1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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 algebraic
12 // 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 "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Support/InstVisitor.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/PatternMatch.h"
52 #include "llvm/Support/Compiler.h"
53 #include "llvm/ADT/DenseMap.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/ADT/STLExtras.h"
61 using namespace llvm::PatternMatch;
63 STATISTIC(NumCombined , "Number of insts combined");
64 STATISTIC(NumConstProp, "Number of constant folds");
65 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
66 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
67 STATISTIC(NumSunkInst , "Number of instructions sunk");
70 class VISIBILITY_HIDDEN InstCombiner
71 : public FunctionPass,
72 public InstVisitor<InstCombiner, Instruction*> {
73 // Worklist of all of the instructions that need to be simplified.
74 std::vector<Instruction*> Worklist;
75 DenseMap<Instruction*, unsigned> WorklistMap;
77 bool MustPreserveLCSSA;
79 /// AddToWorkList - Add the specified instruction to the worklist if it
80 /// isn't already in it.
81 void AddToWorkList(Instruction *I) {
82 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
83 Worklist.push_back(I);
86 // RemoveFromWorkList - remove I from the worklist if it exists.
87 void RemoveFromWorkList(Instruction *I) {
88 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
89 if (It == WorklistMap.end()) return; // Not in worklist.
91 // Don't bother moving everything down, just null out the slot.
92 Worklist[It->second] = 0;
94 WorklistMap.erase(It);
97 Instruction *RemoveOneFromWorkList() {
98 Instruction *I = Worklist.back();
100 WorklistMap.erase(I);
105 /// AddUsersToWorkList - When an instruction is simplified, add all users of
106 /// the instruction to the work lists because they might get more simplified
109 void AddUsersToWorkList(Value &I) {
110 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
112 AddToWorkList(cast<Instruction>(*UI));
115 /// AddUsesToWorkList - When an instruction is simplified, add operands to
116 /// the work lists because they might get more simplified now.
118 void AddUsesToWorkList(Instruction &I) {
119 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
120 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
124 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
125 /// dead. Add all of its operands to the worklist, turning them into
126 /// undef's to reduce the number of uses of those instructions.
128 /// Return the specified operand before it is turned into an undef.
130 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
131 Value *R = I.getOperand(op);
133 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
134 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
136 // Set the operand to undef to drop the use.
137 I.setOperand(i, UndefValue::get(Op->getType()));
144 virtual bool runOnFunction(Function &F);
146 bool DoOneIteration(Function &F, unsigned ItNum);
148 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
149 AU.addRequired<TargetData>();
150 AU.addPreservedID(LCSSAID);
151 AU.setPreservesCFG();
154 TargetData &getTargetData() const { return *TD; }
156 // Visitation implementation - Implement instruction combining for different
157 // instruction types. The semantics are as follows:
159 // null - No change was made
160 // I - Change was made, I is still valid, I may be dead though
161 // otherwise - Change was made, replace I with returned instruction
163 Instruction *visitAdd(BinaryOperator &I);
164 Instruction *visitSub(BinaryOperator &I);
165 Instruction *visitMul(BinaryOperator &I);
166 Instruction *visitURem(BinaryOperator &I);
167 Instruction *visitSRem(BinaryOperator &I);
168 Instruction *visitFRem(BinaryOperator &I);
169 Instruction *commonRemTransforms(BinaryOperator &I);
170 Instruction *commonIRemTransforms(BinaryOperator &I);
171 Instruction *commonDivTransforms(BinaryOperator &I);
172 Instruction *commonIDivTransforms(BinaryOperator &I);
173 Instruction *visitUDiv(BinaryOperator &I);
174 Instruction *visitSDiv(BinaryOperator &I);
175 Instruction *visitFDiv(BinaryOperator &I);
176 Instruction *visitAnd(BinaryOperator &I);
177 Instruction *visitOr (BinaryOperator &I);
178 Instruction *visitXor(BinaryOperator &I);
179 Instruction *visitShl(BinaryOperator &I);
180 Instruction *visitAShr(BinaryOperator &I);
181 Instruction *visitLShr(BinaryOperator &I);
182 Instruction *commonShiftTransforms(BinaryOperator &I);
183 Instruction *visitFCmpInst(FCmpInst &I);
184 Instruction *visitICmpInst(ICmpInst &I);
185 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
187 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
188 ICmpInst::Predicate Cond, Instruction &I);
189 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
191 Instruction *commonCastTransforms(CastInst &CI);
192 Instruction *commonIntCastTransforms(CastInst &CI);
193 Instruction *visitTrunc(CastInst &CI);
194 Instruction *visitZExt(CastInst &CI);
195 Instruction *visitSExt(CastInst &CI);
196 Instruction *visitFPTrunc(CastInst &CI);
197 Instruction *visitFPExt(CastInst &CI);
198 Instruction *visitFPToUI(CastInst &CI);
199 Instruction *visitFPToSI(CastInst &CI);
200 Instruction *visitUIToFP(CastInst &CI);
201 Instruction *visitSIToFP(CastInst &CI);
202 Instruction *visitPtrToInt(CastInst &CI);
203 Instruction *visitIntToPtr(CastInst &CI);
204 Instruction *visitBitCast(CastInst &CI);
205 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
207 Instruction *visitSelectInst(SelectInst &CI);
208 Instruction *visitCallInst(CallInst &CI);
209 Instruction *visitInvokeInst(InvokeInst &II);
210 Instruction *visitPHINode(PHINode &PN);
211 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
212 Instruction *visitAllocationInst(AllocationInst &AI);
213 Instruction *visitFreeInst(FreeInst &FI);
214 Instruction *visitLoadInst(LoadInst &LI);
215 Instruction *visitStoreInst(StoreInst &SI);
216 Instruction *visitBranchInst(BranchInst &BI);
217 Instruction *visitSwitchInst(SwitchInst &SI);
218 Instruction *visitInsertElementInst(InsertElementInst &IE);
219 Instruction *visitExtractElementInst(ExtractElementInst &EI);
220 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
222 // visitInstruction - Specify what to return for unhandled instructions...
