if (LHSCst == RHSCst && LHSCC == RHSCC) {
2472 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
2473 // where C is a power of 2
2474 if (LHSCC == ICmpInst::ICMP_ULT &&
2475 LHSCst->getValue().isPowerOf2()) {
2476 Value *NewOr = Builder->CreateOr(Val, Val2);
2477 return new ICmpInst(LHSCC, NewOr, LHSCst);
2480 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
2481 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
2482 Value *NewOr = Builder->CreateOr(Val, Val2);
2483 return new ICmpInst(LHSCC, NewOr, LHSCst);
2487 // From here on, we only handle:
2488 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
2489 if (Val != Val2) return 0;
2491 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
2492 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
2493 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
2494 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
2495 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
2498 // We can't fold (ugt x, C) & (sgt x, C2).
2499 if (!PredicatesFoldable(LHSCC, RHSCC))
2502 // Ensure that the larger constant is on the RHS.
2504 if (CmpInst::isSigned(LHSCC) ||
2505 (ICmpInst::isEquality(LHSCC) &&
2506 CmpInst::isSigned(RHSCC)))
2507 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
2509 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
2512 std::swap(LHS, RHS);
2513 std::swap(LHSCst, RHSCst);
2514 std::swap(LHSCC, RHSCC);
2517 // At this point, we know we have have two icmp instructions
2518 // comparing a value against two constants and and'ing the result
2519 // together. Because of the above check, we know that we only have
2520 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
2521 // (from the FoldICmpLogical check above), that the two constants
2522 // are not equal and that the larger constant is on the RHS
2523 assert(LHSCst != RHSCst && "Compares not folded above?");
2526 default: llvm_unreachable("Unknown integer condition code!");
2527 case ICmpInst::ICMP_EQ:
2529 default: llvm_unreachable("Unknown integer condition code!");
2530 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
2531 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
2532 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
2533 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2534 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
2535 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
2536 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
2537 return ReplaceInstUsesWith(I, LHS);
2539 case ICmpInst::ICMP_NE:
2541 default: llvm_unreachable("Unknown integer condition code!");
2542 case ICmpInst::ICMP_ULT:
2543 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
2544 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
2545 break; // (X != 13 & X u< 15) -> no change
2546 case ICmpInst::ICMP_SLT:
2547 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
2548 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
2549 break; // (X != 13 & X s< 15) -> no change
2550 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
2551 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
2552 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
2553 return ReplaceInstUsesWith(I, RHS);
2554 case ICmpInst::ICMP_NE:
2555 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
2556 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
2557 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
2558 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
2559 ConstantInt::get(Add->getType(), 1));
2561 break; // (X != 13 & X != 15) -> no change
2564 case ICmpInst::ICMP_ULT:
2566 default: llvm_unreachable("Unknown integer condition code!");
2567 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
2568 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
2569 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2570 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
2572 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
2573 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
2574 return ReplaceInstUsesWith(I, LHS);
2575 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
2579 case ICmpInst::ICMP_SLT:
2581 default: llvm_unreachable("Unknown integer condition code!");
2582 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
2583 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
2584 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2585 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
2587 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
2588 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
2589 return ReplaceInstUsesWith(I, LHS);
2590 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
2594 case ICmpInst::ICMP_UGT:
2596 default: llvm_unreachable("Unknown integer condition code!");
2597 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
2598 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
2599 return ReplaceInstUsesWith(I, RHS);
2600 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
2602 case ICmpInst::ICMP_NE:
2603 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
2604 return new ICmpInst(LHSCC, Val, RHSCst);
2605 break; // (X u> 13 & X != 15) -> no change
2606 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
2607 return InsertRangeTest(Val, AddOne(LHSCst),
2608 RHSCst, false, true, I);
2609 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
2613 case ICmpInst::ICMP_SGT:
2615 default: llvm_unreachable("Unknown integer condition code!");
2616 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
2617 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
2618 return ReplaceInstUsesWith(I, RHS);
2619 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
2621 case ICmpInst::ICMP_NE:
2622 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
2623 return new ICmpInst(LHSCC, Val, RHSCst);
2624 break; // (X s> 13 & X != 15) -> no change
2625 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
2626 return InsertRangeTest(Val, AddOne(LHSCst),
2627 RHSCst, true, true, I);
2628 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
2637 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
2640 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
2641 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
2642 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
2643 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
2644 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
2645 // If either of the constants are nans, then the whole thing returns
2647 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
2648 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2649 return new FCmpInst(FCmpInst::FCMP_ORD,
2650 LHS->getOperand(0), RHS->getOperand(0));
2653 // Handle vector zeros. This occurs because the canonical form of
2654 // "fcmp ord x,x" is "fcmp ord x, 0".
2655 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
2656 isa<ConstantAggregateZero>(RHS->getOperand(1)))
2657 return new FCmpInst(FCmpInst::FCMP_ORD,
2658 LHS->getOperand(0), RHS->getOperand(0));
2662 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
2663 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
2664 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
2667 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
2668 // Swap RHS operands to match LHS.
2669 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
2670 std::swap(Op1LHS, Op1RHS);
2673 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
2674 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
2676 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
2678 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
2679 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2680 if (Op0CC == FCmpInst::FCMP_TRUE)
2681 return ReplaceInstUsesWith(I, RHS);
2682 if (Op1CC == FCmpInst::FCMP_TRUE)
2683 return ReplaceInstUsesWith(I, LHS);
2687 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
2688 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
2690 std::swap(LHS, RHS);
2691 std::swap(Op0Pred, Op1Pred);
2692 std::swap(Op0Ordered, Op1Ordered);
2695 // uno && ueq -> uno && (uno || eq) -> ueq
2696 // ord && olt -> ord && (ord && lt) -> olt
2697 if (Op0Ordered == Op1Ordered)
2698 return ReplaceInstUsesWith(I, RHS);
2700 // uno && oeq -> uno && (ord && eq) -> false
2701 // uno && ord -> false
2703 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2704 // ord && ueq -> ord && (uno || eq) -> oeq
2705 return cast<Instruction>(getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS));
2713 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
2714 bool Changed = SimplifyCommutative(I);
2715 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2717 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
2718 return ReplaceInstUsesWith(I, V);
2720 // See if we can simplify any instructions used by the instruction whose sole
2721 // purpose is to compute bits we don't care about.
2722 if (SimplifyDemandedInstructionBits(I))
2725 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
2726 const APInt &AndRHSMask = AndRHS->getValue();
2727 APInt NotAndRHS(~AndRHSMask);
2729 // Optimize a variety of ((val OP C1) & C2) combinations...
2730 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2731 Value *Op0LHS = Op0I->getOperand(0);
2732 Value *Op0RHS = Op0I->getOperand(1);
2733 switch (Op0I->getOpcode()) {
2735 case Instruction::Xor:
2736 case Instruction::Or:
2737 // If the mask is only needed on one incoming arm, push it up.
2738 if (!Op0I->hasOneUse()) break;
2740 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
2741 // Not masking anything out for the LHS, move to RHS.
2742 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
2743 Op0RHS->getName()+".masked");
2744 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
2746 if (!isa<Constant>(Op0RHS) &&
2747 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
2748 // Not masking anything out for the RHS, move to LHS.
2749 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
2750 Op0LHS->getName()+".masked");
2751 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
2755 case Instruction::Add:
2756 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
2757 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2758 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
2759 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
2760 return BinaryOperator::CreateAnd(V, AndRHS);
2761 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
2762 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
2765 case Instruction::Sub:
2766 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
2767 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2768 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
2769 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
2770 return BinaryOperator::CreateAnd(V, AndRHS);
2772 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
2773 // has 1's for all bits that the subtraction with A might affect.
2774 if (Op0I->hasOneUse()) {
2775 uint32_t BitWidth = AndRHSMask.getBitWidth();
2776 uint32_t Zeros = AndRHSMask.countLeadingZeros();
2777 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
2779 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
2780 if (!(A && A->isZero()) && // avoid infinite recursion.
2781 MaskedValueIsZero(Op0LHS, Mask)) {
2782 Value *NewNeg = Builder->CreateNeg(Op0RHS);
2783 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
2788 case Instruction::Shl:
2789 case Instruction::LShr:
2790 // (1 << x) & 1 --> zext(x == 0)
2791 // (1 >> x) & 1 --> zext(x == 0)
2792 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
2794 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
2795 return new ZExtInst(NewICmp, I.getType());
2800 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
2801 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
2803 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
2804 // If this is an integer truncation or change from signed-to-unsigned, and
2805 // if the source is an and/or with immediate, transform it. This
2806 // frequently occurs for bitfield accesses.
2807 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
2808 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
2809 CastOp->getNumOperands() == 2)
2810 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){
2811 if (CastOp->getOpcode() == Instruction::And) {
2812 // Change: and (cast (and X, C1) to T), C2
2813 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
2814 // This will fold the two constants together, which may allow
2815 // other simplifications.
2816 Value *NewCast = Builder->CreateTruncOrBitCast(
2817 CastOp->getOperand(0), I.getType(),
2818 CastOp->getName()+".shrunk");
2819 // trunc_or_bitcast(C1)&C2
2820 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
2821 C3 = ConstantExpr::getAnd(C3, AndRHS);
2822 return BinaryOperator::CreateAnd(NewCast, C3);
2823 } else if (CastOp->getOpcode() == Instruction::Or) {
2824 // Change: and (cast (or X, C1) to T), C2
2825 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
2826 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
2827 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
2829 return ReplaceInstUsesWith(I, AndRHS);
2835 // Try to fold constant and into select arguments.
2836 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2837 if (Instruction *R = FoldOpIntoSelect(I, SI))
2839 if (isa<PHINode>(Op0))
2840 if (Instruction *NV = FoldOpIntoPhi(I))
2845 // (~A & ~B) == (~(A | B)) - De Morgan's Law
2846 if (Value *Op0NotVal = dyn_castNotVal(Op0))
2847 if (Value *Op1NotVal = dyn_castNotVal(Op1))
2848 if (Op0->hasOneUse() && Op1->hasOneUse()) {
2849 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
2850 I.getName()+".demorgan");
2851 return BinaryOperator::CreateNot(Or);
2855 Value *A = 0, *B = 0, *C = 0, *D = 0;
2856 // (A|B) & ~(A&B) -> A^B
2857 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
2858 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
2859 ((A == C && B == D) || (A == D && B == C)))
2860 return BinaryOperator::CreateXor(A, B);
2862 // ~(A&B) & (A|B) -> A^B
2863 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
2864 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
2865 ((A == C && B == D) || (A == D && B == C)))
2866 return BinaryOperator::CreateXor(A, B);
2868 if (Op0->hasOneUse() &&
2869 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2870 if (A == Op1) { // (A^B)&A -> A&(A^B)
2871 I.swapOperands(); // Simplify below
2872 std::swap(Op0, Op1);
2873 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
2874 cast<BinaryOperator>(Op0)->swapOperands();
2875 I.swapOperands(); // Simplify below
2876 std::swap(Op0, Op1);
2880 if (Op1->hasOneUse() &&
2881 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2882 if (B == Op0) { // B&(A^B) -> B&(B^A)
2883 cast<BinaryOperator>(Op1)->swapOperands();
2886 if (A == Op0) // A&(A^B) -> A & ~B
2887 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
2890 // (A&((~A)|B)) -> A&B
2891 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
2892 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
2893 return BinaryOperator::CreateAnd(A, Op1);
2894 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
2895 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
2896 return BinaryOperator::CreateAnd(A, Op0);
2899 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
2900 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2901 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
2904 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
2905 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
2909 // fold (and (cast A), (cast B)) -> (cast (and A, B))
2910 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
2911 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2912 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
2913 const Type *SrcTy = Op0C->getOperand(0)->getType();
2914 if (SrcTy == Op1C->getOperand(0)->getType() &&
2915 SrcTy->isIntOrIntVector() &&
2916 // Only do this if the casts both really cause code to be generated.
2917 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
2919 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
2921 Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
2922 Op1C->getOperand(0), I.getName());
2923 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2927 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
2928 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
2929 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
2930 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
2931 SI0->getOperand(1) == SI1->getOperand(1) &&
2932 (SI0->hasOneUse() || SI1->hasOneUse())) {
2934 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
2936 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
2937 SI1->getOperand(1));
2941 // If and'ing two fcmp, try combine them into one.
2942 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
2943 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2944 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
2948 return Changed ? &I : 0;
2951 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
2952 /// capable of providing pieces of a bswap. The subexpression provides pieces
2953 /// of a bswap if it is proven that each of the non-zero bytes in the output of
2954 /// the expression came from the corresponding "byte swapped" byte in some other
2955 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
2956 /// we know that the expression deposits the low byte of %X into the high byte
2957 /// of the bswap result and that all other bytes are zero. This expression is
2958 /// accepted, the high byte of ByteValues is set to X to indicate a correct
2961 /// This function returns true if the match was unsuccessful and false if so.
2962 /// On entry to the function the "OverallLeftShift" is a signed integer value
2963 /// indicating the number of bytes that the subexpression is later shifted. For
2964 /// example, if the expression is later right shifted by 16 bits, the
2965 /// OverallLeftShift value would be -2 on entry. This is used to specify which
2966 /// byte of ByteValues is actually being set.
2968 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
2969 /// byte is masked to zero by a user. For example, in (X & 255), X will be
2970 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
2971 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
2972 /// always in the local (OverallLeftShift) coordinate space.
2974 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
2975 SmallVector<Value*, 8> &ByteValues) {
2976 if (Instruction *I = dyn_cast<Instruction>(V)) {
2977 // If this is an or instruction, it may be an inner node of the bswap.
2978 if (I->getOpcode() == Instruction::Or) {
2979 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
2981 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
2985 // If this is a logical shift by a constant multiple of 8, recurse with
2986 // OverallLeftShift and ByteMask adjusted.
2987 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
2989 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
2990 // Ensure the shift amount is defined and of a byte value.
2991 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
2994 unsigned ByteShift = ShAmt >> 3;
2995 if (I->getOpcode() == Instruction::Shl) {
2996 // X << 2 -> collect(X, +2)
2997 OverallLeftShift += ByteShift;
2998 ByteMask >>= ByteShift;
3000 // X >>u 2 -> collect(X, -2)
3001 OverallLeftShift -= ByteShift;
3002 ByteMask <<= ByteShift;
3003 ByteMask &= (~0U >> (32-ByteValues.size()));
3006 if (OverallLeftShift >= (int)ByteValues.size()) return true;
3007 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
3009 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
3013 // If this is a logical 'and' with a mask that clears bytes, clear the
3014 // corresponding bytes in ByteMask.
3015 if (I->getOpcode() == Instruction::And &&
3016 isa<ConstantInt>(I->getOperand(1))) {
3017 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
3018 unsigned NumBytes = ByteValues.size();
3019 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
3020 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
3022 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
3023 // If this byte is masked out by a later operation, we don't care what
3025 if ((ByteMask & (1 << i)) == 0)
3028 // If the AndMask is all zeros for this byte, clear the bit.
3029 APInt MaskB = AndMask & Byte;
3031 ByteMask &= ~(1U << i);
3035 // If the AndMask is not all ones for this byte, it's not a bytezap.
3039 // Otherwise, this byte is kept.
3042 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
3047 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
3048 // the input value to the bswap. Some observations: 1) if more than one byte
3049 // is demanded from this input, then it could not be successfully assembled
3050 // into a byteswap. At least one of the two bytes would not be aligned with
3051 // their ultimate destination.
3052 if (!isPowerOf2_32(ByteMask)) return true;
3053 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
3055 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
3056 // is demanded, it needs to go into byte 0 of the result. This means that the
3057 // byte needs to be shifted until it lands in the right byte bucket. The
3058 // shift amount depends on the position: if the byte is coming from the high
3059 // part of the value (e.g. byte 3) then it must be shifted right. If from the
3060 // low part, it must be shifted left.
3061 unsigned DestByteNo = InputByteNo + OverallLeftShift;
3062 if (InputByteNo < ByteValues.size()/2) {
3063 if (ByteValues.size()-1-DestByteNo != InputByteNo)
3066 if (ByteValues.size()-1-DestByteNo != InputByteNo)
3070 // If the destination byte value is already defined, the values are or'd
3071 // together, which isn't a bswap (unless it's an or of the same bits).
3072 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
3074 ByteValues[DestByteNo] = V;
3078 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3079 /// If so, insert the new bswap intrinsic and return it.
3080 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3081 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3082 if (!ITy || ITy->getBitWidth() % 16 ||
3083 // ByteMask only allows up to 32-byte values.
3084 ITy->getBitWidth() > 32*8)
3085 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3087 /// ByteValues - For each byte of the result, we keep track of which value
3088 /// defines each byte.
3089 SmallVector<Value*, 8> ByteValues;
3090 ByteValues.resize(ITy->getBitWidth()/8);
3092 // Try to find all the pieces corresponding to the bswap.
3093 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
3094 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
3097 // Check to see if all of the bytes come from the same value.
3098 Value *V = ByteValues[0];
3099 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3101 // Check to make sure that all of the bytes come from the same value.
3102 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3103 if (ByteValues[i] != V)
3105 const Type *Tys[] = { ITy };
3106 Module *M = I.getParent()->getParent()->getParent();
3107 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3108 return CallInst::Create(F, V);
3111 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
3112 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
3113 /// we can simplify this expression to "cond ? C : D or B".
3114 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
3115 Value *C, Value *D) {
3116 // If A is not a select of -1/0, this cannot match.
3118 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
3121 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
3122 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
3123 return SelectInst::Create(Cond, C, B);
3124 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
3125 return SelectInst::Create(Cond, C, B);
3126 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
3127 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
3128 return SelectInst::Create(Cond, C, D);
3129 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
3130 return SelectInst::Create(Cond, C, D);
3134 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
3135 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
3136 ICmpInst *LHS, ICmpInst *RHS) {
3138 ConstantInt *LHSCst, *RHSCst;
3139 ICmpInst::Predicate LHSCC, RHSCC;
3141 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
3142 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3143 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3147 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
3148 if (LHSCst == RHSCst && LHSCC == RHSCC &&
3149 LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
3150 Value *NewOr = Builder->CreateOr(Val, Val2);
3151 return new ICmpInst(LHSCC, NewOr, LHSCst);
3154 // From here on, we only handle:
3155 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
3156 if (Val != Val2) return 0;
3158 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3159 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3160 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3161 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3162 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3165 // We can't fold (ugt x, C) | (sgt x, C2).
3166 if (!PredicatesFoldable(LHSCC, RHSCC))
3169 // Ensure that the larger constant is on the RHS.
3171 if (CmpInst::isSigned(LHSCC) ||
3172 (ICmpInst::isEquality(LHSCC) &&
3173 CmpInst::isSigned(RHSCC)))
3174 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3176 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3179 std::swap(LHS, RHS);
3180 std::swap(LHSCst, RHSCst);
3181 std::swap(LHSCC, RHSCC);
3184 // At this point, we know we have have two icmp instructions
3185 // comparing a value against two constants and or'ing the result
3186 // together. Because of the above check, we know that we only have
3187 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3188 // FoldICmpLogical check above), that the two constants are not
3190 assert(LHSCst != RHSCst && "Compares not folded above?");
3193 default: llvm_unreachable("Unknown integer condition code!");
3194 case ICmpInst::ICMP_EQ:
3196 default: llvm_unreachable("Unknown integer condition code!");
3197 case ICmpInst::ICMP_EQ:
3198 if (LHSCst == SubOne(RHSCst)) {
3199 // (X == 13 | X == 14) -> X-13 <u 2
3200 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3201 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
3202 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3203 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3205 break; // (X == 13 | X == 15) -> no change
3206 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3207 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3209 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3210 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3211 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3212 return ReplaceInstUsesWith(I, RHS);
3215 case ICmpInst::ICMP_NE:
3217 default: llvm_unreachable("Unknown integer condition code!");
3218 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3219 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3220 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3221 return ReplaceInstUsesWith(I, LHS);
3222 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3223 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3224 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3225 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
3228 case ICmpInst::ICMP_ULT:
3230 default: llvm_unreachable("Unknown integer condition code!");
3231 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3233 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
3234 // If RHSCst is [us]MAXINT, it is always false. Not handling
3235 // this can cause overflow.
3236 if (RHSCst->isMaxValue(false))
3237 return ReplaceInstUsesWith(I, LHS);
3238 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
3240 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3242 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3243 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3244 return ReplaceInstUsesWith(I, RHS);
3245 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3249 case ICmpInst::ICMP_SLT:
3251 default: llvm_unreachable("Unknown integer condition code!");
3252 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3254 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
3255 // If RHSCst is [us]MAXINT, it is always false. Not handling
3256 // this can cause overflow.
3257 if (RHSCst->isMaxValue(true))
3258 return ReplaceInstUsesWith(I, LHS);
3259 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
3261 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3263 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3264 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3265 return ReplaceInstUsesWith(I, RHS);
3266 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3270 case ICmpInst::ICMP_UGT:
3272 default: llvm_unreachable("Unknown integer condition code!");
3273 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3274 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3275 return ReplaceInstUsesWith(I, LHS);
3276 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3278 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3279 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3280 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
3281 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3285 case ICmpInst::ICMP_SGT:
3287 default: llvm_unreachable("Unknown integer condition code!");
3288 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3289 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3290 return ReplaceInstUsesWith(I, LHS);
3291 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3293 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3294 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3295 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
3296 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3304 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
3306 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
3307 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
3308 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
3309 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3310 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3311 // If either of the constants are nans, then the whole thing returns
3313 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3314 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
3316 // Otherwise, no need to compare the two constants, compare the
3318 return new FCmpInst(FCmpInst::FCMP_UNO,
3319 LHS->getOperand(0), RHS->getOperand(0));
3322 // Handle vector zeros. This occurs because the canonical form of
3323 // "fcmp uno x,x" is "fcmp uno x, 0".
3324 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
3325 isa<ConstantAggregateZero>(RHS->getOperand(1)))
3326 return new FCmpInst(FCmpInst::FCMP_UNO,
3327 LHS->getOperand(0), RHS->getOperand(0));
3332 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
3333 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
3334 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
3336 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
3337 // Swap RHS operands to match LHS.