223 Instruction *visitInstruction(Instruction &I) { return 0; }
226 Instruction *visitCallSite(CallSite CS);
227 bool transformConstExprCastCall(CallSite CS);
230 // InsertNewInstBefore - insert an instruction New before instruction Old
231 // in the program. Add the new instruction to the worklist.
233 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
234 assert(New && New->getParent() == 0 &&
235 "New instruction already inserted into a basic block!");
236 BasicBlock *BB = Old.getParent();
237 BB->getInstList().insert(&Old, New); // Insert inst
242 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
243 /// This also adds the cast to the worklist. Finally, this returns the
245 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
247 if (V->getType() == Ty) return V;
249 if (Constant *CV = dyn_cast<Constant>(V))
250 return ConstantExpr::getCast(opc, CV, Ty);
252 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
257 // ReplaceInstUsesWith - This method is to be used when an instruction is
258 // found to be dead, replacable with another preexisting expression. Here
259 // we add all uses of I to the worklist, replace all uses of I with the new
260 // value, then return I, so that the inst combiner will know that I was
263 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
264 AddUsersToWorkList(I); // Add all modified instrs to worklist
266 I.replaceAllUsesWith(V);
269 // If we are replacing the instruction with itself, this must be in a
270 // segment of unreachable code, so just clobber the instruction.
271 I.replaceAllUsesWith(UndefValue::get(I.getType()));
276 // UpdateValueUsesWith - This method is to be used when an value is
277 // found to be replacable with another preexisting expression or was
278 // updated. Here we add all uses of I to the worklist, replace all uses of
279 // I with the new value (unless the instruction was just updated), then
280 // return true, so that the inst combiner will know that I was modified.
282 bool UpdateValueUsesWith(Value *Old, Value *New) {
283 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
285 Old->replaceAllUsesWith(New);
286 if (Instruction *I = dyn_cast<Instruction>(Old))
288 if (Instruction *I = dyn_cast<Instruction>(New))
293 // EraseInstFromFunction - When dealing with an instruction that has side
294 // effects or produces a void value, we can't rely on DCE to delete the
295 // instruction. Instead, visit methods should return the value returned by
297 Instruction *EraseInstFromFunction(Instruction &I) {
298 assert(I.use_empty() && "Cannot erase instruction that is used!");
299 AddUsesToWorkList(I);
300 RemoveFromWorkList(&I);
302 return 0; // Don't do anything with FI
306 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
307 /// InsertBefore instruction. This is specialized a bit to avoid inserting
308 /// casts that are known to not do anything...
310 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
311 Value *V, const Type *DestTy,
312 Instruction *InsertBefore);
314 /// SimplifyCommutative - This performs a few simplifications for
315 /// commutative operators.
316 bool SimplifyCommutative(BinaryOperator &I);
318 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
319 /// most-complex to least-complex order.
320 bool SimplifyCompare(CmpInst &I);
322 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
323 uint64_t &KnownZero, uint64_t &KnownOne,
326 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
327 uint64_t &UndefElts, unsigned Depth = 0);
329 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
330 // PHI node as operand #0, see if we can fold the instruction into the PHI
331 // (which is only possible if all operands to the PHI are constants).
332 Instruction *FoldOpIntoPhi(Instruction &I);
334 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
335 // operator and they all are only used by the PHI, PHI together their
336 // inputs, and do the operation once, to the result of the PHI.
337 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
338 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
341 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
342 ConstantInt *AndRHS, BinaryOperator &TheAnd);
344 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
345 bool isSub, Instruction &I);
346 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
347 bool isSigned, bool Inside, Instruction &IB);
348 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
349 Instruction *MatchBSwap(BinaryOperator &I);
351 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
354 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
357 // getComplexity: Assign a complexity or rank value to LLVM Values...