3338 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
3339 std::swap(Op1LHS, Op1RHS);
3341 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
3342 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
3344 return new FCmpInst((FCmpInst::Predicate)Op0CC,
3346 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
3347 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
3348 if (Op0CC == FCmpInst::FCMP_FALSE)
3349 return ReplaceInstUsesWith(I, RHS);
3350 if (Op1CC == FCmpInst::FCMP_FALSE)
3351 return ReplaceInstUsesWith(I, LHS);
3354 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
3355 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
3356 if (Op0Ordered == Op1Ordered) {
3357 // If both are ordered or unordered, return a new fcmp with
3358 // or'ed predicates.
3359 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS);
3360 if (Instruction *I = dyn_cast<Instruction>(RV))
3362 // Otherwise, it's a constant boolean value...
3363 return ReplaceInstUsesWith(I, RV);
3369 /// FoldOrWithConstants - This helper function folds:
3371 /// ((A | B) & C1) | (B & C2)
3377 /// when the XOR of the two constants is "all ones" (-1).
3378 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
3379 Value *A, Value *B, Value *C) {
3380 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
3384 ConstantInt *CI2 = 0;
3385 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
3387 APInt Xor = CI1->getValue() ^ CI2->getValue();
3388 if (!Xor.isAllOnesValue()) return 0;
3390 if (V1 == A || V1 == B) {
3391 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
3392 return BinaryOperator::CreateOr(NewOp, V1);
3398 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3399 bool Changed = SimplifyCommutative(I);
3400 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3402 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
3403 return ReplaceInstUsesWith(I, V);
3406 // See if we can simplify any instructions used by the instruction whose sole
3407 // purpose is to compute bits we don't care about.
3408 if (SimplifyDemandedInstructionBits(I))
3411 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3412 ConstantInt *C1 = 0; Value *X = 0;
3413 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3414 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
3416 Value *Or = Builder->CreateOr(X, RHS);
3418 return BinaryOperator::CreateAnd(Or,
3419 ConstantInt::get(I.getContext(),
3420 RHS->getValue() | C1->getValue()));
3423 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3424 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
3426 Value *Or = Builder->CreateOr(X, RHS);
3428 return BinaryOperator::CreateXor(Or,
3429 ConstantInt::get(I.getContext(),
3430 C1->getValue() & ~RHS->getValue()));
3433 // Try to fold constant and into select arguments.
3434 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3435 if (Instruction *R = FoldOpIntoSelect(I, SI))
3437 if (isa<PHINode>(Op0))
3438 if (Instruction *NV = FoldOpIntoPhi(I))
3442 Value *A = 0, *B = 0;
3443 ConstantInt *C1 = 0, *C2 = 0;
3445 // (A | B) | C and A | (B | C) -> bswap if possible.
3446 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3447 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3448 match(Op1, m_Or(m_Value(), m_Value())) ||
3449 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3450 match(Op1, m_Shift(m_Value(), m_Value())))) {
3451 if (Instruction *BSwap = MatchBSwap(I))
3455 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3456 if (Op0->hasOneUse() &&
3457 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3458 MaskedValueIsZero(Op1, C1->getValue())) {
3459 Value *NOr = Builder->CreateOr(A, Op1);
3461 return BinaryOperator::CreateXor(NOr, C1);
3464 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3465 if (Op1->hasOneUse() &&
3466 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3467 MaskedValueIsZero(Op0, C1->getValue())) {
3468 Value *NOr = Builder->CreateOr(A, Op0);
3470 return BinaryOperator::CreateXor(NOr, C1);
3474 Value *C = 0, *D = 0;
3475 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3476 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3477 Value *V1 = 0, *V2 = 0, *V3 = 0;
3478 C1 = dyn_cast<ConstantInt>(C);
3479 C2 = dyn_cast<ConstantInt>(D);
3480 if (C1 && C2) { // (A & C1)|(B & C2)
3481 // If we have: ((V + N) & C1) | (V & C2)
3482 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3483 // replace with V+N.
3484 if (C1->getValue() == ~C2->getValue()) {
3485 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3486 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3487 // Add commutes, try both ways.
3488 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3489 return ReplaceInstUsesWith(I, A);
3490 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3491 return ReplaceInstUsesWith(I, A);
3493 // Or commutes, try both ways.
3494 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3495 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3496 // Add commutes, try both ways.
3497 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3498 return ReplaceInstUsesWith(I, B);
3499 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3500 return ReplaceInstUsesWith(I, B);
3504 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
3505 // iff (C1&C2) == 0 and (N&~C1) == 0
3506 if ((C1->getValue() & C2->getValue()) == 0) {
3507 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
3508 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
3509 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
3510 return BinaryOperator::CreateAnd(A,
3511 ConstantInt::get(A->getContext(),
3512 C1->getValue()|C2->getValue()));
3513 // Or commutes, try both ways.
3514 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
3515 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
3516 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
3517 return BinaryOperator::CreateAnd(B,
3518 ConstantInt::get(B->getContext(),
3519 C1->getValue()|C2->getValue()));
3523 // Check to see if we have any common things being and'ed. If so, find the
3524 // terms for V1 & (V2|V3).
3525 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3527 if (A == B) // (A & C)|(A & D) == A & (C|D)
3528 V1 = A, V2 = C, V3 = D;
3529 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3530 V1 = A, V2 = B, V3 = C;
3531 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3532 V1 = C, V2 = A, V3 = D;
3533 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3534 V1 = C, V2 = A, V3 = B;
3537 Value *Or = Builder->CreateOr(V2, V3, "tmp");
3538 return BinaryOperator::CreateAnd(V1, Or);
3542 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
3543 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
3545 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
3547 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
3549 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
3552 // ((A&~B)|(~A&B)) -> A^B
3553 if ((match(C, m_Not(m_Specific(D))) &&
3554 match(B, m_Not(m_Specific(A)))))
3555 return BinaryOperator::CreateXor(A, D);
3556 // ((~B&A)|(~A&B)) -> A^B
3557 if ((match(A, m_Not(m_Specific(D))) &&
3558 match(B, m_Not(m_Specific(C)))))
3559 return BinaryOperator::CreateXor(C, D);
3560 // ((A&~B)|(B&~A)) -> A^B
3561 if ((match(C, m_Not(m_Specific(B))) &&
3562 match(D, m_Not(m_Specific(A)))))
3563 return BinaryOperator::CreateXor(A, B);
3564 // ((~B&A)|(B&~A)) -> A^B
3565 if ((match(A, m_Not(m_Specific(B))) &&
3566 match(D, m_Not(m_Specific(C)))))
3567 return BinaryOperator::CreateXor(C, B);
3570 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3571 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3572 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3573 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3574 SI0->getOperand(1) == SI1->getOperand(1) &&
3575 (SI0->hasOneUse() || SI1->hasOneUse())) {
3576 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
3578 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3579 SI1->getOperand(1));
3583 // ((A|B)&1)|(B&-2) -> (A&1) | B
3584 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
3585 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
3586 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
3587 if (Ret) return Ret;
3589 // (B&-2)|((A|B)&1) -> (A&1) | B
3590 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
3591 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
3592 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
3593 if (Ret) return Ret;
3596 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3597 if (Value *Op0NotVal = dyn_castNotVal(Op0))
3598 if (Value *Op1NotVal = dyn_castNotVal(Op1))
3599 if (Op0->hasOneUse() && Op1->hasOneUse()) {
3600 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
3601 I.getName()+".demorgan");
3602 return BinaryOperator::CreateNot(And);
3605 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3606 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3607 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3610 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
3611 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
3615 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3616 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3617 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3618 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3619 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
3620 !isa<ICmpInst>(Op1C->getOperand(0))) {
3621 const Type *SrcTy = Op0C->getOperand(0)->getType();
3622 if (SrcTy == Op1C->getOperand(0)->getType() &&
3623 SrcTy->isIntOrIntVector() &&
3624 // Only do this if the casts both really cause code to be
3626 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3628 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3630 Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
3631 Op1C->getOperand(0), I.getName());
3632 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3639 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
3640 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3641 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
3642 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
3646 return Changed ? &I : 0;
3651 // XorSelf - Implements: X ^ X --> 0
3654 XorSelf(Value *rhs) : RHS(rhs) {}
3655 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3656 Instruction *apply(BinaryOperator &Xor) const {
3663 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3664 bool Changed = SimplifyCommutative(I);
3665 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3667 if (isa<UndefValue>(Op1)) {
3668 if (isa<UndefValue>(Op0))
3669 // Handle undef ^ undef -> 0 special case. This is a common
3671 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3672 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3675 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3676 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3677 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
3678 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3681 // See if we can simplify any instructions used by the instruction whose sole
3682 // purpose is to compute bits we don't care about.
3683 if (SimplifyDemandedInstructionBits(I))
3685 if (isa<VectorType>(I.getType()))
3686 if (isa<ConstantAggregateZero>(Op1))
3687 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
3689 // Is this a ~ operation?
3690 if (Value *NotOp = dyn_castNotVal(&I)) {
3691 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
3692 if (Op0I->getOpcode() == Instruction::And ||
3693 Op0I->getOpcode() == Instruction::Or) {
3694 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
3695 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
3696 if (dyn_castNotVal(Op0I->getOperand(1)))
3697 Op0I->swapOperands();
3698 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3700 Builder->CreateNot(Op0I->getOperand(1),
3701 Op0I->getOperand(1)->getName()+".not");
3702 if (Op0I->getOpcode() == Instruction::And)
3703 return BinaryOperator::CreateOr(Op0NotVal, NotY);
3704 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
3707 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
3708 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
3709 if (isFreeToInvert(Op0I->getOperand(0)) &&
3710 isFreeToInvert(Op0I->getOperand(1))) {
3712 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
3714 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
3715 if (Op0I->getOpcode() == Instruction::And)
3716 return BinaryOperator::CreateOr(NotX, NotY);
3717 return BinaryOperator::CreateAnd(NotX, NotY);
3724 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3725 if (RHS->isOne() && Op0->hasOneUse()) {
3726 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
3727 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3728 return new ICmpInst(ICI->getInversePredicate(),
3729 ICI->getOperand(0), ICI->getOperand(1));
3731 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
3732 return new FCmpInst(FCI->getInversePredicate(),
3733 FCI->getOperand(0), FCI->getOperand(1));
3736 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
3737 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3738 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
3739 if (CI->hasOneUse() && Op0C->hasOneUse()) {
3740 Instruction::CastOps Opcode = Op0C->getOpcode();
3741 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
3742 (RHS == ConstantExpr::getCast(Opcode,
3743 ConstantInt::getTrue(I.getContext()),
3744 Op0C->getDestTy()))) {
3745 CI->setPredicate(CI->getInversePredicate());
3746 return CastInst::Create(Opcode, CI, Op0C->getType());
3752 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3753 // ~(c-X) == X-c-1 == X+(-c-1)
3754 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3755 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3756 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3757 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3758 ConstantInt::get(I.getType(), 1));
3759 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
3762 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
3763 if (Op0I->getOpcode() == Instruction::Add) {
3764 // ~(X-c) --> (-c-1)-X
3765 if (RHS->isAllOnesValue()) {
3766 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3767 return BinaryOperator::CreateSub(
3768 ConstantExpr::getSub(NegOp0CI,
3769 ConstantInt::get(I.getType(), 1)),
3770 Op0I->getOperand(0));
3771 } else if (RHS->getValue().isSignBit()) {
3772 // (X + C) ^ signbit -> (X + C + signbit)
3773 Constant *C = ConstantInt::get(I.getContext(),
3774 RHS->getValue() + Op0CI->getValue());
3775 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
3778 } else if (Op0I->getOpcode() == Instruction::Or) {
3779 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3780 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
3781 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3782 // Anything in both C1 and C2 is known to be zero, remove it from
3784 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3785 NewRHS = ConstantExpr::getAnd(NewRHS,
3786 ConstantExpr::getNot(CommonBits));
3788 I.setOperand(0, Op0I->getOperand(0));
3789 I.setOperand(1, NewRHS);
3796 // Try to fold constant and into select arguments.
3797 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3798 if (Instruction *R = FoldOpIntoSelect(I, SI))
3800 if (isa<PHINode>(Op0))
3801 if (Instruction *NV = FoldOpIntoPhi(I))
3805 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3807 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3809 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3811 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3814 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
3817 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
3818 if (A == Op0) { // B^(B|A) == (A|B)^B
3819 Op1I->swapOperands();
3821 std::swap(Op0, Op1);
3822 } else if (B == Op0) { // B^(A|B) == (A|B)^B
3823 I.swapOperands(); // Simplified below.
3824 std::swap(Op0, Op1);
3826 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
3827 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
3828 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
3829 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
3830 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
3832 if (A == Op0) { // A^(A&B) -> A^(B&A)
3833 Op1I->swapOperands();
3836 if (B == Op0) { // A^(B&A) -> (B&A)^A
3837 I.swapOperands(); // Simplified below.
3838 std::swap(Op0, Op1);
3843 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
3846 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
3847 Op0I->hasOneUse()) {
3848 if (A == Op1) // (B|A)^B == (A|B)^B
3850 if (B == Op1) // (A|B)^B == A & ~B
3851 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
3852 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
3853 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
3854 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
3855 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
3856 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
3858 if (A == Op1) // (A&B)^A -> (B&A)^A
3860 if (B == Op1 && // (B&A)^A == ~B & A
3861 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
3862 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
3867 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
3868 if (Op0I && Op1I && Op0I->isShift() &&
3869 Op0I->getOpcode() == Op1I->getOpcode() &&
3870 Op0I->getOperand(1) == Op1I->getOperand(1) &&
3871 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
3873 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
3875 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
3876 Op1I->getOperand(1));
3880 Value *A, *B, *C, *D;
3881 // (A & B)^(A | B) -> A ^ B
3882 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
3883 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
3884 if ((A == C && B == D) || (A == D && B == C))
3885 return BinaryOperator::CreateXor(A, B);
3887 // (A | B)^(A & B) -> A ^ B
3888 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
3889 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
3890 if ((A == C && B == D) || (A == D && B == C))
3891 return BinaryOperator::CreateXor(A, B);
3895 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
3896 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
3897 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
3898 // (X & Y)^(X & Y) -> (Y^Z) & X
3899 Value *X = 0, *Y = 0, *Z = 0;
3901 X = A, Y = B, Z = D;
3903 X = A, Y = B, Z = C;
3905 X = B, Y = A, Z = D;
3907 X = B, Y = A, Z = C;
3910 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
3911 return BinaryOperator::CreateAnd(NewOp, X);
3916 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
3917 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
3918 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3921 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
3922 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
3923 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3924 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
3925 const Type *SrcTy = Op0C->getOperand(0)->getType();
3926 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3927 // Only do this if the casts both really cause code to be generated.
3928 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3930 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3932 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
3933 Op1C->getOperand(0), I.getName());
3934 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3939 return Changed ? &I : 0;
3943 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
3944 return commonShiftTransforms(I);
3947 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
3948 return commonShiftTransforms(I);
3951 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
3952 if (Instruction *R = commonShiftTransforms(I))
3955 Value *Op0 = I.getOperand(0);
3957 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
3958 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
3959 if (CSI->isAllOnesValue())
3960 return ReplaceInstUsesWith(I, CSI);
3962 // See if we can turn a signed shr into an unsigned shr.
3963 if (MaskedValueIsZero(Op0,
3964 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
3965 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
3967 // Arithmetic shifting an all-sign-bit value is a no-op.
3968 unsigned NumSignBits = ComputeNumSignBits(Op0);
3969 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
3970 return ReplaceInstUsesWith(I, Op0);
3975 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
3976 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
3977 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3979 // shl X, 0 == X and shr X, 0 == X
3980 // shl 0, X == 0 and shr 0, X == 0
3981 if (Op1 == Constant::getNullValue(Op1->getType()) ||
3982 Op0 == Constant::getNullValue(Op0->getType()))
3983 return ReplaceInstUsesWith(I, Op0);
3985 if (isa<UndefValue>(Op0)) {
3986 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
3987 return ReplaceInstUsesWith(I, Op0);
3988 else // undef << X -> 0, undef >>u X -> 0
3989 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3991 if (isa<UndefValue>(Op1)) {
3992 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
3993 return ReplaceInstUsesWith(I, Op0);
3994 else // X << undef, X >>u undef -> 0
3995 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3998 // See if we can fold away this shift.
3999 if (SimplifyDemandedInstructionBits(I))
4002 // Try to fold constant and into select arguments.
4003 if (isa<Constant>(Op0))
4004 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
4005 if (Instruction *R = FoldOpIntoSelect(I, SI))
4008 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
4009 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
4014 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
4015 BinaryOperator &I) {
4016 bool isLeftShift = I.getOpcode() == Instruction::Shl;
4018 // See if we can simplify any instructions used by the instruction whose sole
4019 // purpose is to compute bits we don't care about.
4020 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
4022 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
4025 if (Op1->uge(TypeBits)) {
4026 if (I.getOpcode() != Instruction::AShr)
4027 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
4029 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
4034 // ((X*C1) << C2) == (X * (C1 << C2))
4035 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
4036 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
4037 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
4038 return BinaryOperator::CreateMul(BO->getOperand(0),
4039 ConstantExpr::getShl(BOOp, Op1));
4041 // Try to fold constant and into select arguments.
4042 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4043 if (Instruction *R = FoldOpIntoSelect(I, SI))
4045 if (isa<PHINode>(Op0))
4046 if (Instruction *NV = FoldOpIntoPhi(I))
4049 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
4050 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
4051 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
4052 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
4053 // place. Don't try to do this transformation in this case. Also, we
4054 // require that the input operand is a shift-by-constant so that we have
4055 // confidence that the shifts will get folded together. We could do this
4056 // xform in more cases, but it is unlikely to be profitable.
4057 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
4058 isa<ConstantInt>(TrOp->getOperand(1))) {
4059 // Okay, we'll do this xform. Make the shift of shift.
4060 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
4061 // (shift2 (shift1 & 0x00FF), c2)
4062 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
4064 // For logical shifts, the truncation has the effect of making the high
4065 // part of the register be zeros. Emulate this by inserting an AND to
4066 // clear the top bits as needed. This 'and' will usually be zapped by
4067 // other xforms later if dead.
4068 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
4069 unsigned DstSize = TI->getType()->getScalarSizeInBits();
4070 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
4072 // The mask we constructed says what the trunc would do if occurring
4073 // between the shifts. We want to know the effect *after* the second
4074 // shift. We know that it is a logical shift by a constant, so adjust the
4075 // mask as appropriate.
4076 if (I.getOpcode() == Instruction::Shl)
4077 MaskV <<= Op1->getZExtValue();
4079 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
4080 MaskV = MaskV.lshr(Op1->getZExtValue());
4084 Value *And = Builder->CreateAnd(NSh,
4085 ConstantInt::get(I.getContext(), MaskV),
4088 // Return the value truncated to the interesting size.
4089 return new TruncInst(And, I.getType());
4093 if (Op0->hasOneUse()) {
4094 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
4095 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
4098 switch (Op0BO->getOpcode()) {
4100 case Instruction::Add:
4101 case Instruction::And:
4102 case Instruction::Or:
4103 case Instruction::Xor: {
4104 // These operators commute.
4105 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
4106 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
4107 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
4108 m_Specific(Op1)))) {
4109 Value *YS = // (Y << C)
4110 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
4112 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
4113 Op0BO->getOperand(1)->getName());
4114 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
4115 return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
4116 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
4119 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
4120 Value *Op0BOOp1 = Op0BO->getOperand(1);
4121 if (isLeftShift && Op0BOOp1->hasOneUse() &&
4123 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
4124 m_ConstantInt(CC))) &&
4125 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
4126 Value *YS = // (Y << C)
4127 Builder->CreateShl(Op0BO->getOperand(0), Op1,
4130 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
4131 V1->getName()+".mask");
4132 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
4137 case Instruction::Sub: {
4138 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
4139 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
4140 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
4141 m_Specific(Op1)))) {
4142 Value *YS = // (Y << C)
4143 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
4145 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
4146 Op0BO->getOperand(0)->getName());
4147 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
4148 return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
4149 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
4152 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
4153 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
4154 match(Op0BO->getOperand(0),
4155 m_And(m_Shr(m_Value(V1), m_Value(V2)),
4156 m_ConstantInt(CC))) && V2 == Op1 &&
4157 cast<BinaryOperator>(Op0BO->getOperand(0))
4158 ->getOperand(0)->hasOneUse()) {
4159 Value *YS = // (Y << C)
4160 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
4162 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
4163 V1->getName()+".mask");
4165 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
4173 // If the operand is an bitwise operator with a constant RHS, and the
4174 // shift is the only use, we can pull it out of the shift.
4175 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
4176 bool isValid = true; // Valid only for And, Or, Xor
4177 bool highBitSet = false; // Transform if high bit of constant set?
4179 switch (Op0BO->getOpcode()) {
4180 default: isValid = false; break; // Do not perform transform!
4181 case Instruction::Add:
4182 isValid = isLeftShift;
4184 case Instruction::Or:
4185 case Instruction::Xor:
4188 case Instruction::And:
4193 // If this is a signed shift right, and the high bit is modified
4194 // by the logical operation, do not perform the transformation.
4195 // The highBitSet boolean indicates the value of the high bit of
4196 // the constant which would cause it to be modified for this
4199 if (isValid && I.getOpcode() == Instruction::AShr)
4200 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
4203 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
4206 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
4207 NewShift->takeName(Op0BO);
4209 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
4216 // Find out if this is a shift of a shift by a constant.
4217 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
4218 if (ShiftOp && !ShiftOp->isShift())
4221 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
4222 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
4223 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
4224 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
4225 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
4226 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
4227 Value *X = ShiftOp->getOperand(0);
4229 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
4231 const IntegerType *Ty = cast<IntegerType>(I.getType());
4233 // Check for (X << c1) << c2 and (X >> c1) >> c2
4234 if (I.getOpcode() == ShiftOp->getOpcode()) {
4235 // If this is oversized composite shift, then unsigned shifts get 0, ashr
4237 if (AmtSum >= TypeBits) {
4238 if (I.getOpcode() != Instruction::AShr)
4239 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4240 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
4243 return BinaryOperator::Create(I.getOpcode(), X,
4244 ConstantInt::get(Ty, AmtSum));
4247 if (ShiftOp->getOpcode() == Instruction::LShr &&
4248 I.getOpcode() == Instruction::AShr) {
4249 if (AmtSum >= TypeBits)
4250 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4252 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
4253 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
4256 if (ShiftOp->getOpcode() == Instruction::AShr &&
4257 I.getOpcode() == Instruction::LShr) {
4258 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
4259 if (AmtSum >= TypeBits)
4260 AmtSum = TypeBits-1;
4262 Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
4264 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
4265 return BinaryOperator::CreateAnd(Shift,
4266 ConstantInt::get(I.getContext(), Mask));
4269 // Okay, if we get here, one shift must be left, and the other shift must be
4270 // right. See if the amounts are equal.