358 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
359 static unsigned getComplexity(Value *V) {
360 if (isa<Instruction>(V)) {
361 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
365 if (isa<Argument>(V)) return 3;
366 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
369 // isOnlyUse - Return true if this instruction will be deleted if we stop using
371 static bool isOnlyUse(Value *V) {
372 return V->hasOneUse() || isa<Constant>(V);
375 // getPromotedType - Return the specified type promoted as it would be to pass
376 // though a va_arg area...
377 static const Type *getPromotedType(const Type *Ty) {
378 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
379 if (ITy->getBitWidth() < 32)
380 return Type::Int32Ty;
381 } else if (Ty == Type::FloatTy)
382 return Type::DoubleTy;
386 /// getBitCastOperand - If the specified operand is a CastInst or a constant
387 /// expression bitcast, return the operand value, otherwise return null.
388 static Value *getBitCastOperand(Value *V) {
389 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
390 return I->getOperand(0);
391 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
392 if (CE->getOpcode() == Instruction::BitCast)
393 return CE->getOperand(0);
397 /// This function is a wrapper around CastInst::isEliminableCastPair. It
398 /// simply extracts arguments and returns what that function returns.
399 static Instruction::CastOps
400 isEliminableCastPair(
401 const CastInst *CI, ///< The first cast instruction
402 unsigned opcode, ///< The opcode of the second cast instruction
403 const Type *DstTy, ///< The target type for the second cast instruction
404 TargetData *TD ///< The target data for pointer size
407 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
408 const Type *MidTy = CI->getType(); // B from above
410 // Get the opcodes of the two Cast instructions
411 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
412 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
414 return Instruction::CastOps(
415 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
416 DstTy, TD->getIntPtrType()));
419 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
420 /// in any code being generated. It does not require codegen if V is simple
421 /// enough or if the cast can be folded into other casts.
422 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
423 const Type *Ty, TargetData *TD) {
424 if (V->getType() == Ty || isa<Constant>(V)) return false;
426 // If this is another cast that can be eliminated, it isn't codegen either.
427 if (const CastInst *CI = dyn_cast<CastInst>(V))
428 if (isEliminableCastPair(CI, opcode, Ty, TD))
433 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
434 /// InsertBefore instruction. This is specialized a bit to avoid inserting
435 /// casts that are known to not do anything...
437 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
438 Value *V, const Type *DestTy,
439 Instruction *InsertBefore) {
440 if (V->getType() == DestTy) return V;
441 if (Constant *C = dyn_cast<Constant>(V))
442 return ConstantExpr::getCast(opcode, C, DestTy);
444 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
447 // SimplifyCommutative - This performs a few simplifications for commutative
450 // 1. Order operands such that they are listed from right (least complex) to
451 // left (most complex). This puts constants before unary operators before
454 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
455 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
457 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
458 bool Changed = false;
459 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
460 Changed = !I.swapOperands();
462 if (!I.isAssociative()) return Changed;
463 Instruction::BinaryOps Opcode = I.getOpcode();
464 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
465 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
466 if (isa<Constant>(I.getOperand(1))) {
467 Constant *Folded = ConstantExpr::get(I.getOpcode(),
468 cast<Constant>(I.getOperand(1)),
469 cast<Constant>(Op->getOperand(1)));
470 I.setOperand(0, Op->getOperand(0));
471 I.setOperand(1, Folded);
473 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
474 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
475 isOnlyUse(Op) && isOnlyUse(Op1)) {
476 Constant *C1 = cast<Constant>(Op->getOperand(1));
477 Constant *C2 = cast<Constant>(Op1->getOperand(1));
479 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
480 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
481 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
485 I.setOperand(0, New);
486 I.setOperand(1, Folded);
493 /// SimplifyCompare - For a CmpInst this function just orders the operands
494 /// so that theyare listed from right (least complex) to left (most complex).
495 /// This puts constants before unary operators before binary operators.
496 bool InstCombiner::SimplifyCompare(CmpInst &I) {
497 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
500 // Compare instructions are not associative so there's nothing else we can do.
504 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
505 // if the LHS is a constant zero (which is the 'negate' form).
507 static inline Value *dyn_castNegVal(Value *V) {
508 if (BinaryOperator::isNeg(V))
509 return BinaryOperator::getNegArgument(V);
511 // Constants can be considered to be negated values if they can be folded.
512 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
513 return ConstantExpr::getNeg(C);
517 static inline Value *dyn_castNotVal(Value *V) {
518 if (BinaryOperator::isNot(V))
519 return BinaryOperator::getNotArgument(V);
521 // Constants can be considered to be not'ed values...
522 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
523 return ConstantExpr::getNot(C);
527 // dyn_castFoldableMul - If this value is a multiply that can be folded into
528 // other computations (because it has a constant operand), return the
529 // non-constant operand of the multiply, and set CST to point to the multiplier.