4271 if (ShiftAmt1 == ShiftAmt2) {
4272 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
4273 if (I.getOpcode() == Instruction::Shl) {
4274 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
4275 return BinaryOperator::CreateAnd(X,
4276 ConstantInt::get(I.getContext(),Mask));
4278 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
4279 if (I.getOpcode() == Instruction::LShr) {
4280 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
4281 return BinaryOperator::CreateAnd(X,
4282 ConstantInt::get(I.getContext(), Mask));
4284 // We can simplify ((X << C) >>s C) into a trunc + sext.
4285 // NOTE: we could do this for any C, but that would make 'unusual' integer
4286 // types. For now, just stick to ones well-supported by the code
4288 const Type *SExtType = 0;
4289 switch (Ty->getBitWidth() - ShiftAmt1) {
4296 SExtType = IntegerType::get(I.getContext(),
4297 Ty->getBitWidth() - ShiftAmt1);
4302 return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
4303 // Otherwise, we can't handle it yet.
4304 } else if (ShiftAmt1 < ShiftAmt2) {
4305 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
4307 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
4308 if (I.getOpcode() == Instruction::Shl) {
4309 assert(ShiftOp->getOpcode() == Instruction::LShr ||
4310 ShiftOp->getOpcode() == Instruction::AShr);
4311 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
4313 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
4314 return BinaryOperator::CreateAnd(Shift,
4315 ConstantInt::get(I.getContext(),Mask));
4318 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
4319 if (I.getOpcode() == Instruction::LShr) {
4320 assert(ShiftOp->getOpcode() == Instruction::Shl);
4321 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
4323 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
4324 return BinaryOperator::CreateAnd(Shift,
4325 ConstantInt::get(I.getContext(),Mask));
4328 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
4330 assert(ShiftAmt2 < ShiftAmt1);
4331 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
4333 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
4334 if (I.getOpcode() == Instruction::Shl) {
4335 assert(ShiftOp->getOpcode() == Instruction::LShr ||
4336 ShiftOp->getOpcode() == Instruction::AShr);
4337 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
4338 ConstantInt::get(Ty, ShiftDiff));
4340 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
4341 return BinaryOperator::CreateAnd(Shift,
4342 ConstantInt::get(I.getContext(),Mask));
4345 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
4346 if (I.getOpcode() == Instruction::LShr) {
4347 assert(ShiftOp->getOpcode() == Instruction::Shl);
4348 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
4350 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
4351 return BinaryOperator::CreateAnd(Shift,
4352 ConstantInt::get(I.getContext(),Mask));
4355 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
4363 /// FindElementAtOffset - Given a type and a constant offset, determine whether
4364 /// or not there is a sequence of GEP indices into the type that will land us at
4365 /// the specified offset. If so, fill them into NewIndices and return the
4366 /// resultant element type, otherwise return null.
4367 const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset,
4368 SmallVectorImpl<Value*> &NewIndices) {
4370 if (!Ty->isSized()) return 0;
4372 // Start with the index over the outer type. Note that the type size
4373 // might be zero (even if the offset isn't zero) if the indexed type
4374 // is something like [0 x {int, int}]
4375 const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
4376 int64_t FirstIdx = 0;
4377 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
4378 FirstIdx = Offset/TySize;
4379 Offset -= FirstIdx*TySize;
4381 // Handle hosts where % returns negative instead of values [0..TySize).
4385 assert(Offset >= 0);
4387 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
4390 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
4392 // Index into the types. If we fail, set OrigBase to null.
4394 // Indexing into tail padding between struct/array elements.
4395 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
4398 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
4399 const StructLayout *SL = TD->getStructLayout(STy);
4400 assert(Offset < (int64_t)SL->getSizeInBytes() &&
4401 "Offset must stay within the indexed type");
4403 unsigned Elt = SL->getElementContainingOffset(Offset);
4404 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
4407 Offset -= SL->getElementOffset(Elt);
4408 Ty = STy->getElementType(Elt);
4409 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
4410 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
4411 assert(EltSize && "Cannot index into a zero-sized array");
4412 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
4414 Ty = AT->getElementType();
4416 // Otherwise, we can't index into the middle of this atomic type, bail.
4425 /// GetSelectFoldableOperands - We want to turn code that looks like this:
4427 /// %D = select %cond, %C, %A
4429 /// %C = select %cond, %B, 0
4432 /// Assuming that the specified instruction is an operand to the select, return
4433 /// a bitmask indicating which operands of this instruction are foldable if they
4434 /// equal the other incoming value of the select.
4436 static unsigned GetSelectFoldableOperands(Instruction *I) {
4437 switch (I->getOpcode()) {
4438 case Instruction::Add:
4439 case Instruction::Mul:
4440 case Instruction::And:
4441 case Instruction::Or:
4442 case Instruction::Xor:
4443 return 3; // Can fold through either operand.
4444 case Instruction::Sub: // Can only fold on the amount subtracted.
4445 case Instruction::Shl: // Can only fold on the shift amount.
4446 case Instruction::LShr:
4447 case Instruction::AShr:
4450 return 0; // Cannot fold
4454 /// GetSelectFoldableConstant - For the same transformation as the previous
4455 /// function, return the identity constant that goes into the select.
4456 static Constant *GetSelectFoldableConstant(Instruction *I) {
4457 switch (I->getOpcode()) {
4458 default: llvm_unreachable("This cannot happen!");
4459 case Instruction::Add:
4460 case Instruction::Sub:
4461 case Instruction::Or:
4462 case Instruction::Xor:
4463 case Instruction::Shl:
4464 case Instruction::LShr:
4465 case Instruction::AShr:
4466 return Constant::getNullValue(I->getType());
4467 case Instruction::And:
4468 return Constant::getAllOnesValue(I->getType());
4469 case Instruction::Mul:
4470 return ConstantInt::get(I->getType(), 1);
4474 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
4475 /// have the same opcode and only one use each. Try to simplify this.
4476 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
4478 if (TI->getNumOperands() == 1) {
4479 // If this is a non-volatile load or a cast from the same type,
4482 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
4485 return 0; // unknown unary op.
4488 // Fold this by inserting a select from the input values.
4489 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
4490 FI->getOperand(0), SI.getName()+".v");
4491 InsertNewInstBefore(NewSI, SI);
4492 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
4496 // Only handle binary operators here.
4497 if (!isa<BinaryOperator>(TI))
4500 // Figure out if the operations have any operands in common.
4501 Value *MatchOp, *OtherOpT, *OtherOpF;
4503 if (TI->getOperand(0) == FI->getOperand(0)) {
4504 MatchOp = TI->getOperand(0);
4505 OtherOpT = TI->getOperand(1);
4506 OtherOpF = FI->getOperand(1);
4507 MatchIsOpZero = true;
4508 } else if (TI->getOperand(1) == FI->getOperand(1)) {
4509 MatchOp = TI->getOperand(1);
4510 OtherOpT = TI->getOperand(0);
4511 OtherOpF = FI->getOperand(0);
4512 MatchIsOpZero = false;
4513 } else if (!TI->isCommutative()) {
4515 } else if (TI->getOperand(0) == FI->getOperand(1)) {
4516 MatchOp = TI->getOperand(0);
4517 OtherOpT = TI->getOperand(1);
4518 OtherOpF = FI->getOperand(0);
4519 MatchIsOpZero = true;
4520 } else if (TI->getOperand(1) == FI->getOperand(0)) {
4521 MatchOp = TI->getOperand(1);
4522 OtherOpT = TI->getOperand(0);
4523 OtherOpF = FI->getOperand(1);
4524 MatchIsOpZero = true;
4529 // If we reach here, they do have operations in common.
4530 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
4531 OtherOpF, SI.getName()+".v");
4532 InsertNewInstBefore(NewSI, SI);
4534 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
4536 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
4538 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
4540 llvm_unreachable("Shouldn't get here");
4544 static bool isSelect01(Constant *C1, Constant *C2) {
4545 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
4548 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
4551 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
4554 /// FoldSelectIntoOp - Try fold the select into one of the operands to
4555 /// facilitate further optimization.
4556 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
4558 // See the comment above GetSelectFoldableOperands for a description of the
4559 // transformation we are doing here.
4560 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
4561 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
4562 !isa<Constant>(FalseVal)) {
4563 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
4564 unsigned OpToFold = 0;
4565 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
4567 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
4572 Constant *C = GetSelectFoldableConstant(TVI);
4573 Value *OOp = TVI->getOperand(2-OpToFold);
4574 // Avoid creating select between 2 constants unless it's selecting
4576 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
4577 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
4578 InsertNewInstBefore(NewSel, SI);
4579 NewSel->takeName(TVI);
4580 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
4581 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
4582 llvm_unreachable("Unknown instruction!!");
4589 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
4590 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
4591 !isa<Constant>(TrueVal)) {
4592 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
4593 unsigned OpToFold = 0;
4594 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
4596 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
4601 Constant *C = GetSelectFoldableConstant(FVI);
4602 Value *OOp = FVI->getOperand(2-OpToFold);
4603 // Avoid creating select between 2 constants unless it's selecting
4605 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
4606 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
4607 InsertNewInstBefore(NewSel, SI);
4608 NewSel->takeName(FVI);
4609 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
4610 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
4611 llvm_unreachable("Unknown instruction!!");
4621 /// visitSelectInstWithICmp - Visit a SelectInst that has an
4622 /// ICmpInst as its first operand.
4624 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
4626 bool Changed = false;
4627 ICmpInst::Predicate Pred = ICI->getPredicate();
4628 Value *CmpLHS = ICI->getOperand(0);
4629 Value *CmpRHS = ICI->getOperand(1);
4630 Value *TrueVal = SI.getTrueValue();
4631 Value *FalseVal = SI.getFalseValue();
4633 // Check cases where the comparison is with a constant that
4634 // can be adjusted to fit the min/max idiom. We may edit ICI in
4635 // place here, so make sure the select is the only user.
4636 if (ICI->hasOneUse())
4637 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
4640 case ICmpInst::ICMP_ULT:
4641 case ICmpInst::ICMP_SLT: {
4642 // X < MIN ? T : F --> F
4643 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
4644 return ReplaceInstUsesWith(SI, FalseVal);
4645 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
4646 Constant *AdjustedRHS = SubOne(CI);
4647 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
4648 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
4649 Pred = ICmpInst::getSwappedPredicate(Pred);
4650 CmpRHS = AdjustedRHS;
4651 std::swap(FalseVal, TrueVal);
4652 ICI->setPredicate(Pred);
4653 ICI->setOperand(1, CmpRHS);
4654 SI.setOperand(1, TrueVal);
4655 SI.setOperand(2, FalseVal);
4660 case ICmpInst::ICMP_UGT:
4661 case ICmpInst::ICMP_SGT: {
4662 // X > MAX ? T : F --> F
4663 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
4664 return ReplaceInstUsesWith(SI, FalseVal);
4665 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
4666 Constant *AdjustedRHS = AddOne(CI);
4667 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
4668 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
4669 Pred = ICmpInst::getSwappedPredicate(Pred);
4670 CmpRHS = AdjustedRHS;
4671 std::swap(FalseVal, TrueVal);
4672 ICI->setPredicate(Pred);
4673 ICI->setOperand(1, CmpRHS);
4674 SI.setOperand(1, TrueVal);
4675 SI.setOperand(2, FalseVal);
4682 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
4683 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
4684 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
4685 if (match(TrueVal, m_ConstantInt<-1>()) &&
4686 match(FalseVal, m_ConstantInt<0>()))
4687 Pred = ICI->getPredicate();
4688 else if (match(TrueVal, m_ConstantInt<0>()) &&
4689 match(FalseVal, m_ConstantInt<-1>()))
4690 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
4692 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
4693 // If we are just checking for a icmp eq of a single bit and zext'ing it
4694 // to an integer, then shift the bit to the appropriate place and then
4695 // cast to integer to avoid the comparison.
4696 const APInt &Op1CV = CI->getValue();
4698 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
4699 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
4700 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
4701 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
4702 Value *In = ICI->getOperand(0);
4703 Value *Sh = ConstantInt::get(In->getType(),
4704 In->getType()->getScalarSizeInBits()-1);
4705 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
4706 In->getName()+".lobit"),
4708 if (In->getType() != SI.getType())
4709 In = CastInst::CreateIntegerCast(In, SI.getType(),
4710 true/*SExt*/, "tmp", ICI);
4712 if (Pred == ICmpInst::ICMP_SGT)
4713 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
4714 In->getName()+".not"), *ICI);
4716 return ReplaceInstUsesWith(SI, In);
4721 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
4722 // Transform (X == Y) ? X : Y -> Y
4723 if (Pred == ICmpInst::ICMP_EQ)
4724 return ReplaceInstUsesWith(SI, FalseVal);
4725 // Transform (X != Y) ? X : Y -> X
4726 if (Pred == ICmpInst::ICMP_NE)
4727 return ReplaceInstUsesWith(SI, TrueVal);
4728 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
4730 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
4731 // Transform (X == Y) ? Y : X -> X
4732 if (Pred == ICmpInst::ICMP_EQ)
4733 return ReplaceInstUsesWith(SI, FalseVal);
4734 // Transform (X != Y) ? Y : X -> Y
4735 if (Pred == ICmpInst::ICMP_NE)
4736 return ReplaceInstUsesWith(SI, TrueVal);
4737 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
4739 return Changed ? &SI : 0;
4743 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a
4744 /// PHI node (but the two may be in different blocks). See if the true/false
4745 /// values (V) are live in all of the predecessor blocks of the PHI. For
4746 /// example, cases like this cannot be mapped:
4748 /// X = phi [ C1, BB1], [C2, BB2]
4750 /// Z = select X, Y, 0
4752 /// because Y is not live in BB1/BB2.
4754 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V,
4755 const SelectInst &SI) {
4756 // If the value is a non-instruction value like a constant or argument, it
4757 // can always be mapped.
4758 const Instruction *I = dyn_cast<Instruction>(V);
4759 if (I == 0) return true;
4761 // If V is a PHI node defined in the same block as the condition PHI, we can
4762 // map the arguments.
4763 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
4765 if (const PHINode *VP = dyn_cast<PHINode>(I))
4766 if (VP->getParent() == CondPHI->getParent())
4769 // Otherwise, if the PHI and select are defined in the same block and if V is
4770 // defined in a different block, then we can transform it.
4771 if (SI.getParent() == CondPHI->getParent() &&
4772 I->getParent() != CondPHI->getParent())
4775 // Otherwise we have a 'hard' case and we can't tell without doing more
4776 // detailed dominator based analysis, punt.
4780 /// FoldSPFofSPF - We have an SPF (e.g. a min or max) of an SPF of the form:
4781 /// SPF2(SPF1(A, B), C)
4782 Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner,
4783 SelectPatternFlavor SPF1,
4786 SelectPatternFlavor SPF2, Value *C) {
4787 if (C == A || C == B) {
4788 // MAX(MAX(A, B), B) -> MAX(A, B)
4789 // MIN(MIN(a, b), a) -> MIN(a, b)
4791 return ReplaceInstUsesWith(Outer, Inner);
4793 // MAX(MIN(a, b), a) -> a
4794 // MIN(MAX(a, b), a) -> a
4795 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
4796 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
4797 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
4798 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
4799 return ReplaceInstUsesWith(Outer, C);
4802 // TODO: MIN(MIN(A, 23), 97)
4809 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
4810 Value *CondVal = SI.getCondition();
4811 Value *TrueVal = SI.getTrueValue();
4812 Value *FalseVal = SI.getFalseValue();
4814 // select true, X, Y -> X
4815 // select false, X, Y -> Y
4816 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
4817 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
4819 // select C, X, X -> X
4820 if (TrueVal == FalseVal)
4821 return ReplaceInstUsesWith(SI, TrueVal);
4823 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
4824 return ReplaceInstUsesWith(SI, FalseVal);
4825 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
4826 return ReplaceInstUsesWith(SI, TrueVal);
4827 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
4828 if (isa<Constant>(TrueVal))
4829 return ReplaceInstUsesWith(SI, TrueVal);
4831 return ReplaceInstUsesWith(SI, FalseVal);
4834 if (SI.getType() == Type::getInt1Ty(SI.getContext())) {
4835 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
4836 if (C->getZExtValue()) {
4837 // Change: A = select B, true, C --> A = or B, C
4838 return BinaryOperator::CreateOr(CondVal, FalseVal);
4840 // Change: A = select B, false, C --> A = and !B, C
4842 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
4843 "not."+CondVal->getName()), SI);
4844 return BinaryOperator::CreateAnd(NotCond, FalseVal);
4846 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
4847 if (C->getZExtValue() == false) {
4848 // Change: A = select B, C, false --> A = and B, C
4849 return BinaryOperator::CreateAnd(CondVal, TrueVal);
4851 // Change: A = select B, C, true --> A = or !B, C
4853 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
4854 "not."+CondVal->getName()), SI);
4855 return BinaryOperator::CreateOr(NotCond, TrueVal);
4859 // select a, b, a -> a&b
4860 // select a, a, b -> a|b
4861 if (CondVal == TrueVal)
4862 return BinaryOperator::CreateOr(CondVal, FalseVal);
4863 else if (CondVal == FalseVal)
4864 return BinaryOperator::CreateAnd(CondVal, TrueVal);
4867 // Selecting between two integer constants?
4868 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
4869 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
4870 // select C, 1, 0 -> zext C to int
4871 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
4872 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
4873 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
4874 // select C, 0, 1 -> zext !C to int
4876 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
4877 "not."+CondVal->getName()), SI);
4878 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
4881 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
4882 // If one of the constants is zero (we know they can't both be) and we
4883 // have an icmp instruction with zero, and we have an 'and' with the
4884 // non-constant value, eliminate this whole mess. This corresponds to
4885 // cases like this: ((X & 27) ? 27 : 0)
4886 if (TrueValC->isZero() || FalseValC->isZero())
4887 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
4888 cast<Constant>(IC->getOperand(1))->isNullValue())
4889 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
4890 if (ICA->getOpcode() == Instruction::And &&
4891 isa<ConstantInt>(ICA->getOperand(1)) &&
4892 (ICA->getOperand(1) == TrueValC ||
4893 ICA->getOperand(1) == FalseValC) &&
4894 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
4895 // Okay, now we know that everything is set up, we just don't
4896 // know whether we have a icmp_ne or icmp_eq and whether the
4897 // true or false val is the zero.
4898 bool ShouldNotVal = !TrueValC->isZero();
4899 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
4902 V = InsertNewInstBefore(BinaryOperator::Create(
4903 Instruction::Xor, V, ICA->getOperand(1)), SI);
4904 return ReplaceInstUsesWith(SI, V);
4909 // See if we are selecting two values based on a comparison of the two values.
4910 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
4911 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
4912 // Transform (X == Y) ? X : Y -> Y
4913 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
4914 // This is not safe in general for floating point:
4915 // consider X== -0, Y== +0.
4916 // It becomes safe if either operand is a nonzero constant.
4917 ConstantFP *CFPt, *CFPf;
4918 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
4919 !CFPt->getValueAPF().isZero()) ||
4920 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
4921 !CFPf->getValueAPF().isZero()))
4922 return ReplaceInstUsesWith(SI, FalseVal);
4924 // Transform (X != Y) ? X : Y -> X
4925 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
4926 return ReplaceInstUsesWith(SI, TrueVal);
4927 // NOTE: if we wanted to, this is where to detect MIN/MAX
4929 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
4930 // Transform (X == Y) ? Y : X -> X
4931 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
4932 // This is not safe in general for floating point:
4933 // consider X== -0, Y== +0.
4934 // It becomes safe if either operand is a nonzero constant.
4935 ConstantFP *CFPt, *CFPf;
4936 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
4937 !CFPt->getValueAPF().isZero()) ||
4938 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
4939 !CFPf->getValueAPF().isZero()))
4940 return ReplaceInstUsesWith(SI, FalseVal);
4942 // Transform (X != Y) ? Y : X -> Y
4943 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
4944 return ReplaceInstUsesWith(SI, TrueVal);
4945 // NOTE: if we wanted to, this is where to detect MIN/MAX
4947 // NOTE: if we wanted to, this is where to detect ABS
4950 // See if we are selecting two values based on a comparison of the two values.
4951 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
4952 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
4955 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
4956 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
4957 if (TI->hasOneUse() && FI->hasOneUse()) {
4958 Instruction *AddOp = 0, *SubOp = 0;
4960 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
4961 if (TI->getOpcode() == FI->getOpcode())
4962 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
4965 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
4966 // even legal for FP.
4967 if ((TI->getOpcode() == Instruction::Sub &&
4968 FI->getOpcode() == Instruction::Add) ||
4969 (TI->getOpcode() == Instruction::FSub &&
4970 FI->getOpcode() == Instruction::FAdd)) {
4971 AddOp = FI; SubOp = TI;
4972 } else if ((FI->getOpcode() == Instruction::Sub &&
4973 TI->getOpcode() == Instruction::Add) ||
4974 (FI->getOpcode() == Instruction::FSub &&
4975 TI->getOpcode() == Instruction::FAdd)) {
4976 AddOp = TI; SubOp = FI;
4980 Value *OtherAddOp = 0;
4981 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
4982 OtherAddOp = AddOp->getOperand(1);
4983 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
4984 OtherAddOp = AddOp->getOperand(0);
4988 // So at this point we know we have (Y -> OtherAddOp):
4989 // select C, (add X, Y), (sub X, Z)
4990 Value *NegVal; // Compute -Z
4991 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
4992 NegVal = ConstantExpr::getNeg(C);
4994 NegVal = InsertNewInstBefore(
4995 BinaryOperator::CreateNeg(SubOp->getOperand(1),
4999 Value *NewTrueOp = OtherAddOp;
5000 Value *NewFalseOp = NegVal;
5002 std::swap(NewTrueOp, NewFalseOp);
5003 Instruction *NewSel =
5004 SelectInst::Create(CondVal, NewTrueOp,
5005 NewFalseOp, SI.getName() + ".p");
5007 NewSel = InsertNewInstBefore(NewSel, SI);
5008 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
5013 // See if we can fold the select into one of our operands.
5014 if (SI.getType()->isInteger()) {
5015 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal))
5018 // MAX(MAX(a, b), a) -> MAX(a, b)
5019 // MIN(MIN(a, b), a) -> MIN(a, b)
5020 // MAX(MIN(a, b), a) -> a
5021 // MIN(MAX(a, b), a) -> a
5022 Value *LHS, *RHS, *LHS2, *RHS2;
5023 if (SelectPatternFlavor SPF = MatchSelectPattern(&SI, LHS, RHS)) {
5024 if (SelectPatternFlavor SPF2 = MatchSelectPattern(LHS, LHS2, RHS2))
5025 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2,
5028 if (SelectPatternFlavor SPF2 = MatchSelectPattern(RHS, LHS2, RHS2))
5029 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2,
5035 // ABS(-X) -> ABS(X)
5036 // ABS(ABS(X)) -> ABS(X)
5039 // See if we can fold the select into a phi node if the condition is a select.
5040 if (isa<PHINode>(SI.getCondition()))
5041 // The true/false values have to be live in the PHI predecessor's blocks.
5042 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
5043 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
5044 if (Instruction *NV = FoldOpIntoPhi(SI))
5047 if (BinaryOperator::isNot(CondVal)) {
5048 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
5049 SI.setOperand(1, FalseVal);
5050 SI.setOperand(2, TrueVal);
5057 /// EnforceKnownAlignment - If the specified pointer points to an object that
5058 /// we control, modify the object's alignment to PrefAlign. This isn't
5059 /// often possible though. If alignment is important, a more reliable approach
5060 /// is to simply align all global variables and allocation instructions to
5061 /// their preferred alignment from the beginning.
5063 static unsigned EnforceKnownAlignment(Value *V,
5064 unsigned Align, unsigned PrefAlign) {
5066 User *U = dyn_cast<User>(V);
5067 if (!U) return Align;
5069 switch (Operator::getOpcode(U)) {
5071 case Instruction::BitCast:
5072 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
5073 case Instruction::GetElementPtr: {
5074 // If all indexes are zero, it is just the alignment of the base pointer.
5075 bool AllZeroOperands = true;
5076 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
5077 if (!isa<Constant>(*i) ||
5078 !cast<Constant>(*i)->isNullValue()) {
5079 AllZeroOperands = false;
5083 if (AllZeroOperands) {
5084 // Treat this like a bitcast.
5085 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
5091 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
5092 // If there is a large requested alignment and we can, bump up the alignment
5094 if (!GV->isDeclaration()) {
5095 if (GV->getAlignment() >= PrefAlign)
5096 Align = GV->getAlignment();
5098 GV->setAlignment(PrefAlign);
5102 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
5103 // If there is a requested alignment and if this is an alloca, round up.
5104 if (AI->getAlignment() >= PrefAlign)
5105 Align = AI->getAlignment();
5107 AI->setAlignment(PrefAlign);
5115 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
5116 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
5117 /// and it is more than the alignment of the ultimate object, see if we can
5118 /// increase the alignment of the ultimate object, making this check succeed.
5119 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
5120 unsigned PrefAlign) {
5121 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
5122 sizeof(PrefAlign) * CHAR_BIT;
5123 APInt Mask = APInt::getAllOnesValue(BitWidth);
5124 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5125 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
5126 unsigned TrailZ = KnownZero.countTrailingOnes();
5127 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
5129 if (PrefAlign > Align)
5130 Align = EnforceKnownAlignment(V, Align, PrefAlign);
5132 // We don't need to make any adjustment.
5136 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
5137 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
5138 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
5139 unsigned MinAlign = std::min(DstAlign, SrcAlign);
5140 unsigned CopyAlign = MI->getAlignment();
5142 if (CopyAlign < MinAlign) {
5143 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
5148 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
5150 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
5151 if (MemOpLength == 0) return 0;
5153 // Source and destination pointer types are always "i8*" for intrinsic. See
5154 // if the size is something we can handle with a single primitive load/store.
5155 // A single load+store correctly handles overlapping memory in the memmove
5157 unsigned Size = MemOpLength->getZExtValue();
5158 if (Size == 0) return MI; // Delete this mem transfer.
5160 if (Size > 8 || (Size&(Size-1)))
5161 return 0; // If not 1/2/4/8 bytes, exit.
5163 // Use an integer load+store unless we can find something better.
5165 PointerType::getUnqual(IntegerType::get(MI->getContext(), Size<<3));
5167 // Memcpy forces the use of i8* for the source and destination. That means
5168 // that if you're using memcpy to move one double around, you'll get a cast
5169 // from double* to i8*. We'd much rather use a double load+store rather than
5170 // an i64 load+store, here because this improves the odds that the source or
5171 // dest address will be promotable. See if we can find a better type than the
5172 // integer datatype.
5173 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
5174 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
5175 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
5176 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
5177 // down through these levels if so.
5178 while (!SrcETy->isSingleValueType()) {
5179 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
5180 if (STy->getNumElements() == 1)
5181 SrcETy = STy->getElementType(0);
5184 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
5185 if (ATy->getNumElements() == 1)
5186 SrcETy = ATy->getElementType();
5193 if (SrcETy->isSingleValueType())
5194 NewPtrTy = PointerType::getUnqual(SrcETy);
5199 // If the memcpy/memmove provides better alignment info than we can
5201 SrcAlign = std::max(SrcAlign, CopyAlign);
5202 DstAlign = std::max(DstAlign, CopyAlign);
5204 Value *Src = Builder->CreateBitCast(MI->getOperand(2), NewPtrTy);
5205 Value *Dest = Builder->CreateBitCast(MI->getOperand(1), NewPtrTy);
5206 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
5207 InsertNewInstBefore(L, *MI);
5208 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
5210 // Set the size of the copy to 0, it will be deleted on the next iteration.
5211 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
5215 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
5216 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
5217 if (MI->getAlignment() < Alignment) {
5218 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
5223 // Extract the length and alignment and fill if they are constant.
5224 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
5225 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
5226 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(MI->getContext()))
5228 uint64_t Len = LenC->getZExtValue();
5229 Alignment = MI->getAlignment();
5231 // If the length is zero, this is a no-op
5232 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
5234 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
5235 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
5236 const Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
5238 Value *Dest = MI->getDest();
5239 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
5241 // Alignment 0 is identity for alignment 1 for memset, but not store.
5242 if (Alignment == 0) Alignment = 1;
5244 // Extract the fill value and store.
5245 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
5246 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
5247 Dest, false, Alignment), *MI);
5249 // Set the size of the copy to 0, it will be deleted on the next iteration.
5250 MI->setLength(Constant::getNullValue(LenC->getType()));
5258 /// visitCallInst - CallInst simplification. This mostly only handles folding
5259 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
5260 /// the heavy lifting.
5262 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
5263 if (isFreeCall(&CI))
5264 return visitFree(CI);
5266 // If the caller function is nounwind, mark the call as nounwind, even if the
5268 if (CI.getParent()->getParent()->doesNotThrow() &&
5269 !CI.doesNotThrow()) {
5270 CI.setDoesNotThrow();
5274 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
5275 if (!II) return visitCallSite(&CI);
5277 // Intrinsics cannot occur in an invoke, so handle them here instead of in
5279 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
5280 bool Changed = false;
5282 // memmove/cpy/set of zero bytes is a noop.
5283 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
5284 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
5286 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
5287 if (CI->getZExtValue() == 1) {
5288 // Replace the instruction with just byte operations. We would
5289 // transform other cases to loads/stores, but we don't know if
5290 // alignment is sufficient.
5294 // If we have a memmove and the source operation is a constant global,
5295 // then the source and dest pointers can't alias, so we can change this
5296 // into a call to memcpy.
5297 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
5298 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
5299 if (GVSrc->isConstant()) {
5300 Module *M = CI.getParent()->getParent()->getParent();
5301 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
5303 Tys[0] = CI.getOperand(3)->getType();
5305 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
5310 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
5311 // memmove(x,x,size) -> noop.
5312 if (MTI->getSource() == MTI->getDest())
5313 return EraseInstFromFunction(CI);
5316 // If we can determine a pointer alignment that is bigger than currently
5317 // set, update the alignment.
5318 if (isa<MemTransferInst>(MI)) {
5319 if (Instruction *I = SimplifyMemTransfer(MI))
5321 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
5322 if (Instruction *I = SimplifyMemSet(MSI))
5326 if (Changed) return II;
5329 switch (II->getIntrinsicID()) {
5331 case Intrinsic::bswap:
5332 // bswap(bswap(x)) -> x
5333 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
5334 if (Operand->getIntrinsicID() == Intrinsic::bswap)
5335 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
5337 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
5338 if (TruncInst *TI = dyn_cast<TruncInst>(II->getOperand(1))) {
5339 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
5340 if (Operand->getIntrinsicID() == Intrinsic::bswap) {
5341 unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
5342 TI->getType()->getPrimitiveSizeInBits();
5343 Value *CV = ConstantInt::get(Operand->getType(), C);
5344 Value *V = Builder->CreateLShr(Operand->getOperand(1), CV);
5345 return new TruncInst(V, TI->getType());
5350 case Intrinsic::powi:
5351 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getOperand(2))) {
5352 // powi(x, 0) -> 1.0
5353 if (Power->isZero())
5354 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
5357 return ReplaceInstUsesWith(CI, II->getOperand(1));
5358 // powi(x, -1) -> 1/x
5359 if (Power->isAllOnesValue())
5360 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
5365 case Intrinsic::uadd_with_overflow: {
5366 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
5367 const IntegerType *IT = cast<IntegerType>(II->getOperand(1)->getType());
5368 uint32_t BitWidth = IT->getBitWidth();
5369 APInt Mask = APInt::getSignBit(BitWidth);
5370 APInt LHSKnownZero(BitWidth, 0);
5371 APInt LHSKnownOne(BitWidth, 0);
5372 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
5373 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
5374 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
5376 if (LHSKnownNegative || LHSKnownPositive) {
5377 APInt RHSKnownZero(BitWidth, 0);
5378 APInt RHSKnownOne(BitWidth, 0);
5379 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
5380 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
5381 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
5382 if (LHSKnownNegative && RHSKnownNegative) {
5383 // The sign bit is set in both cases: this MUST overflow.
5384 // Create a simple add instruction, and insert it into the struct.
5385 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI);
5388 UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext())
5390 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
5391 return InsertValueInst::Create(Struct, Add, 0);
5394 if (LHSKnownPositive && RHSKnownPositive) {
5395 // The sign bit is clear in both cases: this CANNOT overflow.
5396 // Create a simple add instruction, and insert it into the struct.
5397 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI);
5400 UndefValue::get(LHS->getType()),
5401 ConstantInt::getFalse(II->getContext())
5403 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
5404 return InsertValueInst::Create(Struct, Add, 0);
5408 // FALL THROUGH uadd into sadd
5409 case Intrinsic::sadd_with_overflow:
5410 // Canonicalize constants into the RHS.
5411 if (isa<Constant>(II->getOperand(1)) &&
5412 !isa<Constant>(II->getOperand(2))) {
5413 Value *LHS = II->getOperand(1);
5414 II->setOperand(1, II->getOperand(2));
5415 II->setOperand(2, LHS);
5419 // X + undef -> undef
5420 if (isa<UndefValue>(II->getOperand(2)))
5421 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
5423 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(2))) {
5424 // X + 0 -> {X, false}
5425 if (RHS->isZero()) {
5427 UndefValue::get(II->getOperand(0)->getType()),
5428 ConstantInt::getFalse(II->getContext())
5430 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
5431 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
5435 case Intrinsic::usub_with_overflow:
5436 case Intrinsic::ssub_with_overflow:
5437 // undef - X -> undef
5438 // X - undef -> undef
5439 if (isa<UndefValue>(II->getOperand(1)) ||
5440 isa<UndefValue>(II->getOperand(2)))
5441 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
5443 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(2))) {
5444 // X - 0 -> {X, false}
5445 if (RHS->isZero()) {
5447 UndefValue::get(II->getOperand(1)->getType()),
5448 ConstantInt::getFalse(II->getContext())
5450 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
5451 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
5455 case Intrinsic::umul_with_overflow:
5456 case Intrinsic::smul_with_overflow:
5457 // Canonicalize constants into the RHS.
5458 if (isa<Constant>(II->getOperand(1)) &&
5459 !isa<Constant>(II->getOperand(2))) {
5460 Value *LHS = II->getOperand(1);
5461 II->setOperand(1, II->getOperand(2));
5462 II->setOperand(2, LHS);
5466 // X * undef -> undef
5467 if (isa<UndefValue>(II->getOperand(2)))
5468 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
5470 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getOperand(2))) {
5471 // X*0 -> {0, false}
5473 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
5475 // X * 1 -> {X, false}
5476 if (RHSI->equalsInt(1)) {
5478 UndefValue::get(II->getOperand(1)->getType()),
5479 ConstantInt::getFalse(II->getContext())
5481 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
5482 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
5486 case Intrinsic::ppc_altivec_lvx:
5487 case Intrinsic::ppc_altivec_lvxl:
5488 case Intrinsic::x86_sse_loadu_ps:
5489 case Intrinsic::x86_sse2_loadu_pd:
5490 case Intrinsic::x86_sse2_loadu_dq:
5491 // Turn PPC lvx -> load if the pointer is known aligned.
5492 // Turn X86 loadups -> load if the pointer is known aligned.
5493 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
5494 Value *Ptr = Builder->CreateBitCast(II->getOperand(1),
5495 PointerType::getUnqual(II->getType()));
5496 return new LoadInst(Ptr);
5499 case Intrinsic::ppc_altivec_stvx:
5500 case Intrinsic::ppc_altivec_stvxl:
5501 // Turn stvx -> store if the pointer is known aligned.
5502 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
5503 const Type *OpPtrTy =
5504 PointerType::getUnqual(II->getOperand(1)->getType());
5505 Value *Ptr = Builder->CreateBitCast(II->getOperand(2), OpPtrTy);
5506 return new StoreInst(II->getOperand(1), Ptr);
5509 case Intrinsic::x86_sse_storeu_ps:
5510 case Intrinsic::x86_sse2_storeu_pd:
5511 case Intrinsic::x86_sse2_storeu_dq:
5512 // Turn X86 storeu -> store if the pointer is known aligned.
5513 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
5514 const Type *OpPtrTy =
5515 PointerType::getUnqual(II->getOperand(2)->getType());
5516 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy);
5517 return new StoreInst(II->getOperand(2), Ptr);
5521 case Intrinsic::x86_sse_cvttss2si: {
5522 // These intrinsics only demands the 0th element of its input vector. If
5523 // we can simplify the input based on that, do so now.
5525 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
5526 APInt DemandedElts(VWidth, 1);
5527 APInt UndefElts(VWidth, 0);
5528 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
5530 II->setOperand(1, V);
5536 case Intrinsic::ppc_altivec_vperm:
5537 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
5538 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
5539 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
5541 // Check that all of the elements are integer constants or undefs.
5542 bool AllEltsOk = true;
5543 for (unsigned i = 0; i != 16; ++i) {
5544 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
5545 !isa<UndefValue>(Mask->getOperand(i))) {
5552 // Cast the input vectors to byte vectors.
5553 Value *Op0 = Builder->CreateBitCast(II->getOperand(1), Mask->getType());
5554 Value *Op1 = Builder->CreateBitCast(II->getOperand(2), Mask->getType());
5555 Value *Result = UndefValue::get(Op0->getType());
5557 // Only extract each element once.
5558 Value *ExtractedElts[32];
5559 memset(ExtractedElts, 0, sizeof(ExtractedElts));
5561 for (unsigned i = 0; i != 16; ++i) {
5562 if (isa<UndefValue>(Mask->getOperand(i)))
5564 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
5565 Idx &= 31; // Match the hardware behavior.
5567 if (ExtractedElts[Idx] == 0) {
5568 ExtractedElts[Idx] =
5569 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
5570 ConstantInt::get(Type::getInt32Ty(II->getContext()),
5571 Idx&15, false), "tmp");
5574 // Insert this value into the result vector.
5575 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
5576 ConstantInt::get(Type::getInt32Ty(II->getContext()),
5579 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
5584 case Intrinsic::stackrestore: {
5585 // If the save is right next to the restore, remove the restore. This can
5586 // happen when variable allocas are DCE'd.
5587 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
5588 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
5589 BasicBlock::iterator BI = SS;
5591 return EraseInstFromFunction(CI);
5595 // Scan down this block to see if there is another stack restore in the
5596 // same block without an intervening call/alloca.
5597 BasicBlock::iterator BI = II;
5598 TerminatorInst *TI = II->getParent()->getTerminator();
5599 bool CannotRemove = false;
5600 for (++BI; &*BI != TI; ++BI) {
5601 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
5602 CannotRemove = true;
5605 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
5606 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
5607 // If there is a stackrestore below this one, remove this one.
5608 if (II->getIntrinsicID() == Intrinsic::stackrestore)
5609 return EraseInstFromFunction(CI);
5610 // Otherwise, ignore the intrinsic.
5612 // If we found a non-intrinsic call, we can't remove the stack
5614 CannotRemove = true;
5620 // If the stack restore is in a return/unwind block and if there are no
5621 // allocas or calls between the restore and the return, nuke the restore.
5622 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
5623 return EraseInstFromFunction(CI);
5628 return visitCallSite(II);
5631 // InvokeInst simplification
5633 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
5634 return visitCallSite(&II);
5637 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
5638 /// passed through the varargs area, we can eliminate the use of the cast.
5639 static bool isSafeToEliminateVarargsCast(const CallSite CS,
5640 const CastInst * const CI,
5641 const TargetData * const TD,
5643 if (!CI->isLosslessCast())
5646 // The size of ByVal arguments is derived from the type, so we
5647 // can't change to a type with a different size. If the size were
5648 // passed explicitly we could avoid this check.
5649 if (!CS.paramHasAttr(ix, Attribute::ByVal))
5653 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
5654 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
5655 if (!SrcTy->isSized() || !DstTy->isSized())
5657 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
5662 // visitCallSite - Improvements for call and invoke instructions.
5664 Instruction *InstCombiner::visitCallSite(CallSite CS) {
5665 bool Changed = false;
5667 // If the callee is a constexpr cast of a function, attempt to move the cast
5668 // to the arguments of the call/invoke.
5669 if (transformConstExprCastCall(CS)) return 0;
5671 Value *Callee = CS.getCalledValue();
5673 if (Function *CalleeF = dyn_cast<Function>(Callee))
5674 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
5675 Instruction *OldCall = CS.getInstruction();
5676 // If the call and callee calling conventions don't match, this call must
5677 // be unreachable, as the call is undefined.
5678 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
5679 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
5681 // If OldCall dues not return void then replaceAllUsesWith undef.
5682 // This allows ValueHandlers and custom metadata to adjust itself.
5683 if (!OldCall->getType()->isVoidTy())
5684 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
5685 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
5686 return EraseInstFromFunction(*OldCall);
5690 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
5691 // This instruction is not reachable, just remove it. We insert a store to
5692 // undef so that we know that this code is not reachable, despite the fact
5693 // that we can't modify the CFG here.
5694 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
5695 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
5696 CS.getInstruction());
5698 // If CS dues not return void then replaceAllUsesWith undef.
5699 // This allows ValueHandlers and custom metadata to adjust itself.
5700 if (!CS.getInstruction()->getType()->isVoidTy())
5701 CS.getInstruction()->
5702 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
5704 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
5705 // Don't break the CFG, insert a dummy cond branch.
5706 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
5707 ConstantInt::getTrue(Callee->getContext()), II);
5709 return EraseInstFromFunction(*CS.getInstruction());
5712 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
5713 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
5714 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
5715 return transformCallThroughTrampoline(CS);
5717 const PointerType *PTy = cast<PointerType>(Callee->getType());
5718 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
5719 if (FTy->isVarArg()) {
5720 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
5721 // See if we can optimize any arguments passed through the varargs area of
5723 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
5724 E = CS.arg_end(); I != E; ++I, ++ix) {
5725 CastInst *CI = dyn_cast<CastInst>(*I);
5726 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
5727 *I = CI->getOperand(0);
5733 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
5734 // Inline asm calls cannot throw - mark them 'nounwind'.
5735 CS.setDoesNotThrow();
5739 return Changed ? CS.getInstruction() : 0;
5742 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
5743 // attempt to move the cast to the arguments of the call/invoke.
5745 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
5746 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
5747 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
5748 if (CE->getOpcode() != Instruction::BitCast ||
5749 !isa<Function>(CE->getOperand(0)))
5751 Function *Callee = cast<Function>(CE->getOperand(0));
5752 Instruction *Caller = CS.getInstruction();
5753 const AttrListPtr &CallerPAL = CS.getAttributes();
5755 // Okay, this is a cast from a function to a different type. Unless doing so
5756 // would cause a type conversion of one of our arguments, change this call to
5757 // be a direct call with arguments casted to the appropriate types.
5759 const FunctionType *FT = Callee->getFunctionType();
5760 const Type *OldRetTy = Caller->getType();
5761 const Type *NewRetTy = FT->getReturnType();
5763 if (isa<StructType>(NewRetTy))
5764 return false; // TODO: Handle multiple return values.
5766 // Check to see if we are changing the return type...
5767 if (OldRetTy != NewRetTy) {
5768 if (Callee->isDeclaration() &&
5769 // Conversion is ok if changing from one pointer type to another or from
5770 // a pointer to an integer of the same size.
5771 !((isa<PointerType>(OldRetTy) || !TD ||
5772 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
5773 (isa<PointerType>(NewRetTy) || !TD ||
5774 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
5775 return false; // Cannot transform this return value.