530 // Otherwise, return null.
532 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
533 if (V->hasOneUse() && V->getType()->isInteger())
534 if (Instruction *I = dyn_cast<Instruction>(V)) {
535 if (I->getOpcode() == Instruction::Mul)
536 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
537 return I->getOperand(0);
538 if (I->getOpcode() == Instruction::Shl)
539 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
540 // The multiplier is really 1 << CST.
541 Constant *One = ConstantInt::get(V->getType(), 1);
542 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
543 return I->getOperand(0);
549 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
550 /// expression, return it.
551 static User *dyn_castGetElementPtr(Value *V) {
552 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
553 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
554 if (CE->getOpcode() == Instruction::GetElementPtr)
555 return cast<User>(V);
559 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
560 static ConstantInt *AddOne(ConstantInt *C) {
561 return cast<ConstantInt>(ConstantExpr::getAdd(C,
562 ConstantInt::get(C->getType(), 1)));
564 static ConstantInt *SubOne(ConstantInt *C) {
565 return cast<ConstantInt>(ConstantExpr::getSub(C,
566 ConstantInt::get(C->getType(), 1)));
569 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
570 /// known to be either zero or one and return them in the KnownZero/KnownOne
571 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
573 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
574 uint64_t &KnownOne, unsigned Depth = 0) {
575 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
576 // we cannot optimize based on the assumption that it is zero without changing
577 // it to be an explicit zero. If we don't change it to zero, other code could
578 // optimized based on the contradictory assumption that it is non-zero.
579 // Because instcombine aggressively folds operations with undef args anyway,
580 // this won't lose us code quality.
581 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
582 // We know all of the bits for a constant!
583 KnownOne = CI->getZExtValue() & Mask;
584 KnownZero = ~KnownOne & Mask;
588 KnownZero = KnownOne = 0; // Don't know anything.
589 if (Depth == 6 || Mask == 0)
590 return; // Limit search depth.
592 uint64_t KnownZero2, KnownOne2;
593 Instruction *I = dyn_cast<Instruction>(V);
596 Mask &= cast<IntegerType>(V->getType())->getBitMask();
598 switch (I->getOpcode()) {
599 case Instruction::And:
600 // If either the LHS or the RHS are Zero, the result is zero.
601 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
603 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
604 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
605 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
607 // Output known-1 bits are only known if set in both the LHS & RHS.
608 KnownOne &= KnownOne2;
609 // Output known-0 are known to be clear if zero in either the LHS | RHS.
610 KnownZero |= KnownZero2;
612 case Instruction::Or:
613 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
615 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
616 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
617 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
619 // Output known-0 bits are only known if clear in both the LHS & RHS.
620 KnownZero &= KnownZero2;
621 // Output known-1 are known to be set if set in either the LHS | RHS.
622 KnownOne |= KnownOne2;
624 case Instruction::Xor: {
625 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
626 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
627 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
628 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
630 // Output known-0 bits are known if clear or set in both the LHS & RHS.
631 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
632 // Output known-1 are known to be set if set in only one of the LHS, RHS.
633 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
634 KnownZero = KnownZeroOut;
637 case Instruction::Select:
638 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
639 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
640 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
641 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
643 // Only known if known in both the LHS and RHS.
644 KnownOne &= KnownOne2;
645 KnownZero &= KnownZero2;
647 case Instruction::FPTrunc:
648 case Instruction::FPExt:
649 case Instruction::FPToUI:
650 case Instruction::FPToSI:
651 case Instruction::SIToFP:
652 case Instruction::PtrToInt:
653 case Instruction::UIToFP:
654 case Instruction::IntToPtr:
655 return; // Can't work with floating point or pointers
656 case Instruction::Trunc:
657 // All these have integer operands
658 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
660 case Instruction::BitCast: {
661 const Type *SrcTy = I->getOperand(0)->getType();
662 if (SrcTy->isInteger()) {
663 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
668 case Instruction::ZExt: {
669 // Compute the bits in the result that are not present in the input.
670 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
671 uint64_t NotIn = ~SrcTy->getBitMask();
672 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
674 Mask &= SrcTy->getBitMask();
675 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
676 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
677 // The top bits are known to be zero.
678 KnownZero |= NewBits;
681 case Instruction::SExt: {
682 // Compute the bits in the result that are not present in the input.
683 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
684 uint64_t NotIn = ~SrcTy->getBitMask();
685 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
687 Mask &= SrcTy->getBitMask();
688 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
689 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
691 // If the sign bit of the input is known set or clear, then we know the
692 // top bits of the result.