5777 if (!Caller->use_empty() &&
5778 // void -> non-void is handled specially
5779 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
5780 return false; // Cannot transform this return value.
5782 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
5783 Attributes RAttrs = CallerPAL.getRetAttributes();
5784 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
5785 return false; // Attribute not compatible with transformed value.
5788 // If the callsite is an invoke instruction, and the return value is used by
5789 // a PHI node in a successor, we cannot change the return type of the call
5790 // because there is no place to put the cast instruction (without breaking
5791 // the critical edge). Bail out in this case.
5792 if (!Caller->use_empty())
5793 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
5794 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
5796 if (PHINode *PN = dyn_cast<PHINode>(*UI))
5797 if (PN->getParent() == II->getNormalDest() ||
5798 PN->getParent() == II->getUnwindDest())
5802 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
5803 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
5805 CallSite::arg_iterator AI = CS.arg_begin();
5806 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
5807 const Type *ParamTy = FT->getParamType(i);
5808 const Type *ActTy = (*AI)->getType();
5810 if (!CastInst::isCastable(ActTy, ParamTy))
5811 return false; // Cannot transform this parameter value.
5813 if (CallerPAL.getParamAttributes(i + 1)
5814 & Attribute::typeIncompatible(ParamTy))
5815 return false; // Attribute not compatible with transformed value.
5817 // Converting from one pointer type to another or between a pointer and an
5818 // integer of the same size is safe even if we do not have a body.
5819 bool isConvertible = ActTy == ParamTy ||
5820 (TD && ((isa<PointerType>(ParamTy) ||
5821 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
5822 (isa<PointerType>(ActTy) ||
5823 ActTy == TD->getIntPtrType(Caller->getContext()))));
5824 if (Callee->isDeclaration() && !isConvertible) return false;
5827 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
5828 Callee->isDeclaration())
5829 return false; // Do not delete arguments unless we have a function body.
5831 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
5832 !CallerPAL.isEmpty())
5833 // In this case we have more arguments than the new function type, but we
5834 // won't be dropping them. Check that these extra arguments have attributes
5835 // that are compatible with being a vararg call argument.
5836 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
5837 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
5839 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
5840 if (PAttrs & Attribute::VarArgsIncompatible)
5844 // Okay, we decided that this is a safe thing to do: go ahead and start
5845 // inserting cast instructions as necessary...
5846 std::vector<Value*> Args;
5847 Args.reserve(NumActualArgs);
5848 SmallVector<AttributeWithIndex, 8> attrVec;
5849 attrVec.reserve(NumCommonArgs);
5851 // Get any return attributes.
5852 Attributes RAttrs = CallerPAL.getRetAttributes();
5854 // If the return value is not being used, the type may not be compatible
5855 // with the existing attributes. Wipe out any problematic attributes.
5856 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
5858 // Add the new return attributes.
5860 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
5862 AI = CS.arg_begin();
5863 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
5864 const Type *ParamTy = FT->getParamType(i);
5865 if ((*AI)->getType() == ParamTy) {
5866 Args.push_back(*AI);
5868 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
5869 false, ParamTy, false);
5870 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
5873 // Add any parameter attributes.
5874 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
5875 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
5878 // If the function takes more arguments than the call was taking, add them
5880 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
5881 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
5883 // If we are removing arguments to the function, emit an obnoxious warning.
5884 if (FT->getNumParams() < NumActualArgs) {
5885 if (!FT->isVarArg()) {
5886 errs() << "WARNING: While resolving call to function '"
5887 << Callee->getName() << "' arguments were dropped!\n";
5889 // Add all of the arguments in their promoted form to the arg list.
5890 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
5891 const Type *PTy = getPromotedType((*AI)->getType());
5892 if (PTy != (*AI)->getType()) {
5893 // Must promote to pass through va_arg area!
5894 Instruction::CastOps opcode =
5895 CastInst::getCastOpcode(*AI, false, PTy, false);
5896 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
5898 Args.push_back(*AI);
5901 // Add any parameter attributes.
5902 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
5903 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
5908 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
5909 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
5911 if (NewRetTy->isVoidTy())
5912 Caller->setName(""); // Void type should not have a name.
5914 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
5918 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
5919 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
5920 Args.begin(), Args.end(),
5921 Caller->getName(), Caller);
5922 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
5923 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
5925 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
5926 Caller->getName(), Caller);
5927 CallInst *CI = cast<CallInst>(Caller);
5928 if (CI->isTailCall())
5929 cast<CallInst>(NC)->setTailCall();
5930 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
5931 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
5934 // Insert a cast of the return type as necessary.
5936 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
5937 if (!NV->getType()->isVoidTy()) {
5938 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
5940 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
5942 // If this is an invoke instruction, we should insert it after the first
5943 // non-phi, instruction in the normal successor block.
5944 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
5945 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
5946 InsertNewInstBefore(NC, *I);
5948 // Otherwise, it's a call, just insert cast right after the call instr
5949 InsertNewInstBefore(NC, *Caller);
5951 Worklist.AddUsersToWorkList(*Caller);
5953 NV = UndefValue::get(Caller->getType());
5958 if (!Caller->use_empty())
5959 Caller->replaceAllUsesWith(NV);
5961 EraseInstFromFunction(*Caller);
5965 // transformCallThroughTrampoline - Turn a call to a function created by the
5966 // init_trampoline intrinsic into a direct call to the underlying function.
5968 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
5969 Value *Callee = CS.getCalledValue();
5970 const PointerType *PTy = cast<PointerType>(Callee->getType());
5971 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
5972 const AttrListPtr &Attrs = CS.getAttributes();
5974 // If the call already has the 'nest' attribute somewhere then give up -
5975 // otherwise 'nest' would occur twice after splicing in the chain.
5976 if (Attrs.hasAttrSomewhere(Attribute::Nest))
5979 IntrinsicInst *Tramp =
5980 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
5982 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
5983 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
5984 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
5986 const AttrListPtr &NestAttrs = NestF->getAttributes();
5987 if (!NestAttrs.isEmpty()) {
5988 unsigned NestIdx = 1;
5989 const Type *NestTy = 0;
5990 Attributes NestAttr = Attribute::None;
5992 // Look for a parameter marked with the 'nest' attribute.
5993 for (FunctionType::param_iterator I = NestFTy->param_begin(),
5994 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
5995 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
5996 // Record the parameter type and any other attributes.
5998 NestAttr = NestAttrs.getParamAttributes(NestIdx);
6003 Instruction *Caller = CS.getInstruction();
6004 std::vector<Value*> NewArgs;
6005 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
6007 SmallVector<AttributeWithIndex, 8> NewAttrs;
6008 NewAttrs.reserve(Attrs.getNumSlots() + 1);
6010 // Insert the nest argument into the call argument list, which may
6011 // mean appending it. Likewise for attributes.
6013 // Add any result attributes.
6014 if (Attributes Attr = Attrs.getRetAttributes())
6015 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
6019 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
6021 if (Idx == NestIdx) {
6022 // Add the chain argument and attributes.
6023 Value *NestVal = Tramp->getOperand(3);
6024 if (NestVal->getType() != NestTy)
6025 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
6026 NewArgs.push_back(NestVal);
6027 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
6033 // Add the original argument and attributes.
6034 NewArgs.push_back(*I);
6035 if (Attributes Attr = Attrs.getParamAttributes(Idx))
6037 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
6043 // Add any function attributes.
6044 if (Attributes Attr = Attrs.getFnAttributes())
6045 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
6047 // The trampoline may have been bitcast to a bogus type (FTy).
6048 // Handle this by synthesizing a new function type, equal to FTy
6049 // with the chain parameter inserted.
6051 std::vector<const Type*> NewTypes;
6052 NewTypes.reserve(FTy->getNumParams()+1);
6054 // Insert the chain's type into the list of parameter types, which may
6055 // mean appending it.
6058 FunctionType::param_iterator I = FTy->param_begin(),
6059 E = FTy->param_end();
6063 // Add the chain's type.
6064 NewTypes.push_back(NestTy);
6069 // Add the original type.
6070 NewTypes.push_back(*I);
6076 // Replace the trampoline call with a direct call. Let the generic
6077 // code sort out any function type mismatches.
6078 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
6080 Constant *NewCallee =
6081 NestF->getType() == PointerType::getUnqual(NewFTy) ?
6082 NestF : ConstantExpr::getBitCast(NestF,
6083 PointerType::getUnqual(NewFTy));
6084 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
6087 Instruction *NewCaller;
6088 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
6089 NewCaller = InvokeInst::Create(NewCallee,
6090 II->getNormalDest(), II->getUnwindDest(),
6091 NewArgs.begin(), NewArgs.end(),
6092 Caller->getName(), Caller);
6093 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
6094 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
6096 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
6097 Caller->getName(), Caller);
6098 if (cast<CallInst>(Caller)->isTailCall())
6099 cast<CallInst>(NewCaller)->setTailCall();
6100 cast<CallInst>(NewCaller)->
6101 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
6102 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
6104 if (!Caller->getType()->isVoidTy())
6105 Caller->replaceAllUsesWith(NewCaller);
6106 Caller->eraseFromParent();
6107 Worklist.Remove(Caller);
6112 // Replace the trampoline call with a direct call. Since there is no 'nest'
6113 // parameter, there is no need to adjust the argument list. Let the generic
6114 // code sort out any function type mismatches.
6115 Constant *NewCallee =
6116 NestF->getType() == PTy ? NestF :
6117 ConstantExpr::getBitCast(NestF, PTy);
6118 CS.setCalledFunction(NewCallee);
6119 return CS.getInstruction();
6122 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
6123 /// and if a/b/c and the add's all have a single use, turn this into a phi
6124 /// and a single binop.
6125 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
6126 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
6127 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
6128 unsigned Opc = FirstInst->getOpcode();
6129 Value *LHSVal = FirstInst->getOperand(0);
6130 Value *RHSVal = FirstInst->getOperand(1);
6132 const Type *LHSType = LHSVal->getType();
6133 const Type *RHSType = RHSVal->getType();
6135 // Scan to see if all operands are the same opcode, and all have one use.
6136 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
6137 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
6138 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
6139 // Verify type of the LHS matches so we don't fold cmp's of different
6140 // types or GEP's with different index types.
6141 I->getOperand(0)->getType() != LHSType ||
6142 I->getOperand(1)->getType() != RHSType)
6145 // If they are CmpInst instructions, check their predicates
6146 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
6147 if (cast<CmpInst>(I)->getPredicate() !=
6148 cast<CmpInst>(FirstInst)->getPredicate())
6151 // Keep track of which operand needs a phi node.
6152 if (I->getOperand(0) != LHSVal) LHSVal = 0;
6153 if (I->getOperand(1) != RHSVal) RHSVal = 0;
6156 // If both LHS and RHS would need a PHI, don't do this transformation,
6157 // because it would increase the number of PHIs entering the block,
6158 // which leads to higher register pressure. This is especially
6159 // bad when the PHIs are in the header of a loop.
6160 if (!LHSVal && !RHSVal)
6163 // Otherwise, this is safe to transform!
6165 Value *InLHS = FirstInst->getOperand(0);
6166 Value *InRHS = FirstInst->getOperand(1);
6167 PHINode *NewLHS = 0, *NewRHS = 0;
6169 NewLHS = PHINode::Create(LHSType,
6170 FirstInst->getOperand(0)->getName() + ".pn");
6171 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
6172 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
6173 InsertNewInstBefore(NewLHS, PN);
6178 NewRHS = PHINode::Create(RHSType,
6179 FirstInst->getOperand(1)->getName() + ".pn");
6180 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
6181 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
6182 InsertNewInstBefore(NewRHS, PN);
6186 // Add all operands to the new PHIs.
6187 if (NewLHS || NewRHS) {
6188 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6189 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
6191 Value *NewInLHS = InInst->getOperand(0);
6192 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
6195 Value *NewInRHS = InInst->getOperand(1);
6196 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
6201 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
6202 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
6203 CmpInst *CIOp = cast<CmpInst>(FirstInst);
6204 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
6208 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
6209 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
6211 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
6212 FirstInst->op_end());
6213 // This is true if all GEP bases are allocas and if all indices into them are
6215 bool AllBasePointersAreAllocas = true;
6217 // We don't want to replace this phi if the replacement would require
6218 // more than one phi, which leads to higher register pressure. This is
6219 // especially bad when the PHIs are in the header of a loop.
6220 bool NeededPhi = false;
6222 // Scan to see if all operands are the same opcode, and all have one use.
6223 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
6224 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
6225 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
6226 GEP->getNumOperands() != FirstInst->getNumOperands())
6229 // Keep track of whether or not all GEPs are of alloca pointers.
6230 if (AllBasePointersAreAllocas &&
6231 (!isa<AllocaInst>(GEP->getOperand(0)) ||
6232 !GEP->hasAllConstantIndices()))
6233 AllBasePointersAreAllocas = false;
6235 // Compare the operand lists.
6236 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
6237 if (FirstInst->getOperand(op) == GEP->getOperand(op))
6240 // Don't merge two GEPs when two operands differ (introducing phi nodes)
6241 // if one of the PHIs has a constant for the index. The index may be
6242 // substantially cheaper to compute for the constants, so making it a
6243 // variable index could pessimize the path. This also handles the case
6244 // for struct indices, which must always be constant.
6245 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
6246 isa<ConstantInt>(GEP->getOperand(op)))
6249 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
6252 // If we already needed a PHI for an earlier operand, and another operand
6253 // also requires a PHI, we'd be introducing more PHIs than we're
6254 // eliminating, which increases register pressure on entry to the PHI's
6259 FixedOperands[op] = 0; // Needs a PHI.
6264 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
6265 // bother doing this transformation. At best, this will just save a bit of
6266 // offset calculation, but all the predecessors will have to materialize the
6267 // stack address into a register anyway. We'd actually rather *clone* the
6268 // load up into the predecessors so that we have a load of a gep of an alloca,
6269 // which can usually all be folded into the load.
6270 if (AllBasePointersAreAllocas)
6273 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
6274 // that is variable.
6275 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
6277 bool HasAnyPHIs = false;
6278 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
6279 if (FixedOperands[i]) continue; // operand doesn't need a phi.
6280 Value *FirstOp = FirstInst->getOperand(i);
6281 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
6282 FirstOp->getName()+".pn");
6283 InsertNewInstBefore(NewPN, PN);
6285 NewPN->reserveOperandSpace(e);
6286 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
6287 OperandPhis[i] = NewPN;
6288 FixedOperands[i] = NewPN;
6293 // Add all operands to the new PHIs.
6295 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6296 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
6297 BasicBlock *InBB = PN.getIncomingBlock(i);
6299 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
6300 if (PHINode *OpPhi = OperandPhis[op])
6301 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
6305 Value *Base = FixedOperands[0];
6306 return cast<GEPOperator>(FirstInst)->isInBounds() ?
6307 GetElementPtrInst::CreateInBounds(Base, FixedOperands.begin()+1,
6308 FixedOperands.end()) :
6309 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
6310 FixedOperands.end());
6314 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
6315 /// sink the load out of the block that defines it. This means that it must be
6316 /// obvious the value of the load is not changed from the point of the load to
6317 /// the end of the block it is in.
6319 /// Finally, it is safe, but not profitable, to sink a load targetting a
6320 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
6322 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
6323 BasicBlock::iterator BBI = L, E = L->getParent()->end();
6325 for (++BBI; BBI != E; ++BBI)
6326 if (BBI->mayWriteToMemory())
6329 // Check for non-address taken alloca. If not address-taken already, it isn't
6330 // profitable to do this xform.
6331 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
6332 bool isAddressTaken = false;
6333 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
6335 if (isa<LoadInst>(UI)) continue;
6336 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
6337 // If storing TO the alloca, then the address isn't taken.
6338 if (SI->getOperand(1) == AI) continue;
6340 isAddressTaken = true;
6344 if (!isAddressTaken && AI->isStaticAlloca())
6348 // If this load is a load from a GEP with a constant offset from an alloca,
6349 // then we don't want to sink it. In its present form, it will be
6350 // load [constant stack offset]. Sinking it will cause us to have to
6351 // materialize the stack addresses in each predecessor in a register only to
6352 // do a shared load from register in the successor.
6353 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
6354 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
6355 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
6361 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) {
6362 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
6364 // When processing loads, we need to propagate two bits of information to the
6365 // sunk load: whether it is volatile, and what its alignment is. We currently
6366 // don't sink loads when some have their alignment specified and some don't.
6367 // visitLoadInst will propagate an alignment onto the load when TD is around,
6368 // and if TD isn't around, we can't handle the mixed case.
6369 bool isVolatile = FirstLI->isVolatile();
6370 unsigned LoadAlignment = FirstLI->getAlignment();
6372 // We can't sink the load if the loaded value could be modified between the
6373 // load and the PHI.
6374 if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
6375 !isSafeAndProfitableToSinkLoad(FirstLI))
6378 // If the PHI is of volatile loads and the load block has multiple
6379 // successors, sinking it would remove a load of the volatile value from
6380 // the path through the other successor.
6382 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
6385 // Check to see if all arguments are the same operation.
6386 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6387 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
6388 if (!LI || !LI->hasOneUse())
6391 // We can't sink the load if the loaded value could be modified between
6392 // the load and the PHI.
6393 if (LI->isVolatile() != isVolatile ||
6394 LI->getParent() != PN.getIncomingBlock(i) ||
6395 !isSafeAndProfitableToSinkLoad(LI))
6398 // If some of the loads have an alignment specified but not all of them,
6399 // we can't do the transformation.
6400 if ((LoadAlignment != 0) != (LI->getAlignment() != 0))
6403 LoadAlignment = std::min(LoadAlignment, LI->getAlignment());
6405 // If the PHI is of volatile loads and the load block has multiple
6406 // successors, sinking it would remove a load of the volatile value from
6407 // the path through the other successor.
6409 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
6413 // Okay, they are all the same operation. Create a new PHI node of the
6414 // correct type, and PHI together all of the LHS's of the instructions.
6415 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
6416 PN.getName()+".in");
6417 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
6419 Value *InVal = FirstLI->getOperand(0);
6420 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
6422 // Add all operands to the new PHI.
6423 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6424 Value *NewInVal = cast<LoadInst>(PN.getIncomingValue(i))->getOperand(0);
6425 if (NewInVal != InVal)
6427 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
6432 // The new PHI unions all of the same values together. This is really
6433 // common, so we handle it intelligently here for compile-time speed.
6437 InsertNewInstBefore(NewPN, PN);
6441 // If this was a volatile load that we are merging, make sure to loop through
6442 // and mark all the input loads as non-volatile. If we don't do this, we will
6443 // insert a new volatile load and the old ones will not be deletable.
6445 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
6446 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
6448 return new LoadInst(PhiVal, "", isVolatile, LoadAlignment);
6453 /// FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
6454 /// operator and they all are only used by the PHI, PHI together their
6455 /// inputs, and do the operation once, to the result of the PHI.
6456 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
6457 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
6459 if (isa<GetElementPtrInst>(FirstInst))
6460 return FoldPHIArgGEPIntoPHI(PN);
6461 if (isa<LoadInst>(FirstInst))
6462 return FoldPHIArgLoadIntoPHI(PN);
6464 // Scan the instruction, looking for input operations that can be folded away.
6465 // If all input operands to the phi are the same instruction (e.g. a cast from
6466 // the same type or "+42") we can pull the operation through the PHI, reducing
6467 // code size and simplifying code.
6468 Constant *ConstantOp = 0;
6469 const Type *CastSrcTy = 0;
6471 if (isa<CastInst>(FirstInst)) {
6472 CastSrcTy = FirstInst->getOperand(0)->getType();
6474 // Be careful about transforming integer PHIs. We don't want to pessimize
6475 // the code by turning an i32 into an i1293.
6476 if (isa<IntegerType>(PN.getType()) && isa<IntegerType>(CastSrcTy)) {
6477 if (!ShouldChangeType(PN.getType(), CastSrcTy))
6480 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
6481 // Can fold binop, compare or shift here if the RHS is a constant,
6482 // otherwise call FoldPHIArgBinOpIntoPHI.
6483 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
6484 if (ConstantOp == 0)
6485 return FoldPHIArgBinOpIntoPHI(PN);
6487 return 0; // Cannot fold this operation.
6490 // Check to see if all arguments are the same operation.
6491 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6492 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
6493 if (I == 0 || !I->hasOneUse() || !I->isSameOperationAs(FirstInst))
6496 if (I->getOperand(0)->getType() != CastSrcTy)
6497 return 0; // Cast operation must match.
6498 } else if (I->getOperand(1) != ConstantOp) {
6503 // Okay, they are all the same operation. Create a new PHI node of the
6504 // correct type, and PHI together all of the LHS's of the instructions.
6505 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
6506 PN.getName()+".in");
6507 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
6509 Value *InVal = FirstInst->getOperand(0);
6510 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
6512 // Add all operands to the new PHI.
6513 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
6514 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
6515 if (NewInVal != InVal)
6517 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
6522 // The new PHI unions all of the same values together. This is really
6523 // common, so we handle it intelligently here for compile-time speed.
6527 InsertNewInstBefore(NewPN, PN);
6531 // Insert and return the new operation.
6532 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst))
6533 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
6535 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
6536 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
6538 CmpInst *CIOp = cast<CmpInst>(FirstInst);
6539 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
6540 PhiVal, ConstantOp);
6543 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
6545 static bool DeadPHICycle(PHINode *PN,
6546 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
6547 if (PN->use_empty()) return true;
6548 if (!PN->hasOneUse()) return false;
6550 // Remember this node, and if we find the cycle, return.
6551 if (!PotentiallyDeadPHIs.insert(PN))
6554 // Don't scan crazily complex things.
6555 if (PotentiallyDeadPHIs.size() == 16)
6558 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
6559 return DeadPHICycle(PU, PotentiallyDeadPHIs);
6564 /// PHIsEqualValue - Return true if this phi node is always equal to
6565 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
6566 /// z = some value; x = phi (y, z); y = phi (x, z)
6567 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
6568 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
6569 // See if we already saw this PHI node.
6570 if (!ValueEqualPHIs.insert(PN))
6573 // Don't scan crazily complex things.