693 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
694 if (KnownZero & InSignBit) { // Input sign bit known zero
695 KnownZero |= NewBits;
696 KnownOne &= ~NewBits;
697 } else if (KnownOne & InSignBit) { // Input sign bit known set
699 KnownZero &= ~NewBits;
700 } else { // Input sign bit unknown
701 KnownZero &= ~NewBits;
702 KnownOne &= ~NewBits;
706 case Instruction::Shl:
707 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
708 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
709 uint64_t ShiftAmt = SA->getZExtValue();
711 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
712 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
713 KnownZero <<= ShiftAmt;
714 KnownOne <<= ShiftAmt;
715 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
719 case Instruction::LShr:
720 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
721 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
722 // Compute the new bits that are at the top now.
723 uint64_t ShiftAmt = SA->getZExtValue();
724 uint64_t HighBits = (1ULL << ShiftAmt)-1;
725 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
727 // Unsigned shift right.
729 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
730 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
731 KnownZero >>= ShiftAmt;
732 KnownOne >>= ShiftAmt;
733 KnownZero |= HighBits; // high bits known zero.
737 case Instruction::AShr:
738 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
739 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
740 // Compute the new bits that are at the top now.
741 uint64_t ShiftAmt = SA->getZExtValue();
742 uint64_t HighBits = (1ULL << ShiftAmt)-1;
743 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
745 // Signed shift right.
747 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
748 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
749 KnownZero >>= ShiftAmt;
750 KnownOne >>= ShiftAmt;
752 // Handle the sign bits.
753 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
754 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
756 if (KnownZero & SignBit) { // New bits are known zero.
757 KnownZero |= HighBits;
758 } else if (KnownOne & SignBit) { // New bits are known one.
759 KnownOne |= HighBits;
767 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
768 /// this predicate to simplify operations downstream. Mask is known to be zero
769 /// for bits that V cannot have.
770 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
771 uint64_t KnownZero, KnownOne;
772 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
773 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
774 return (KnownZero & Mask) == Mask;
777 /// ShrinkDemandedConstant - Check to see if the specified operand of the
778 /// specified instruction is a constant integer. If so, check to see if there
779 /// are any bits set in the constant that are not demanded. If so, shrink the
780 /// constant and return true.
781 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
783 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
784 if (!OpC) return false;
786 // If there are no bits set that aren't demanded, nothing to do.
787 if ((~Demanded & OpC->getZExtValue()) == 0)
790 // This is producing any bits that are not needed, shrink the RHS.
791 uint64_t Val = Demanded & OpC->getZExtValue();
792 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Val));
796 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
797 // set of known zero and one bits, compute the maximum and minimum values that
798 // could have the specified known zero and known one bits, returning them in
800 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
803 int64_t &Min, int64_t &Max) {
804 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
805 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
807 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
809 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
810 // bit if it is unknown.
812 Max = KnownOne|UnknownBits;
814 if (SignBit & UnknownBits) { // Sign bit is unknown
819 // Sign extend the min/max values.
820 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
821 Min = (Min << ShAmt) >> ShAmt;
822 Max = (Max << ShAmt) >> ShAmt;
825 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
826 // a set of known zero and one bits, compute the maximum and minimum values that
827 // could have the specified known zero and known one bits, returning them in
829 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
834 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
835 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
837 // The minimum value is when the unknown bits are all zeros.
839 // The maximum value is when the unknown bits are all ones.
840 Max = KnownOne|UnknownBits;
844 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
845 /// DemandedMask bits of the result of V are ever used downstream. If we can
846 /// use this information to simplify V, do so and return true. Otherwise,
847 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
848 /// the expression (used to simplify the caller). The KnownZero/One bits may
849 /// only be accurate for those bits in the DemandedMask.
850 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
851 uint64_t &KnownZero, uint64_t &KnownOne,
853 const IntegerType *VTy = cast<IntegerType>(V->getType());
854 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
855 // We know all of the bits for a constant!
856 KnownOne = CI->getZExtValue() & DemandedMask;
857 KnownZero = ~KnownOne & DemandedMask;
861 KnownZero = KnownOne = 0;
862 if (!V->hasOneUse()) { // Other users may use these bits.
863 if (Depth != 0) { // Not at the root.
864 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
865 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
868 // If this is the root being simplified, allow it to have multiple uses,
869 // just set the DemandedMask to all bits.
870 DemandedMask = VTy->getBitMask();
871 } else if (DemandedMask == 0) { // Not demanding any bits from V.
872 if (V != UndefValue::get(VTy))
873 return UpdateValueUsesWith(V, UndefValue::get(VTy));
875 } else if (Depth == 6) { // Limit search depth.
879 Instruction *I = dyn_cast<Instruction>(V);
880 if (!I) return false; // Only analyze instructions.
882 DemandedMask &= VTy->getBitMask();
884 uint64_t KnownZero2 = 0, KnownOne2 = 0;
885 switch (I->getOpcode()) {
887 case Instruction::And:
888 // If either the LHS or the RHS are Zero, the result is zero.