6574 if (ValueEqualPHIs.size() == 16)
6577 // Scan the operands to see if they are either phi nodes or are equal to
6579 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
6580 Value *Op = PN->getIncomingValue(i);
6581 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
6582 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
6584 } else if (Op != NonPhiInVal)
6593 struct PHIUsageRecord {
6594 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on)
6595 unsigned Shift; // The amount shifted.
6596 Instruction *Inst; // The trunc instruction.
6598 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
6599 : PHIId(pn), Shift(Sh), Inst(User) {}
6601 bool operator<(const PHIUsageRecord &RHS) const {
6602 if (PHIId < RHS.PHIId) return true;
6603 if (PHIId > RHS.PHIId) return false;
6604 if (Shift < RHS.Shift) return true;
6605 if (Shift > RHS.Shift) return false;
6606 return Inst->getType()->getPrimitiveSizeInBits() <
6607 RHS.Inst->getType()->getPrimitiveSizeInBits();
6611 struct LoweredPHIRecord {
6612 PHINode *PN; // The PHI that was lowered.
6613 unsigned Shift; // The amount shifted.
6614 unsigned Width; // The width extracted.
6616 LoweredPHIRecord(PHINode *pn, unsigned Sh, const Type *Ty)
6617 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
6619 // Ctor form used by DenseMap.
6620 LoweredPHIRecord(PHINode *pn, unsigned Sh)
6621 : PN(pn), Shift(Sh), Width(0) {}
6627 struct DenseMapInfo<LoweredPHIRecord> {
6628 static inline LoweredPHIRecord getEmptyKey() {
6629 return LoweredPHIRecord(0, 0);
6631 static inline LoweredPHIRecord getTombstoneKey() {
6632 return LoweredPHIRecord(0, 1);
6634 static unsigned getHashValue(const LoweredPHIRecord &Val) {
6635 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
6638 static bool isEqual(const LoweredPHIRecord &LHS,
6639 const LoweredPHIRecord &RHS) {
6640 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
6641 LHS.Width == RHS.Width;
6645 struct isPodLike<LoweredPHIRecord> { static const bool value = true; };
6649 /// SliceUpIllegalIntegerPHI - This is an integer PHI and we know that it has an
6650 /// illegal type: see if it is only used by trunc or trunc(lshr) operations. If
6651 /// so, we split the PHI into the various pieces being extracted. This sort of
6652 /// thing is introduced when SROA promotes an aggregate to large integer values.
6654 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
6655 /// inttoptr. We should produce new PHIs in the right type.
6657 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
6658 // PHIUsers - Keep track of all of the truncated values extracted from a set
6659 // of PHIs, along with their offset. These are the things we want to rewrite.
6660 SmallVector<PHIUsageRecord, 16> PHIUsers;
6662 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
6663 // nodes which are extracted from. PHIsToSlice is a set we use to avoid
6664 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
6665 // check the uses of (to ensure they are all extracts).
6666 SmallVector<PHINode*, 8> PHIsToSlice;
6667 SmallPtrSet<PHINode*, 8> PHIsInspected;
6669 PHIsToSlice.push_back(&FirstPhi);
6670 PHIsInspected.insert(&FirstPhi);
6672 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
6673 PHINode *PN = PHIsToSlice[PHIId];
6675 // Scan the input list of the PHI. If any input is an invoke, and if the
6676 // input is defined in the predecessor, then we won't be split the critical
6677 // edge which is required to insert a truncate. Because of this, we have to
6679 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
6680 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
6681 if (II == 0) continue;
6682 if (II->getParent() != PN->getIncomingBlock(i))
6685 // If we have a phi, and if it's directly in the predecessor, then we have
6686 // a critical edge where we need to put the truncate. Since we can't
6687 // split the edge in instcombine, we have to bail out.
6692 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
6694 Instruction *User = cast<Instruction>(*UI);
6696 // If the user is a PHI, inspect its uses recursively.
6697 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
6698 if (PHIsInspected.insert(UserPN))
6699 PHIsToSlice.push_back(UserPN);
6703 // Truncates are always ok.
6704 if (isa<TruncInst>(User)) {
6705 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, User));
6709 // Otherwise it must be a lshr which can only be used by one trunc.
6710 if (User->getOpcode() != Instruction::LShr ||
6711 !User->hasOneUse() || !isa<TruncInst>(User->use_back()) ||
6712 !isa<ConstantInt>(User->getOperand(1)))
6715 unsigned Shift = cast<ConstantInt>(User->getOperand(1))->getZExtValue();
6716 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, User->use_back()));
6720 // If we have no users, they must be all self uses, just nuke the PHI.
6721 if (PHIUsers.empty())
6722 return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
6724 // If this phi node is transformable, create new PHIs for all the pieces
6725 // extracted out of it. First, sort the users by their offset and size.
6726 array_pod_sort(PHIUsers.begin(), PHIUsers.end());
6728 DEBUG(errs() << "SLICING UP PHI: " << FirstPhi << '\n';
6729 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
6730 errs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] <<'\n';
6733 // PredValues - This is a temporary used when rewriting PHI nodes. It is
6734 // hoisted out here to avoid construction/destruction thrashing.
6735 DenseMap<BasicBlock*, Value*> PredValues;
6737 // ExtractedVals - Each new PHI we introduce is saved here so we don't
6738 // introduce redundant PHIs.
6739 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
6741 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
6742 unsigned PHIId = PHIUsers[UserI].PHIId;
6743 PHINode *PN = PHIsToSlice[PHIId];
6744 unsigned Offset = PHIUsers[UserI].Shift;
6745 const Type *Ty = PHIUsers[UserI].Inst->getType();
6749 // If we've already lowered a user like this, reuse the previously lowered
6751 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == 0) {
6753 // Otherwise, Create the new PHI node for this user.
6754 EltPHI = PHINode::Create(Ty, PN->getName()+".off"+Twine(Offset), PN);
6755 assert(EltPHI->getType() != PN->getType() &&
6756 "Truncate didn't shrink phi?");
6758 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
6759 BasicBlock *Pred = PN->getIncomingBlock(i);
6760 Value *&PredVal = PredValues[Pred];
6762 // If we already have a value for this predecessor, reuse it.
6764 EltPHI->addIncoming(PredVal, Pred);
6768 // Handle the PHI self-reuse case.
6769 Value *InVal = PN->getIncomingValue(i);
6772 EltPHI->addIncoming(PredVal, Pred);
6776 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
6777 // If the incoming value was a PHI, and if it was one of the PHIs we
6778 // already rewrote it, just use the lowered value.
6779 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
6781 EltPHI->addIncoming(PredVal, Pred);
6786 // Otherwise, do an extract in the predecessor.
6787 Builder->SetInsertPoint(Pred, Pred->getTerminator());
6790 Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(),
6791 Offset), "extract");
6792 Res = Builder->CreateTrunc(Res, Ty, "extract.t");
6794 EltPHI->addIncoming(Res, Pred);
6796 // If the incoming value was a PHI, and if it was one of the PHIs we are
6797 // rewriting, we will ultimately delete the code we inserted. This
6798 // means we need to revisit that PHI to make sure we extract out the
6800 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
6801 if (PHIsInspected.count(OldInVal)) {
6802 unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(),
6803 OldInVal)-PHIsToSlice.begin();
6804 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
6805 cast<Instruction>(Res)));
6811 DEBUG(errs() << " Made element PHI for offset " << Offset << ": "
6812 << *EltPHI << '\n');
6813 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
6816 // Replace the use of this piece with the PHI node.
6817 ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
6820 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
6822 Value *Undef = UndefValue::get(FirstPhi.getType());
6823 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
6824 ReplaceInstUsesWith(*PHIsToSlice[i], Undef);
6825 return ReplaceInstUsesWith(FirstPhi, Undef);
6828 // PHINode simplification
6830 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
6831 // If LCSSA is around, don't mess with Phi nodes
6832 if (MustPreserveLCSSA) return 0;
6834 if (Value *V = PN.hasConstantValue())
6835 return ReplaceInstUsesWith(PN, V);
6837 // If all PHI operands are the same operation, pull them through the PHI,
6838 // reducing code size.
6839 if (isa<Instruction>(PN.getIncomingValue(0)) &&
6840 isa<Instruction>(PN.getIncomingValue(1)) &&
6841 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
6842 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
6843 // FIXME: The hasOneUse check will fail for PHIs that use the value more
6844 // than themselves more than once.
6845 PN.getIncomingValue(0)->hasOneUse())
6846 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
6849 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
6850 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
6851 // PHI)... break the cycle.
6852 if (PN.hasOneUse()) {
6853 Instruction *PHIUser = cast<Instruction>(PN.use_back());
6854 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
6855 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
6856 PotentiallyDeadPHIs.insert(&PN);
6857 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
6858 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
6861 // If this phi has a single use, and if that use just computes a value for
6862 // the next iteration of a loop, delete the phi. This occurs with unused
6863 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
6864 // common case here is good because the only other things that catch this
6865 // are induction variable analysis (sometimes) and ADCE, which is only run
6867 if (PHIUser->hasOneUse() &&
6868 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
6869 PHIUser->use_back() == &PN) {
6870 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
6874 // We sometimes end up with phi cycles that non-obviously end up being the
6875 // same value, for example:
6876 // z = some value; x = phi (y, z); y = phi (x, z)
6877 // where the phi nodes don't necessarily need to be in the same block. Do a
6878 // quick check to see if the PHI node only contains a single non-phi value, if
6879 // so, scan to see if the phi cycle is actually equal to that value.
6881 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
6882 // Scan for the first non-phi operand.
6883 while (InValNo != NumOperandVals &&
6884 isa<PHINode>(PN.getIncomingValue(InValNo)))
6887 if (InValNo != NumOperandVals) {
6888 Value *NonPhiInVal = PN.getOperand(InValNo);
6890 // Scan the rest of the operands to see if there are any conflicts, if so
6891 // there is no need to recursively scan other phis.
6892 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
6893 Value *OpVal = PN.getIncomingValue(InValNo);
6894 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
6898 // If we scanned over all operands, then we have one unique value plus
6899 // phi values. Scan PHI nodes to see if they all merge in each other or
6901 if (InValNo == NumOperandVals) {
6902 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
6903 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
6904 return ReplaceInstUsesWith(PN, NonPhiInVal);
6909 // If there are multiple PHIs, sort their operands so that they all list
6910 // the blocks in the same order. This will help identical PHIs be eliminated
6911 // by other passes. Other passes shouldn't depend on this for correctness
6913 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
6915 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
6916 BasicBlock *BBA = PN.getIncomingBlock(i);
6917 BasicBlock *BBB = FirstPN->getIncomingBlock(i);
6919 Value *VA = PN.getIncomingValue(i);
6920 unsigned j = PN.getBasicBlockIndex(BBB);
6921 Value *VB = PN.getIncomingValue(j);
6922 PN.setIncomingBlock(i, BBB);
6923 PN.setIncomingValue(i, VB);
6924 PN.setIncomingBlock(j, BBA);
6925 PN.setIncomingValue(j, VA);
6926 // NOTE: Instcombine normally would want us to "return &PN" if we
6927 // modified any of the operands of an instruction. However, since we
6928 // aren't adding or removing uses (just rearranging them) we don't do
6929 // this in this case.
6933 // If this is an integer PHI and we know that it has an illegal type, see if
6934 // it is only used by trunc or trunc(lshr) operations. If so, we split the
6935 // PHI into the various pieces being extracted. This sort of thing is
6936 // introduced when SROA promotes an aggregate to a single large integer type.
6937 if (isa<IntegerType>(PN.getType()) && TD &&
6938 !TD->isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
6939 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))
6945 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
6946 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
6948 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD))
6949 return ReplaceInstUsesWith(GEP, V);
6951 Value *PtrOp = GEP.getOperand(0);
6953 if (isa<UndefValue>(GEP.getOperand(0)))
6954 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
6956 // Eliminate unneeded casts for indices.
6958 bool MadeChange = false;
6959 unsigned PtrSize = TD->getPointerSizeInBits();
6961 gep_type_iterator GTI = gep_type_begin(GEP);
6962 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
6963 I != E; ++I, ++GTI) {
6964 if (!isa<SequentialType>(*GTI)) continue;
6966 // If we are using a wider index than needed for this platform, shrink it
6967 // to what we need. If narrower, sign-extend it to what we need. This
6968 // explicit cast can make subsequent optimizations more obvious.
6969 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
6970 if (OpBits == PtrSize)
6973 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true);
6976 if (MadeChange) return &GEP;
6979 // Combine Indices - If the source pointer to this getelementptr instruction
6980 // is a getelementptr instruction, combine the indices of the two
6981 // getelementptr instructions into a single instruction.
6983 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
6984 // Note that if our source is a gep chain itself that we wait for that
6985 // chain to be resolved before we perform this transformation. This
6986 // avoids us creating a TON of code in some cases.
6988 if (GetElementPtrInst *SrcGEP =
6989 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
6990 if (SrcGEP->getNumOperands() == 2)
6991 return 0; // Wait until our source is folded to completion.
6993 SmallVector<Value*, 8> Indices;
6995 // Find out whether the last index in the source GEP is a sequential idx.
6996 bool EndsWithSequential = false;
6997 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
6999 EndsWithSequential = !isa<StructType>(*I);
7001 // Can we combine the two pointer arithmetics offsets?
7002 if (EndsWithSequential) {
7003 // Replace: gep (gep %P, long B), long A, ...
7004 // With: T = long A+B; gep %P, T, ...
7007 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
7008 Value *GO1 = GEP.getOperand(1);
7009 if (SO1 == Constant::getNullValue(SO1->getType())) {
7011 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7014 // If they aren't the same type, then the input hasn't been processed
7015 // by the loop above yet (which canonicalizes sequential index types to
7016 // intptr_t). Just avoid transforming this until the input has been
7018 if (SO1->getType() != GO1->getType())
7020 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
7023 // Update the GEP in place if possible.
7024 if (Src->getNumOperands() == 2) {
7025 GEP.setOperand(0, Src->getOperand(0));
7026 GEP.setOperand(1, Sum);
7029 Indices.append(Src->op_begin()+1, Src->op_end()-1);
7030 Indices.push_back(Sum);
7031 Indices.append(GEP.op_begin()+2, GEP.op_end());
7032 } else if (isa<Constant>(*GEP.idx_begin()) &&
7033 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7034 Src->getNumOperands() != 1) {
7035 // Otherwise we can do the fold if the first index of the GEP is a zero
7036 Indices.append(Src->op_begin()+1, Src->op_end());
7037 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
7040 if (!Indices.empty())
7041 return (cast<GEPOperator>(&GEP)->isInBounds() &&
7042 Src->isInBounds()) ?
7043 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
7044 Indices.end(), GEP.getName()) :
7045 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
7046 Indices.end(), GEP.getName());
7049 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
7050 if (Value *X = getBitCastOperand(PtrOp)) {
7051 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
7053 // If the input bitcast is actually "bitcast(bitcast(x))", then we don't
7054 // want to change the gep until the bitcasts are eliminated.
7055 if (getBitCastOperand(X)) {
7056 Worklist.AddValue(PtrOp);
7060 bool HasZeroPointerIndex = false;
7061 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
7062 HasZeroPointerIndex = C->isZero();
7064 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
7065 // into : GEP [10 x i8]* X, i32 0, ...
7067 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
7068 // into : GEP i8* X, ...
7070 // This occurs when the program declares an array extern like "int X[];"
7071 if (HasZeroPointerIndex) {
7072 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7073 const PointerType *XTy = cast<PointerType>(X->getType());
7074 if (const ArrayType *CATy =
7075 dyn_cast<ArrayType>(CPTy->getElementType())) {
7076 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
7077 if (CATy->getElementType() == XTy->getElementType()) {
7078 // -> GEP i8* X, ...
7079 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
7080 return cast<GEPOperator>(&GEP)->isInBounds() ?
7081 GetElementPtrInst::CreateInBounds(X, Indices.begin(), Indices.end(),
7083 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
7087 if (const ArrayType *XATy = dyn_cast<ArrayType>(XTy->getElementType())){
7088 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
7089 if (CATy->getElementType() == XATy->getElementType()) {
7090 // -> GEP [10 x i8]* X, i32 0, ...
7091 // At this point, we know that the cast source type is a pointer
7092 // to an array of the same type as the destination pointer
7093 // array. Because the array type is never stepped over (there
7094 // is a leading zero) we can fold the cast into this GEP.
7095 GEP.setOperand(0, X);
7100 } else if (GEP.getNumOperands() == 2) {
7101 // Transform things like:
7102 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
7103 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
7104 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7105 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7106 if (TD && isa<ArrayType>(SrcElTy) &&
7107 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7108 TD->getTypeAllocSize(ResElTy)) {
7110 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
7111 Idx[1] = GEP.getOperand(1);
7112 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
7113 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
7114 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
7115 // V and GEP are both pointer types --> BitCast
7116 return new BitCastInst(NewGEP, GEP.getType());
7119 // Transform things like:
7120 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
7121 // (where tmp = 8*tmp2) into:
7122 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
7124 if (TD && isa<ArrayType>(SrcElTy) &&
7125 ResElTy == Type::getInt8Ty(GEP.getContext())) {
7126 uint64_t ArrayEltSize =
7127 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
7129 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7130 // allow either a mul, shift, or constant here.
7132 ConstantInt *Scale = 0;
7133 if (ArrayEltSize == 1) {
7134 NewIdx = GEP.getOperand(1);
7135 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
7136 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7137 NewIdx = ConstantInt::get(CI->getType(), 1);
7139 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7140 if (Inst->getOpcode() == Instruction::Shl &&
7141 isa<ConstantInt>(Inst->getOperand(1))) {
7142 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
7143 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
7144 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
7146 NewIdx = Inst->getOperand(0);
7147 } else if (Inst->getOpcode() == Instruction::Mul &&
7148 isa<ConstantInt>(Inst->getOperand(1))) {
7149 Scale = cast<ConstantInt>(Inst->getOperand(1));
7150 NewIdx = Inst->getOperand(0);
7154 // If the index will be to exactly the right offset with the scale taken
7155 // out, perform the transformation. Note, we don't know whether Scale is
7156 // signed or not. We'll use unsigned version of division/modulo
7157 // operation after making sure Scale doesn't have the sign bit set.
7158 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
7159 Scale->getZExtValue() % ArrayEltSize == 0) {
7160 Scale = ConstantInt::get(Scale->getType(),
7161 Scale->getZExtValue() / ArrayEltSize);
7162 if (Scale->getZExtValue() != 1) {
7163 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7165 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
7168 // Insert the new GEP instruction.
7170 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
7172 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
7173 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
7174 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
7175 // The NewGEP must be pointer typed, so must the old one -> BitCast
7176 return new BitCastInst(NewGEP, GEP.getType());
7182 /// See if we can simplify:
7183 /// X = bitcast A* to B*
7184 /// Y = gep X, <...constant indices...>
7185 /// into a gep of the original struct. This is important for SROA and alias
7186 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
7187 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
7189 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
7190 // Determine how much the GEP moves the pointer. We are guaranteed to get
7191 // a constant back from EmitGEPOffset.
7192 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP));
7193 int64_t Offset = OffsetV->getSExtValue();
7195 // If this GEP instruction doesn't move the pointer, just replace the GEP
7196 // with a bitcast of the real input to the dest type.
7198 // If the bitcast is of an allocation, and the allocation will be
7199 // converted to match the type of the cast, don't touch this.
7200 if (isa<AllocaInst>(BCI->getOperand(0)) ||
7201 isMalloc(BCI->getOperand(0))) {
7202 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
7203 if (Instruction *I = visitBitCast(*BCI)) {
7206 BCI->getParent()->getInstList().insert(BCI, I);
7207 ReplaceInstUsesWith(*BCI, I);
7212 return new BitCastInst(BCI->getOperand(0), GEP.getType());
7215 // Otherwise, if the offset is non-zero, we need to find out if there is a
7216 // field at Offset in 'A's type. If so, we can pull the cast through the
7218 SmallVector<Value*, 8> NewIndices;
7220 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
7221 if (FindElementAtOffset(InTy, Offset, NewIndices)) {
7222 Value *NGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
7223 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
7225 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
7228 if (NGEP->getType() == GEP.getType())
7229 return ReplaceInstUsesWith(GEP, NGEP);
7230 NGEP->takeName(&GEP);
7231 return new BitCastInst(NGEP, GEP.getType());
7239 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
7240 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
7241 if (AI.isArrayAllocation()) { // Check C != 1
7242 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7244 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7245 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7246 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
7247 New->setAlignment(AI.getAlignment());
7249 // Scan to the end of the allocation instructions, to skip over a block of
7250 // allocas if possible...also skip interleaved debug info
7252 BasicBlock::iterator It = New;
7253 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
7255 // Now that I is pointing to the first non-allocation-inst in the block,
7256 // insert our getelementptr instruction...
7258 Value *NullIdx =Constant::getNullValue(Type::getInt32Ty(AI.getContext()));
7262 Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
7263 New->getName()+".sub", It);
7265 // Now make everything use the getelementptr instead of the original
7267 return ReplaceInstUsesWith(AI, V);
7268 } else if (isa<UndefValue>(AI.getArraySize())) {
7269 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7273 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
7274 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7275 // Note that we only do this for alloca's, because malloc should allocate
7276 // and return a unique pointer, even for a zero byte allocation.
7277 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
7278 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7280 // If the alignment is 0 (unspecified), assign it the preferred alignment.
7281 if (AI.getAlignment() == 0)
7282 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
7288 Instruction *InstCombiner::visitFree(Instruction &FI) {
7289 Value *Op = FI.getOperand(1);
7291 // free undef -> unreachable.
7292 if (isa<UndefValue>(Op)) {
7293 // Insert a new store to null because we cannot modify the CFG here.
7294 new StoreInst(ConstantInt::getTrue(FI.getContext()),
7295 UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI);
7296 return EraseInstFromFunction(FI);
7299 // If we have 'free null' delete the instruction. This can happen in stl code
7300 // when lots of inlining happens.
7301 if (isa<ConstantPointerNull>(Op))
7302 return EraseInstFromFunction(FI);
7304 // If we have a malloc call whose only use is a free call, delete both.