889 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
890 KnownZero, KnownOne, Depth+1))
892 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
894 // If something is known zero on the RHS, the bits aren't demanded on the
896 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
897 KnownZero2, KnownOne2, Depth+1))
899 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
901 // If all of the demanded bits are known 1 on one side, return the other.
902 // These bits cannot contribute to the result of the 'and'.
903 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
904 return UpdateValueUsesWith(I, I->getOperand(0));
905 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
906 return UpdateValueUsesWith(I, I->getOperand(1));
908 // If all of the demanded bits in the inputs are known zeros, return zero.
909 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
910 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
912 // If the RHS is a constant, see if we can simplify it.
913 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
914 return UpdateValueUsesWith(I, I);
916 // Output known-1 bits are only known if set in both the LHS & RHS.
917 KnownOne &= KnownOne2;
918 // Output known-0 are known to be clear if zero in either the LHS | RHS.
919 KnownZero |= KnownZero2;
921 case Instruction::Or:
922 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
923 KnownZero, KnownOne, Depth+1))
925 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
926 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
927 KnownZero2, KnownOne2, Depth+1))
929 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
931 // If all of the demanded bits are known zero on one side, return the other.
932 // These bits cannot contribute to the result of the 'or'.
933 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
934 return UpdateValueUsesWith(I, I->getOperand(0));
935 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
936 return UpdateValueUsesWith(I, I->getOperand(1));
938 // If all of the potentially set bits on one side are known to be set on
939 // the other side, just use the 'other' side.
940 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
941 (DemandedMask & (~KnownZero)))
942 return UpdateValueUsesWith(I, I->getOperand(0));
943 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
944 (DemandedMask & (~KnownZero2)))
945 return UpdateValueUsesWith(I, I->getOperand(1));
947 // If the RHS is a constant, see if we can simplify it.
948 if (ShrinkDemandedConstant(I, 1, DemandedMask))
949 return UpdateValueUsesWith(I, I);
951 // Output known-0 bits are only known if clear in both the LHS & RHS.
952 KnownZero &= KnownZero2;
953 // Output known-1 are known to be set if set in either the LHS | RHS.
954 KnownOne |= KnownOne2;
956 case Instruction::Xor: {
957 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
958 KnownZero, KnownOne, Depth+1))
960 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
961 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
962 KnownZero2, KnownOne2, Depth+1))
964 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
966 // If all of the demanded bits are known zero on one side, return the other.
967 // These bits cannot contribute to the result of the 'xor'.
968 if ((DemandedMask & KnownZero) == DemandedMask)
969 return UpdateValueUsesWith(I, I->getOperand(0));
970 if ((DemandedMask & KnownZero2) == DemandedMask)
971 return UpdateValueUsesWith(I, I->getOperand(1));
973 // Output known-0 bits are known if clear or set in both the LHS & RHS.
974 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
975 // Output known-1 are known to be set if set in only one of the LHS, RHS.
976 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
978 // If all of the demanded bits are known to be zero on one side or the
979 // other, turn this into an *inclusive* or.
980 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
981 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
983 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
985 InsertNewInstBefore(Or, *I);
986 return UpdateValueUsesWith(I, Or);
989 // If all of the demanded bits on one side are known, and all of the set
990 // bits on that side are also known to be set on the other side, turn this
991 // into an AND, as we know the bits will be cleared.
992 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
993 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
994 if ((KnownOne & KnownOne2) == KnownOne) {
995 Constant *AndC = ConstantInt::get(VTy, ~KnownOne & DemandedMask);
997 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
998 InsertNewInstBefore(And, *I);
999 return UpdateValueUsesWith(I, And);
1003 // If the RHS is a constant, see if we can simplify it.
1004 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1005 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1006 return UpdateValueUsesWith(I, I);
1008 KnownZero = KnownZeroOut;
1009 KnownOne = KnownOneOut;
1012 case Instruction::Select:
1013 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1014 KnownZero, KnownOne, Depth+1))
1016 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1017 KnownZero2, KnownOne2, Depth+1))
1019 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1020 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1022 // If the operands are constants, see if we can simplify them.
1023 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1024 return UpdateValueUsesWith(I, I);
1025 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1026 return UpdateValueUsesWith(I, I);
1028 // Only known if known in both the LHS and RHS.
1029 KnownOne &= KnownOne2;
1030 KnownZero &= KnownZero2;
1032 case Instruction::Trunc:
1033 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1034 KnownZero, KnownOne, Depth+1))
1036 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1038 case Instruction::BitCast:
1039 if (!I->getOperand(0)->getType()->isInteger())
1042 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1043 KnownZero, KnownOne, Depth+1))
1045 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1047 case Instruction::ZExt: {
1048 // Compute the bits in the result that are not present in the input.
1049 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1050 uint64_t NotIn = ~SrcTy->getBitMask();
1051 uint64_t NewBits = VTy->getBitMask() & NotIn;
1053 DemandedMask &= SrcTy->getBitMask();
1054 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1055 KnownZero, KnownOne, Depth+1))
1057 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1058 // The top bits are known to be zero.
1059 KnownZero |= NewBits;
1062 case Instruction::SExt: {
1063 // Compute the bits in the result that are not present in the input.
1064 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1065 uint64_t NotIn = ~SrcTy->getBitMask();
1066 uint64_t NewBits = VTy->getBitMask() & NotIn;
1068 // Get the sign bit for the source type
1069 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1070 int64_t InputDemandedBits = DemandedMask & SrcTy->getBitMask();
1072 // If any of the sign extended bits are demanded, we know that the sign
1074 if (NewBits & DemandedMask)
1075 InputDemandedBits |= InSignBit;
1077 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1078 KnownZero, KnownOne, Depth+1))
1080 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1082 // If the sign bit of the input is known set or clear, then we know the
1083 // top bits of the result.
1085 // If the input sign bit is known zero, or if the NewBits are not demanded
1086 // convert this into a zero extension.
1087 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1088 // Convert to ZExt cast
1089 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1090 return UpdateValueUsesWith(I, NewCast);
1091 } else if (KnownOne & InSignBit) { // Input sign bit known set
1092 KnownOne |= NewBits;
1093 KnownZero &= ~NewBits;
1094 } else { // Input sign bit unknown
1095 KnownZero &= ~NewBits;
1096 KnownOne &= ~NewBits;
1100 case Instruction::Add:
1101 // If there is a constant on the RHS, there are a variety of xformations
1103 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1104 // If null, this should be simplified elsewhere. Some of the xforms here
1105 // won't work if the RHS is zero.
1106 if (RHS->isNullValue())
1109 // Figure out what the input bits are. If the top bits of the and result
1110 // are not demanded, then the add doesn't demand them from its input
1113 // Shift the demanded mask up so that it's at the top of the uint64_t.
1114 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
1115 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1117 // If the top bit of the output is demanded, demand everything from the
1118 // input. Otherwise, we demand all the input bits except NLZ top bits.
1119 uint64_t InDemandedBits = ~0ULL >> (64-BitWidth+NLZ);
1121 // Find information about known zero/one bits in the input.
1122 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1123 KnownZero2, KnownOne2, Depth+1))
1126 // If the RHS of the add has bits set that can't affect the input, reduce
1128 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1129 return UpdateValueUsesWith(I, I);
1131 // Avoid excess work.
1132 if (KnownZero2 == 0 && KnownOne2 == 0)
1135 // Turn it into OR if input bits are zero.
1136 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1138 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1140 InsertNewInstBefore(Or, *I);
1141 return UpdateValueUsesWith(I, Or);
1144 // We can say something about the output known-zero and known-one bits,
1145 // depending on potential carries from the input constant and the
1146 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1147 // bits set and the RHS constant is 0x01001, then we know we have a known
1148 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1150 // To compute this, we first compute the potential carry bits. These are
1151 // the bits which may be modified. I'm not aware of a better way to do
1153 uint64_t RHSVal = RHS->getZExtValue();
1155 bool CarryIn = false;
1156 uint64_t CarryBits = 0;
1157 uint64_t CurBit = 1;
1158 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1159 // Record the current carry in.
1160 if (CarryIn) CarryBits |= CurBit;
1164 // This bit has a carry out unless it is "zero + zero" or
1165 // "zero + anything" with no carry in.
1166 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1167 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1168 } else if (!CarryIn &&
1169 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1170 CarryOut = false; // 0 + anything has no carry out if no carry in.
1172 // Otherwise, we have to assume we have a carry out.
1176 // This stage's carry out becomes the next stage's carry-in.
1180 // Now that we know which bits have carries, compute the known-1/0 sets.
1182 // Bits are known one if they are known zero in one operand and one in the
1183 // other, and there is no input carry.
1184 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1186 // Bits are known zero if they are known zero in both operands and there
1187 // is no input carry.
1188 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1190 // If the high-bits of this ADD are not demanded, then it does not demand
1191 // the high bits of its LHS or RHS.
1192 if ((DemandedMask & VTy->getSignBit()) == 0) {
1193 // Right fill the mask of bits for this ADD to demand the most
1194 // significant bit and all those below it.
1195 unsigned NLZ = CountLeadingZeros_64(DemandedMask);
1196 uint64_t DemandedFromOps = ~0ULL >> NLZ;
1197 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1198 KnownZero2, KnownOne2, Depth+1))
1200 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1201 KnownZero2, KnownOne2, Depth+1))
1206 case Instruction::Sub:
1207 // If the high-bits of this SUB are not demanded, then it does not demand
1208 // the high bits of its LHS or RHS.
1209 if ((DemandedMask & VTy->getSignBit()) == 0) {
1210 // Right fill the mask of bits for this SUB to demand the most
1211 // significant bit and all those below it.