7306 if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
7307 if (Op->hasOneUse() && CI->hasOneUse()) {
7308 EraseInstFromFunction(FI);
7309 EraseInstFromFunction(*CI);
7310 return EraseInstFromFunction(*cast<Instruction>(Op));
7313 // Op is a call to malloc
7314 if (Op->hasOneUse()) {
7315 EraseInstFromFunction(FI);
7316 return EraseInstFromFunction(*cast<Instruction>(Op));
7324 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
7325 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
7326 const TargetData *TD) {
7327 User *CI = cast<User>(LI.getOperand(0));
7328 Value *CastOp = CI->getOperand(0);
7330 const PointerType *DestTy = cast<PointerType>(CI->getType());
7331 const Type *DestPTy = DestTy->getElementType();
7332 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7334 // If the address spaces don't match, don't eliminate the cast.
7335 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
7338 const Type *SrcPTy = SrcTy->getElementType();
7340 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
7341 isa<VectorType>(DestPTy)) {
7342 // If the source is an array, the code below will not succeed. Check to
7343 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7345 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7346 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7347 if (ASrcTy->getNumElements() != 0) {
7349 Idxs[0] = Constant::getNullValue(Type::getInt32Ty(LI.getContext()));
7351 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
7352 SrcTy = cast<PointerType>(CastOp->getType());
7353 SrcPTy = SrcTy->getElementType();
7356 if (IC.getTargetData() &&
7357 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
7358 isa<VectorType>(SrcPTy)) &&
7359 // Do not allow turning this into a load of an integer, which is then
7360 // casted to a pointer, this pessimizes pointer analysis a lot.
7361 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
7362 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
7363 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
7365 // Okay, we are casting from one integer or pointer type to another of
7366 // the same size. Instead of casting the pointer before the load, cast
7367 // the result of the loaded value.
7369 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
7370 // Now cast the result of the load.
7371 return new BitCastInst(NewLoad, LI.getType());
7378 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
7379 Value *Op = LI.getOperand(0);
7381 // Attempt to improve the alignment.
7383 unsigned KnownAlign =
7384 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
7386 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
7388 LI.setAlignment(KnownAlign);
7391 // load (cast X) --> cast (load X) iff safe.
7392 if (isa<CastInst>(Op))
7393 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
7396 // None of the following transforms are legal for volatile loads.
7397 if (LI.isVolatile()) return 0;
7399 // Do really simple store-to-load forwarding and load CSE, to catch cases
7400 // where there are several consequtive memory accesses to the same location,
7401 // separated by a few arithmetic operations.
7402 BasicBlock::iterator BBI = &LI;
7403 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
7404 return ReplaceInstUsesWith(LI, AvailableVal);
7406 // load(gep null, ...) -> unreachable
7407 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
7408 const Value *GEPI0 = GEPI->getOperand(0);
7409 // TODO: Consider a target hook for valid address spaces for this xform.
7410 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
7411 // Insert a new store to null instruction before the load to indicate
7412 // that this code is not reachable. We do this instead of inserting
7413 // an unreachable instruction directly because we cannot modify the
7415 new StoreInst(UndefValue::get(LI.getType()),
7416 Constant::getNullValue(Op->getType()), &LI);
7417 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7421 // load null/undef -> unreachable
7422 // TODO: Consider a target hook for valid address spaces for this xform.
7423 if (isa<UndefValue>(Op) ||
7424 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
7425 // Insert a new store to null instruction before the load to indicate that
7426 // this code is not reachable. We do this instead of inserting an
7427 // unreachable instruction directly because we cannot modify the CFG.
7428 new StoreInst(UndefValue::get(LI.getType()),
7429 Constant::getNullValue(Op->getType()), &LI);
7430 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7433 // Instcombine load (constantexpr_cast global) -> cast (load global)
7434 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
7436 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
7439 if (Op->hasOneUse()) {
7440 // Change select and PHI nodes to select values instead of addresses: this
7441 // helps alias analysis out a lot, allows many others simplifications, and
7442 // exposes redundancy in the code.
7444 // Note that we cannot do the transformation unless we know that the
7445 // introduced loads cannot trap! Something like this is valid as long as
7446 // the condition is always false: load (select bool %C, int* null, int* %G),
7447 // but it would not be valid if we transformed it to load from null
7450 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
7451 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
7452 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
7453 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
7454 Value *V1 = Builder->CreateLoad(SI->getOperand(1),
7455 SI->getOperand(1)->getName()+".val");
7456 Value *V2 = Builder->CreateLoad(SI->getOperand(2),
7457 SI->getOperand(2)->getName()+".val");
7458 return SelectInst::Create(SI->getCondition(), V1, V2);
7461 // load (select (cond, null, P)) -> load P
7462 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
7463 if (C->isNullValue()) {
7464 LI.setOperand(0, SI->getOperand(2));
7468 // load (select (cond, P, null)) -> load P
7469 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
7470 if (C->isNullValue()) {
7471 LI.setOperand(0, SI->getOperand(1));
7479 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
7480 /// when possible. This makes it generally easy to do alias analysis and/or
7481 /// SROA/mem2reg of the memory object.
7482 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
7483 User *CI = cast<User>(SI.getOperand(1));
7484 Value *CastOp = CI->getOperand(0);
7486 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7487 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
7488 if (SrcTy == 0) return 0;
7490 const Type *SrcPTy = SrcTy->getElementType();
7492 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
7495 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
7496 /// to its first element. This allows us to handle things like:
7497 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
7498 /// on 32-bit hosts.
7499 SmallVector<Value*, 4> NewGEPIndices;
7501 // If the source is an array, the code below will not succeed. Check to
7502 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7504 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
7505 // Index through pointer.
7506 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(SI.getContext()));
7507 NewGEPIndices.push_back(Zero);
7510 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
7511 if (!STy->getNumElements()) /* Struct can be empty {} */
7513 NewGEPIndices.push_back(Zero);
7514 SrcPTy = STy->getElementType(0);
7515 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
7516 NewGEPIndices.push_back(Zero);
7517 SrcPTy = ATy->getElementType();
7523 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
7526 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
7529 // If the pointers point into different address spaces or if they point to
7530 // values with different sizes, we can't do the transformation.
7531 if (!IC.getTargetData() ||
7532 SrcTy->getAddressSpace() !=
7533 cast<PointerType>(CI->getType())->getAddressSpace() ||
7534 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
7535 IC.getTargetData()->getTypeSizeInBits(DestPTy))
7538 // Okay, we are casting from one integer or pointer type to another of
7539 // the same size. Instead of casting the pointer before
7540 // the store, cast the value to be stored.
7542 Value *SIOp0 = SI.getOperand(0);
7543 Instruction::CastOps opcode = Instruction::BitCast;
7544 const Type* CastSrcTy = SIOp0->getType();
7545 const Type* CastDstTy = SrcPTy;
7546 if (isa<PointerType>(CastDstTy)) {
7547 if (CastSrcTy->isInteger())
7548 opcode = Instruction::IntToPtr;
7549 } else if (isa<IntegerType>(CastDstTy)) {
7550 if (isa<PointerType>(SIOp0->getType()))
7551 opcode = Instruction::PtrToInt;
7554 // SIOp0 is a pointer to aggregate and this is a store to the first field,
7555 // emit a GEP to index into its first field.
7556 if (!NewGEPIndices.empty())
7557 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
7558 NewGEPIndices.end());
7560 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
7561 SIOp0->getName()+".c");
7562 return new StoreInst(NewCast, CastOp);
7565 /// equivalentAddressValues - Test if A and B will obviously have the same
7566 /// value. This includes recognizing that %t0 and %t1 will have the same
7567 /// value in code like this:
7568 /// %t0 = getelementptr \@a, 0, 3
7569 /// store i32 0, i32* %t0
7570 /// %t1 = getelementptr \@a, 0, 3
7571 /// %t2 = load i32* %t1
7573 static bool equivalentAddressValues(Value *A, Value *B) {
7574 // Test if the values are trivially equivalent.
7575 if (A == B) return true;
7577 // Test if the values come form identical arithmetic instructions.
7578 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
7579 // its only used to compare two uses within the same basic block, which
7580 // means that they'll always either have the same value or one of them
7581 // will have an undefined value.
7582 if (isa<BinaryOperator>(A) ||
7585 isa<GetElementPtrInst>(A))
7586 if (Instruction *BI = dyn_cast<Instruction>(B))
7587 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
7590 // Otherwise they may not be equivalent.
7594 // If this instruction has two uses, one of which is a llvm.dbg.declare,
7595 // return the llvm.dbg.declare.
7596 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
7597 if (!V->hasNUses(2))
7599 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
7601 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
7603 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
7604 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
7611 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
7612 Value *Val = SI.getOperand(0);
7613 Value *Ptr = SI.getOperand(1);
7615 // If the RHS is an alloca with a single use, zapify the store, making the
7617 // If the RHS is an alloca with a two uses, the other one being a
7618 // llvm.dbg.declare, zapify the store and the declare, making the
7619 // alloca dead. We must do this to prevent declare's from affecting
7621 if (!SI.isVolatile()) {
7622 if (Ptr->hasOneUse()) {
7623 if (isa<AllocaInst>(Ptr)) {
7624 EraseInstFromFunction(SI);
7628 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
7629 if (isa<AllocaInst>(GEP->getOperand(0))) {
7630 if (GEP->getOperand(0)->hasOneUse()) {
7631 EraseInstFromFunction(SI);
7635 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
7636 EraseInstFromFunction(*DI);
7637 EraseInstFromFunction(SI);
7644 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
7645 EraseInstFromFunction(*DI);
7646 EraseInstFromFunction(SI);
7652 // Attempt to improve the alignment.
7654 unsigned KnownAlign =
7655 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
7657 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
7659 SI.setAlignment(KnownAlign);
7662 // Do really simple DSE, to catch cases where there are several consecutive
7663 // stores to the same location, separated by a few arithmetic operations. This
7664 // situation often occurs with bitfield accesses.
7665 BasicBlock::iterator BBI = &SI;
7666 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
7669 // Don't count debug info directives, lest they affect codegen,
7670 // and we skip pointer-to-pointer bitcasts, which are NOPs.
7671 // It is necessary for correctness to skip those that feed into a
7672 // llvm.dbg.declare, as these are not present when debugging is off.
7673 if (isa<DbgInfoIntrinsic>(BBI) ||
7674 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
7679 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
7680 // Prev store isn't volatile, and stores to the same location?
7681 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
7682 SI.getOperand(1))) {
7685 EraseInstFromFunction(*PrevSI);
7691 // If this is a load, we have to stop. However, if the loaded value is from
7692 // the pointer we're loading and is producing the pointer we're storing,
7693 // then *this* store is dead (X = load P; store X -> P).
7694 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7695 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
7697 EraseInstFromFunction(SI);
7701 // Otherwise, this is a load from some other location. Stores before it
7706 // Don't skip over loads or things that can modify memory.
7707 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
7712 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
7714 // store X, null -> turns into 'unreachable' in SimplifyCFG
7715 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
7716 if (!isa<UndefValue>(Val)) {
7717 SI.setOperand(0, UndefValue::get(Val->getType()));
7718 if (Instruction *U = dyn_cast<Instruction>(Val))
7719 Worklist.Add(U); // Dropped a use.
7722 return 0; // Do not modify these!
7725 // store undef, Ptr -> noop
7726 if (isa<UndefValue>(Val)) {
7727 EraseInstFromFunction(SI);
7732 // If the pointer destination is a cast, see if we can fold the cast into the
7734 if (isa<CastInst>(Ptr))
7735 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7737 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
7739 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
7743 // If this store is the last instruction in the basic block (possibly
7744 // excepting debug info instructions and the pointer bitcasts that feed
7745 // into them), and if the block ends with an unconditional branch, try
7746 // to move it to the successor block.
7750 } while (isa<DbgInfoIntrinsic>(BBI) ||
7751 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
7752 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
7753 if (BI->isUnconditional())
7754 if (SimplifyStoreAtEndOfBlock(SI))
7755 return 0; // xform done!
7760 /// SimplifyStoreAtEndOfBlock - Turn things like:
7761 /// if () { *P = v1; } else { *P = v2 }
7762 /// into a phi node with a store in the successor.
7764 /// Simplify things like:
7765 /// *P = v1; if () { *P = v2; }
7766 /// into a phi node with a store in the successor.
7768 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
7769 BasicBlock *StoreBB = SI.getParent();
7771 // Check to see if the successor block has exactly two incoming edges. If
7772 // so, see if the other predecessor contains a store to the same location.
7773 // if so, insert a PHI node (if needed) and move the stores down.
7774 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
7776 // Determine whether Dest has exactly two predecessors and, if so, compute
7777 // the other predecessor.
7778 pred_iterator PI = pred_begin(DestBB);
7779 BasicBlock *OtherBB = 0;
7783 if (PI == pred_end(DestBB))
7786 if (*PI != StoreBB) {
7791 if (++PI != pred_end(DestBB))
7794 // Bail out if all the relevant blocks aren't distinct (this can happen,
7795 // for example, if SI is in an infinite loop)
7796 if (StoreBB == DestBB || OtherBB == DestBB)
7799 // Verify that the other block ends in a branch and is not otherwise empty.
7800 BasicBlock::iterator BBI = OtherBB->getTerminator();
7801 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
7802 if (!OtherBr || BBI == OtherBB->begin())
7805 // If the other block ends in an unconditional branch, check for the 'if then
7806 // else' case. there is an instruction before the branch.
7807 StoreInst *OtherStore = 0;
7808 if (OtherBr->isUnconditional()) {
7810 // Skip over debugging info.
7811 while (isa<DbgInfoIntrinsic>(BBI) ||
7812 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
7813 if (BBI==OtherBB->begin())
7817 // If this isn't a store, isn't a store to the same location, or if the
7818 // alignments differ, bail out.
7819 OtherStore = dyn_cast<StoreInst>(BBI);
7820 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
7821 OtherStore->getAlignment() != SI.getAlignment())
7824 // Otherwise, the other block ended with a conditional branch. If one of the
7825 // destinations is StoreBB, then we have the if/then case.
7826 if (OtherBr->getSuccessor(0) != StoreBB &&
7827 OtherBr->getSuccessor(1) != StoreBB)
7830 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
7831 // if/then triangle. See if there is a store to the same ptr as SI that
7832 // lives in OtherBB.
7834 // Check to see if we find the matching store.
7835 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
7836 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
7837 OtherStore->getAlignment() != SI.getAlignment())
7841 // If we find something that may be using or overwriting the stored
7842 // value, or if we run out of instructions, we can't do the xform.
7843 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
7844 BBI == OtherBB->begin())
7848 // In order to eliminate the store in OtherBr, we have to
7849 // make sure nothing reads or overwrites the stored value in
7851 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
7852 // FIXME: This should really be AA driven.
7853 if (I->mayReadFromMemory() || I->mayWriteToMemory())
7858 // Insert a PHI node now if we need it.
7859 Value *MergedVal = OtherStore->getOperand(0);
7860 if (MergedVal != SI.getOperand(0)) {
7861 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
7862 PN->reserveOperandSpace(2);
7863 PN->addIncoming(SI.getOperand(0), SI.getParent());
7864 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
7865 MergedVal = InsertNewInstBefore(PN, DestBB->front());
7868 // Advance to a place where it is safe to insert the new store and
7870 BBI = DestBB->getFirstNonPHI();
7871 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
7872 OtherStore->isVolatile(),
7873 SI.getAlignment()), *BBI);
7875 // Nuke the old stores.
7876 EraseInstFromFunction(SI);
7877 EraseInstFromFunction(*OtherStore);
7883 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
7884 // Change br (not X), label True, label False to: br X, label False, True
7886 BasicBlock *TrueDest;
7887 BasicBlock *FalseDest;
7888 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
7889 !isa<Constant>(X)) {
7890 // Swap Destinations and condition...
7892 BI.setSuccessor(0, FalseDest);
7893 BI.setSuccessor(1, TrueDest);
7897 // Cannonicalize fcmp_one -> fcmp_oeq
7898 FCmpInst::Predicate FPred; Value *Y;
7899 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
7900 TrueDest, FalseDest)) &&
7901 BI.getCondition()->hasOneUse())
7902 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
7903 FPred == FCmpInst::FCMP_OGE) {
7904 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
7905 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
7907 // Swap Destinations and condition.
7908 BI.setSuccessor(0, FalseDest);
7909 BI.setSuccessor(1, TrueDest);
7914 // Cannonicalize icmp_ne -> icmp_eq
7915 ICmpInst::Predicate IPred;
7916 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
7917 TrueDest, FalseDest)) &&
7918 BI.getCondition()->hasOneUse())
7919 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
7920 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
7921 IPred == ICmpInst::ICMP_SGE) {
7922 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
7923 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
7924 // Swap Destinations and condition.
7925 BI.setSuccessor(0, FalseDest);
7926 BI.setSuccessor(1, TrueDest);
7934 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
7935 Value *Cond = SI.getCondition();
7936 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
7937 if (I->getOpcode() == Instruction::Add)
7938 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7939 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
7940 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
7942 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
7944 SI.setOperand(0, I->getOperand(0));
7952 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
7953 Value *Agg = EV.getAggregateOperand();
7955 if (!EV.hasIndices())
7956 return ReplaceInstUsesWith(EV, Agg);
7958 if (Constant *C = dyn_cast<Constant>(Agg)) {
7959 if (isa<UndefValue>(C))
7960 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
7962 if (isa<ConstantAggregateZero>(C))
7963 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
7965 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
7966 // Extract the element indexed by the first index out of the constant
7967 Value *V = C->getOperand(*EV.idx_begin());
7968 if (EV.getNumIndices() > 1)
7969 // Extract the remaining indices out of the constant indexed by the
7971 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
7973 return ReplaceInstUsesWith(EV, V);
7975 return 0; // Can't handle other constants
7977 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
7978 // We're extracting from an insertvalue instruction, compare the indices
7979 const unsigned *exti, *exte, *insi, *inse;
7980 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
7981 exte = EV.idx_end(), inse = IV->idx_end();
7982 exti != exte && insi != inse;
7985 // The insert and extract both reference distinctly different elements.
7986 // This means the extract is not influenced by the insert, and we can
7987 // replace the aggregate operand of the extract with the aggregate
7988 // operand of the insert. i.e., replace
7989 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
7990 // %E = extractvalue { i32, { i32 } } %I, 0
7992 // %E = extractvalue { i32, { i32 } } %A, 0
7993 return ExtractValueInst::Create(IV->getAggregateOperand(),
7994 EV.idx_begin(), EV.idx_end());
7996 if (exti == exte && insi == inse)
7997 // Both iterators are at the end: Index lists are identical. Replace
7998 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
7999 // %C = extractvalue { i32, { i32 } } %B, 1, 0
8001 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
8003 // The extract list is a prefix of the insert list. i.e. replace
8004 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
8005 // %E = extractvalue { i32, { i32 } } %I, 1
8007 // %X = extractvalue { i32, { i32 } } %A, 1
8008 // %E = insertvalue { i32 } %X, i32 42, 0
8009 // by switching the order of the insert and extract (though the
8010 // insertvalue should be left in, since it may have other uses).
8011 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
8012 EV.idx_begin(), EV.idx_end());
8013 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
8017 // The insert list is a prefix of the extract list
8018 // We can simply remove the common indices from the extract and make it
8019 // operate on the inserted value instead of the insertvalue result.
8021 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
8022 // %E = extractvalue { i32, { i32 } } %I, 1, 0
8024 // %E extractvalue { i32 } { i32 42 }, 0
8025 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
8028 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
8029 // We're extracting from an intrinsic, see if we're the only user, which
8030 // allows us to simplify multiple result intrinsics to simpler things that
8031 // just get one value..
8032 if (II->hasOneUse()) {
8033 // Check if we're grabbing the overflow bit or the result of a 'with
8034 // overflow' intrinsic. If it's the latter we can remove the intrinsic
8035 // and replace it with a traditional binary instruction.
8036 switch (II->getIntrinsicID()) {
8037 case Intrinsic::uadd_with_overflow:
8038 case Intrinsic::sadd_with_overflow:
8039 if (*EV.idx_begin() == 0) { // Normal result.
8040 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
8041 II->replaceAllUsesWith(UndefValue::get(II->getType()));
8042 EraseInstFromFunction(*II);
8043 return BinaryOperator::CreateAdd(LHS, RHS);
8046 case Intrinsic::usub_with_overflow:
8047 case Intrinsic::ssub_with_overflow:
8048 if (*EV.idx_begin() == 0) { // Normal result.
8049 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
8050 II->replaceAllUsesWith(UndefValue::get(II->getType()));
8051 EraseInstFromFunction(*II);
8052 return BinaryOperator::CreateSub(LHS, RHS);
8055 case Intrinsic::umul_with_overflow:
8056 case Intrinsic::smul_with_overflow:
8057 if (*EV.idx_begin() == 0) { // Normal result.
8058 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
8059 II->replaceAllUsesWith(UndefValue::get(II->getType()));
8060 EraseInstFromFunction(*II);
8061 return BinaryOperator::CreateMul(LHS, RHS);
8069 // Can't simplify extracts from other values. Note that nested extracts are
8070 // already simplified implicitely by the above (extract ( extract (insert) )
8071 // will be translated into extract ( insert ( extract ) ) first and then just
8072 // the value inserted, if appropriate).
8076 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8077 /// is to leave as a vector operation.
8078 static bool CheapToScalarize(Value *V, bool isConstant) {
8079 if (isa<ConstantAggregateZero>(V))
8081 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
8082 if (isConstant) return true;
8083 // If all elts are the same, we can extract.
8084 Constant *Op0 = C->getOperand(0);
8085 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8086 if (C->getOperand(i) != Op0)
8090 Instruction *I = dyn_cast<Instruction>(V);
8091 if (!I) return false;
8093 // Insert element gets simplified to the inserted element or is deleted if
8094 // this is constant idx extract element and its a constant idx insertelt.
8095 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8096 isa<ConstantInt>(I->getOperand(2)))
8098 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8100 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8101 if (BO->hasOneUse() &&
8102 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8103 CheapToScalarize(BO->getOperand(1), isConstant)))
8105 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8106 if (CI->hasOneUse() &&
8107 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8108 CheapToScalarize(CI->getOperand(1), isConstant)))
8114 /// Read and decode a shufflevector mask.
8116 /// It turns undef elements into values that are larger than the number of
8117 /// elements in the input.
8118 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8119 unsigned NElts = SVI->getType()->getNumElements();
8120 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8121 return std::vector<unsigned>(NElts, 0);
8122 if (isa<UndefValue>(SVI->getOperand(2)))
8123 return std::vector<unsigned>(NElts, 2*NElts);
8125 std::vector<unsigned> Result;
8126 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
8127 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
8128 if (isa<UndefValue>(*i))
8129 Result.push_back(NElts*2); // undef -> 8
8131 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
8135 /// FindScalarElement - Given a vector and an element number, see if the scalar
8136 /// value is already around as a register, for example if it were inserted then
8137 /// extracted from the vector.
8138 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8139 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
8140 const VectorType *PTy = cast<VectorType>(V->getType());
8141 unsigned Width = PTy->getNumElements();
8142 if (EltNo >= Width) // Out of range access.
8143 return UndefValue::get(PTy->getElementType());
8145 if (isa<UndefValue>(V))
8146 return UndefValue::get(PTy->getElementType());
8147 else if (isa<ConstantAggregateZero>(V))
8148 return Constant::getNullValue(PTy->getElementType());
8149 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
8150 return CP->getOperand(EltNo);
8151 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8152 // If this is an insert to a variable element, we don't know what it is.
8153 if (!isa<ConstantInt>(III->getOperand(2)))
8155 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8157 // If this is an insert to the element we are looking for, return the
8160 return III->getOperand(1);
8162 // Otherwise, the insertelement doesn't modify the value, recurse on its
8164 return FindScalarElement(III->getOperand(0), EltNo);
8165 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8167 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
8168 unsigned InEl = getShuffleMask(SVI)[EltNo];
8169 if (InEl < LHSWidth)
8170 return FindScalarElement(SVI->getOperand(0), InEl);
8171 else if (InEl < LHSWidth*2)
8172 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
8174 return UndefValue::get(PTy->getElementType());
8177 // Otherwise, we don't know.
8181 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8182 // If vector val is undef, replace extract with scalar undef.
8183 if (isa<UndefValue>(EI.getOperand(0)))
8184 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8186 // If vector val is constant 0, replace extract with scalar 0.
8187 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8188 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8190 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
8191 // If vector val is constant with all elements the same, replace EI with
8192 // that element. When the elements are not identical, we cannot replace yet
8193 // (we do that below, but only when the index is constant).
8194 Constant *op0 = C->getOperand(0);
8195 for (unsigned i = 1; i != C->getNumOperands(); ++i)
8196 if (C->getOperand(i) != op0) {
8201 return ReplaceInstUsesWith(EI, op0);
8204 // If extracting a specified index from the vector, see if we can recursively
8205 // find a previously computed scalar that was inserted into the vector.
8206 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8207 unsigned IndexVal = IdxC->getZExtValue();
8208 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
8210 // If this is extracting an invalid index, turn this into undef, to avoid
8211 // crashing the code below.
8212 if (IndexVal >= VectorWidth)
8213 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8215 // This instruction only demands the single element from the input vector.
8216 // If the input vector has a single use, simplify it based on this use
8218 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
8219 APInt UndefElts(VectorWidth, 0);
8220 APInt DemandedMask(VectorWidth, 1 << IndexVal);
8221 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8222 DemandedMask, UndefElts)) {
8223 EI.setOperand(0, V);
8228 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8229 return ReplaceInstUsesWith(EI, Elt);
8231 // If the this extractelement is directly using a bitcast from a vector of
8232 // the same number of elements, see if we can find the source element from
8233 // it. In this case, we will end up needing to bitcast the scalars.
8234 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
8235 if (const VectorType *VT =
8236 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
8237 if (VT->getNumElements() == VectorWidth)
8238 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
8239 return new BitCastInst(Elt, EI.getType());
8243 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8244 // Push extractelement into predecessor operation if legal and
8245 // profitable to do so
8246 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8247 if (I->hasOneUse() &&
8248 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
8250 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
8251 EI.getName()+".lhs");
8253 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
8254 EI.getName()+".rhs");
8255 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
8257 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8258 // Extracting the inserted element?
8259 if (IE->getOperand(2) == EI.getOperand(1))
8260 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8261 // If the inserted and extracted elements are constants, they must not
8262 // be the same value, extract from the pre-inserted value instead.
8263 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
8264 Worklist.AddValue(EI.getOperand(0));
8265 EI.setOperand(0, IE->getOperand(0));
8268 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8269 // If this is extracting an element from a shufflevector, figure out where
8270 // it came from and extract from the appropriate input element instead.
8271 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8272 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8275 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
8277 if (SrcIdx < LHSWidth)
8278 Src = SVI->getOperand(0);
8279 else if (SrcIdx < LHSWidth*2) {
8281 Src = SVI->getOperand(1);
8283 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8285 return ExtractElementInst::Create(Src,
8286 ConstantInt::get(Type::getInt32Ty(EI.getContext()),
8290 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
8295 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8296 /// elements from either LHS or RHS, return the shuffle mask and true.
8297 /// Otherwise, return false.
8298 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8299 std::vector<Constant*> &Mask) {
8300 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8301 "Invalid CollectSingleShuffleElements");
8302 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8304 if (isa<UndefValue>(V)) {
8305 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
8310 for (unsigned i = 0; i != NumElts; ++i)
8311 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
8316 for (unsigned i = 0; i != NumElts; ++i)
8317 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
8322 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8323 // If this is an insert of an extract from some other vector, include it.
8324 Value *VecOp = IEI->getOperand(0);
8325 Value *ScalarOp = IEI->getOperand(1);
8326 Value *IdxOp = IEI->getOperand(2);
8328 if (!isa<ConstantInt>(IdxOp))
8330 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8332 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8333 // Okay, we can handle this if the vector we are insertinting into is
8335 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8336 // If so, update the mask to reflect the inserted undef.
8337 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
8340 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8341 if (isa<ConstantInt>(EI->getOperand(1)) &&
8342 EI->getOperand(0)->getType() == V->getType()) {
8343 unsigned ExtractedIdx =
8344 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8346 // This must be extracting from either LHS or RHS.
8347 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8348 // Okay, we can handle this if the vector we are insertinting into is
8350 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8351 // If so, update the mask to reflect the inserted value.
8352 if (EI->getOperand(0) == LHS) {
8353 Mask[InsertedIdx % NumElts] =
8354 ConstantInt::get(Type::getInt32Ty(V->getContext()),
8357 assert(EI->getOperand(0) == RHS);
8358 Mask[InsertedIdx % NumElts] =
8359 ConstantInt::get(Type::getInt32Ty(V->getContext()),
8360 ExtractedIdx+NumElts);
8369 // TODO: Handle shufflevector here!
8374 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8375 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8376 /// that computes V and the LHS value of the shuffle.
8377 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8379 assert(isa<VectorType>(V->getType()) &&
8380 (RHS == 0 || V->getType() == RHS->getType()) &&
8381 "Invalid shuffle!");
8382 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8384 if (isa<UndefValue>(V)) {
8385 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
8387 } else if (isa<ConstantAggregateZero>(V)) {
8388 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
8390 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8391 // If this is an insert of an extract from some other vector, include it.
8392 Value *VecOp = IEI->getOperand(0);
8393 Value *ScalarOp = IEI->getOperand(1);
8394 Value *IdxOp = IEI->getOperand(2);
8396 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8397 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8398 EI->getOperand(0)->getType() == V->getType()) {
8399 unsigned ExtractedIdx =
8400 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8401 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8403 // Either the extracted from or inserted into vector must be RHSVec,
8404 // otherwise we'd end up with a shuffle of three inputs.
8405 if (EI->getOperand(0) == RHS || RHS == 0) {
8406 RHS = EI->getOperand(0);
8407 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8408 Mask[InsertedIdx % NumElts] =
8409 ConstantInt::get(Type::getInt32Ty(V->getContext()),
8410 NumElts+ExtractedIdx);
8415 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8416 // Everything but the extracted element is replaced with the RHS.
8417 for (unsigned i = 0; i != NumElts; ++i) {
8418 if (i != InsertedIdx)
8419 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()),
8425 // If this insertelement is a chain that comes from exactly these two
8426 // vectors, return the vector and the effective shuffle.
8427 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8428 return EI->getOperand(0);
8432 // TODO: Handle shufflevector here!
8434 // Otherwise, can't do anything fancy. Return an identity vector.
8435 for (unsigned i = 0; i != NumElts; ++i)
8436 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
8440 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8441 Value *VecOp = IE.getOperand(0);
8442 Value *ScalarOp = IE.getOperand(1);
8443 Value *IdxOp = IE.getOperand(2);
8445 // Inserting an undef or into an undefined place, remove this.
8446 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
8447 ReplaceInstUsesWith(IE, VecOp);
8449 // If the inserted element was extracted from some other vector, and if the
8450 // indexes are constant, try to turn this into a shufflevector operation.
8451 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8452 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8453 EI->getOperand(0)->getType() == IE.getType()) {
8454 unsigned NumVectorElts = IE.getType()->getNumElements();
8455 unsigned ExtractedIdx =
8456 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8457 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8459 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8460 return ReplaceInstUsesWith(IE, VecOp);
8462 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8463 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8465 // If we are extracting a value from a vector, then inserting it right
8466 // back into the same place, just use the input vector.
8467 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8468 return ReplaceInstUsesWith(IE, VecOp);
8470 // If this insertelement isn't used by some other insertelement, turn it
8471 // (and any insertelements it points to), into one big shuffle.
8472 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8473 std::vector<Constant*> Mask;
8475 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8476 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8477 // We now have a shuffle of LHS, RHS, Mask.
8478 return new ShuffleVectorInst(LHS, RHS,
8479 ConstantVector::get(Mask));
8484 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
8485 APInt UndefElts(VWidth, 0);
8486 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
8487 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
8494 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8495 Value *LHS = SVI.getOperand(0);
8496 Value *RHS = SVI.getOperand(1);
8497 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8499 bool MadeChange = false;
8501 // Undefined shuffle mask -> undefined value.
8502 if (isa<UndefValue>(SVI.getOperand(2)))
8503 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8505 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
8507 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
8510 APInt UndefElts(VWidth, 0);
8511 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
8512 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
8513 LHS = SVI.getOperand(0);
8514 RHS = SVI.getOperand(1);
8518 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8519 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8520 if (LHS == RHS || isa<UndefValue>(LHS)) {
8521 if (isa<UndefValue>(LHS) && LHS == RHS) {
8522 // shuffle(undef,undef,mask) -> undef.
8523 return ReplaceInstUsesWith(SVI, LHS);
8526 // Remap any references to RHS to use LHS.
8527 std::vector<Constant*> Elts;
8528 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8530 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
8532 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8533 (Mask[i] < e && isa<UndefValue>(LHS))) {
8534 Mask[i] = 2*e; // Turn into undef.
8535 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
8537 Mask[i] = Mask[i] % e; // Force to LHS.
8538 Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
8543 SVI.setOperand(0, SVI.getOperand(1));
8544 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8545 SVI.setOperand(2, ConstantVector::get(Elts));
8546 LHS = SVI.getOperand(0);
8547 RHS = SVI.getOperand(1);
8551 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8552 bool isLHSID = true, isRHSID = true;
8554 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8555 if (Mask[i] >= e*2) continue; // Ignore undef values.
8556 // Is this an identity shuffle of the LHS value?
8557 isLHSID &= (Mask[i] == i);
8559 // Is this an identity shuffle of the RHS value?
8560 isRHSID &= (Mask[i]-e == i);
8563 // Eliminate identity shuffles.
8564 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8565 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
8567 // If the LHS is a shufflevector itself, see if we can combine it with this
8568 // one without producing an unusual shuffle. Here we are really conservative:
8569 // we are absolutely afraid of producing a shuffle mask not in the input
8570 // program, because the code gen may not be smart enough to turn a merged
8571 // shuffle into two specific shuffles: it may produce worse code. As such,
8572 // we only merge two shuffles if the result is one of the two input shuffle
8573 // masks. In this case, merging the shuffles just removes one instruction,
8574 // which we know is safe. This is good for things like turning:
8575 // (splat(splat)) -> splat.
8576 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
8577 if (isa<UndefValue>(RHS)) {
8578 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
8580 if (LHSMask.size() == Mask.size()) {
8581 std::vector<unsigned> NewMask;
8582 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
8584 NewMask.push_back(2*e);
8586 NewMask.push_back(LHSMask[Mask[i]]);
8588 // If the result mask is equal to the src shuffle or this
8589 // shuffle mask, do the replacement.
8590 if (NewMask == LHSMask || NewMask == Mask) {
8591 unsigned LHSInNElts =
8592 cast<VectorType>(LHSSVI->getOperand(0)->getType())->
8594 std::vector<Constant*> Elts;
8595 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
8596 if (NewMask[i] >= LHSInNElts*2) {
8597 Elts.push_back(UndefValue::get(
8598 Type::getInt32Ty(SVI.getContext())));
8600 Elts.push_back(ConstantInt::get(
8601 Type::getInt32Ty(SVI.getContext()),
8605 return new ShuffleVectorInst(LHSSVI->getOperand(0),
8606 LHSSVI->getOperand(1),
8607 ConstantVector::get(Elts));
8613 return MadeChange ? &SVI : 0;
8619 /// TryToSinkInstruction - Try to move the specified instruction from its
8620 /// current block into the beginning of DestBlock, which can only happen if it's
8621 /// safe to move the instruction past all of the instructions between it and the
8622 /// end of its block.
8623 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
8624 assert(I->hasOneUse() && "Invariants didn't hold!");
8626 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
8627 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
8630 // Do not sink alloca instructions out of the entry block.
8631 if (isa<AllocaInst>(I) && I->getParent() ==
8632 &DestBlock->getParent()->getEntryBlock())
8635 // We can only sink load instructions if there is nothing between the load and
8636 // the end of block that could change the value.
8637 if (I->mayReadFromMemory()) {
8638 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
8640 if (Scan->mayWriteToMemory())
8644 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
8646 I->moveBefore(InsertPos);
8652 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
8653 /// all reachable code to the worklist.
8655 /// This has a couple of tricks to make the code faster and more powerful. In
8656 /// particular, we constant fold and DCE instructions as we go, to avoid adding
8657 /// them to the worklist (this significantly speeds up instcombine on code where
8658 /// many instructions are dead or constant). Additionally, if we find a branch
8659 /// whose condition is a known constant, we only visit the reachable successors.
8661 static bool AddReachableCodeToWorklist(BasicBlock *BB,
8662 SmallPtrSet<BasicBlock*, 64> &Visited,
8664 const TargetData *TD) {
8665 bool MadeIRChange = false;
8666 SmallVector<BasicBlock*, 256> Worklist;
8667 Worklist.push_back(BB);
8669 std::vector<Instruction*> InstrsForInstCombineWorklist;
8670 InstrsForInstCombineWorklist.reserve(128);
8672 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
8674 while (!Worklist.empty()) {
8675 BB = Worklist.back();
8676 Worklist.pop_back();
8678 // We have now visited this block! If we've already been here, ignore it.
8679 if (!Visited.insert(BB)) continue;
8681 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
8682 Instruction *Inst = BBI++;
8684 // DCE instruction if trivially dead.
8685 if (isInstructionTriviallyDead(Inst)) {
8687 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
8688 Inst->eraseFromParent();
8692 // ConstantProp instruction if trivially constant.
8693 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
8694 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
8695 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
8697 Inst->replaceAllUsesWith(C);
8699 Inst->eraseFromParent();
8706 // See if we can constant fold its operands.
8707 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
8709 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
8710 if (CE == 0) continue;
8712 // If we already folded this constant, don't try again.
8713 if (!FoldedConstants.insert(CE))
8716 Constant *NewC = ConstantFoldConstantExpression(CE, TD);
8717 if (NewC && NewC != CE) {
8719 MadeIRChange = true;
8725 InstrsForInstCombineWorklist.push_back(Inst);
8728 // Recursively visit successors. If this is a branch or switch on a
8729 // constant, only visit the reachable successor.
8730 TerminatorInst *TI = BB->getTerminator();
8731 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
8732 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
8733 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
8734 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
8735 Worklist.push_back(ReachableBB);
8738 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
8739 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
8740 // See if this is an explicit destination.
8741 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
8742 if (SI->getCaseValue(i) == Cond) {
8743 BasicBlock *ReachableBB = SI->getSuccessor(i);
8744 Worklist.push_back(ReachableBB);
8748 // Otherwise it is the default destination.
8749 Worklist.push_back(SI->getSuccessor(0));
8754 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
8755 Worklist.push_back(TI->getSuccessor(i));
8758 // Once we've found all of the instructions to add to instcombine's worklist,
8759 // add them in reverse order. This way instcombine will visit from the top
8760 // of the function down. This jives well with the way that it adds all uses
8761 // of instructions to the worklist after doing a transformation, thus avoiding
8762 // some N^2 behavior in pathological cases.
8763 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
8764 InstrsForInstCombineWorklist.size());
8766 return MadeIRChange;
8769 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
8770 MadeIRChange = false;
8772 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
8773 << F.getNameStr() << "\n");
8776 // Do a depth-first traversal of the function, populate the worklist with
8777 // the reachable instructions. Ignore blocks that are not reachable. Keep
8778 // track of which blocks we visit.
8779 SmallPtrSet<BasicBlock*, 64> Visited;
8780 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
8782 // Do a quick scan over the function. If we find any blocks that are
8783 // unreachable, remove any instructions inside of them. This prevents
8784 // the instcombine code from having to deal with some bad special cases.
8785 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
8786 if (!Visited.count(BB)) {
8787 Instruction *Term = BB->getTerminator();
8788 while (Term != BB->begin()) { // Remove instrs bottom-up
8789 BasicBlock::iterator I = Term; --I;
8791 DEBUG(errs() << "IC: DCE: " << *I << '\n');
8792 // A debug intrinsic shouldn't force another iteration if we weren't
8793 // going to do one without it.
8794 if (!isa<DbgInfoIntrinsic>(I)) {
8796 MadeIRChange = true;
8799 // If I is not void type then replaceAllUsesWith undef.
8800 // This allows ValueHandlers and custom metadata to adjust itself.
8801 if (!I->getType()->isVoidTy())
8802 I->replaceAllUsesWith(UndefValue::get(I->getType()));
8803 I->eraseFromParent();
8808 while (!Worklist.isEmpty()) {
8809 Instruction *I = Worklist.RemoveOne();
8810 if (I == 0) continue; // skip null values.
8812 // Check to see if we can DCE the instruction.
8813 if (isInstructionTriviallyDead(I)) {
8814 DEBUG(errs() << "IC: DCE: " << *I << '\n');
8815 EraseInstFromFunction(*I);
8817 MadeIRChange = true;
8821 // Instruction isn't dead, see if we can constant propagate it.
8822 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
8823 if (Constant *C = ConstantFoldInstruction(I, TD)) {
8824 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
8826 // Add operands to the worklist.
8827 ReplaceInstUsesWith(*I, C);
8829 EraseInstFromFunction(*I);
8830 MadeIRChange = true;
8834 // See if we can trivially sink this instruction to a successor basic block.
8835 if (I->hasOneUse()) {
8836 BasicBlock *BB = I->getParent();
8837 Instruction *UserInst = cast<Instruction>(I->use_back());
8838 BasicBlock *UserParent;
8840 // Get the block the use occurs in.
8841 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
8842 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
8844 UserParent = UserInst->getParent();
8846 if (UserParent != BB) {
8847 bool UserIsSuccessor = false;
8848 // See if the user is one of our successors.
8849 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
8850 if (*SI == UserParent) {
8851 UserIsSuccessor = true;
8855 // If the user is one of our immediate successors, and if that successor
8856 // only has us as a predecessors (we'd have to split the critical edge
8857 // otherwise), we can keep going.
8858 if (UserIsSuccessor && UserParent->getSinglePredecessor())
8859 // Okay, the CFG is simple enough, try to sink this instruction.
8860 MadeIRChange |= TryToSinkInstruction(I, UserParent);
8864 // Now that we have an instruction, try combining it to simplify it.
8865 Builder->SetInsertPoint(I->getParent(), I);
8870 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
8871 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
8873 if (Instruction *Result = visit(*I)) {
8875 // Should we replace the old instruction with a new one?
8877 DEBUG(errs() << "IC: Old = " << *I << '\n'
8878 << " New = " << *Result << '\n');
8880 // Everything uses the new instruction now.
8881 I->replaceAllUsesWith(Result);
8883 // Push the new instruction and any users onto the worklist.
8884 Worklist.Add(Result);
8885 Worklist.AddUsersToWorkList(*Result);
8887 // Move the name to the new instruction first.
8888 Result->takeName(I);
8890 // Insert the new instruction into the basic block...
8891 BasicBlock *InstParent = I->getParent();
8892 BasicBlock::iterator InsertPos = I;
8894 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
8895 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
8898 InstParent->getInstList().insert(InsertPos, Result);
8900 EraseInstFromFunction(*I);
8903 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
8904 << " New = " << *I << '\n');
8907 // If the instruction was modified, it's possible that it is now dead.
8908 // if so, remove it.
8909 if (isInstructionTriviallyDead(I)) {
8910 EraseInstFromFunction(*I);
8913 Worklist.AddUsersToWorkList(*I);
8916 MadeIRChange = true;
8921 return MadeIRChange;
8925 bool InstCombiner::runOnFunction(Function &F) {
8926 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
8927 TD = getAnalysisIfAvailable<TargetData>();
8930 /// Builder - This is an IRBuilder that automatically inserts new
8931 /// instructions into the worklist when they are created.
8932 IRBuilder<true, TargetFolder, InstCombineIRInserter>
8933 TheBuilder(F.getContext(), TargetFolder(TD),
8934 InstCombineIRInserter(Worklist));
8935 Builder = &TheBuilder;
8937 bool EverMadeChange = false;
8939 // Iterate while there is work to do.
8940 unsigned Iteration = 0;
8941 while (DoOneIteration(F, Iteration++))
8942 EverMadeChange = true;
8945 return EverMadeChange;
8948 FunctionPass *llvm::createInstructionCombiningPass() {
8949 return new InstCombiner();