1212 unsigned NLZ = CountLeadingZeros_64(DemandedMask);
1213 uint64_t DemandedFromOps = ~0ULL >> NLZ;
1214 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1215 KnownZero2, KnownOne2, Depth+1))
1217 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1218 KnownZero2, KnownOne2, Depth+1))
1222 case Instruction::Shl:
1223 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1224 uint64_t ShiftAmt = SA->getZExtValue();
1225 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1226 KnownZero, KnownOne, Depth+1))
1228 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1229 KnownZero <<= ShiftAmt;
1230 KnownOne <<= ShiftAmt;
1231 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1234 case Instruction::LShr:
1235 // For a logical shift right
1236 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1237 unsigned ShiftAmt = SA->getZExtValue();
1239 // Compute the new bits that are at the top now.
1240 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1241 HighBits <<= VTy->getBitWidth() - ShiftAmt;
1242 uint64_t TypeMask = VTy->getBitMask();
1243 // Unsigned shift right.
1244 if (SimplifyDemandedBits(I->getOperand(0),
1245 (DemandedMask << ShiftAmt) & TypeMask,
1246 KnownZero, KnownOne, Depth+1))
1248 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1249 KnownZero &= TypeMask;
1250 KnownOne &= TypeMask;
1251 KnownZero >>= ShiftAmt;
1252 KnownOne >>= ShiftAmt;
1253 KnownZero |= HighBits; // high bits known zero.
1256 case Instruction::AShr:
1257 // If this is an arithmetic shift right and only the low-bit is set, we can
1258 // always convert this into a logical shr, even if the shift amount is
1259 // variable. The low bit of the shift cannot be an input sign bit unless
1260 // the shift amount is >= the size of the datatype, which is undefined.
1261 if (DemandedMask == 1) {
1262 // Perform the logical shift right.
1263 Value *NewVal = BinaryOperator::createLShr(
1264 I->getOperand(0), I->getOperand(1), I->getName());
1265 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1266 return UpdateValueUsesWith(I, NewVal);
1269 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1270 unsigned ShiftAmt = SA->getZExtValue();
1272 // Compute the new bits that are at the top now.
1273 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1274 HighBits <<= VTy->getBitWidth() - ShiftAmt;
1275 uint64_t TypeMask = VTy->getBitMask();
1276 // Signed shift right.
1277 if (SimplifyDemandedBits(I->getOperand(0),
1278 (DemandedMask << ShiftAmt) & TypeMask,
1279 KnownZero, KnownOne, Depth+1))
1281 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1282 KnownZero &= TypeMask;
1283 KnownOne &= TypeMask;
1284 KnownZero >>= ShiftAmt;
1285 KnownOne >>= ShiftAmt;
1287 // Handle the sign bits.
1288 uint64_t SignBit = 1ULL << (VTy->getBitWidth()-1);
1289 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1291 // If the input sign bit is known to be zero, or if none of the top bits
1292 // are demanded, turn this into an unsigned shift right.
1293 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1294 // Perform the logical shift right.
1295 Value *NewVal = BinaryOperator::createLShr(
1296 I->getOperand(0), SA, I->getName());
1297 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1298 return UpdateValueUsesWith(I, NewVal);
1299 } else if (KnownOne & SignBit) { // New bits are known one.
1300 KnownOne |= HighBits;
1306 // If the client is only demanding bits that we know, return the known
1308 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1309 return UpdateValueUsesWith(I, ConstantInt::get(VTy, KnownOne));
1314 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1315 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1316 /// actually used by the caller. This method analyzes which elements of the
1317 /// operand are undef and returns that information in UndefElts.
1319 /// If the information about demanded elements can be used to simplify the
1320 /// operation, the operation is simplified, then the resultant value is
1321 /// returned. This returns null if no change was made.
1322 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1323 uint64_t &UndefElts,
1325 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1326 assert(VWidth <= 64 && "Vector too wide to analyze!");
1327 uint64_t EltMask = ~0ULL >> (64-VWidth);
1328 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1329 "Invalid DemandedElts!");
1331 if (isa<UndefValue>(V)) {
1332 // If the entire vector is undefined, just return this info.
1333 UndefElts = EltMask;
1335 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1336 UndefElts = EltMask;
1337 return UndefValue::get(V->getType());
1341 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1342 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1343 Constant *Undef = UndefValue::get(EltTy);
1345 std::vector<Constant*> Elts;
1346 for (unsigned i = 0; i != VWidth; ++i)
1347 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1348 Elts.push_back(Undef);
1349 UndefElts |= (1ULL << i);
1350 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1351 Elts.push_back(Undef);
1352 UndefElts |= (1ULL << i);
1353 } else { // Otherwise, defined.
1354 Elts.push_back(CP->getOperand(i));
1357 // If we changed the constant, return it.
1358 Constant *NewCP = ConstantVector::get(Elts);
1359 return NewCP != CP ? NewCP : 0;
1360 } else if (isa<ConstantAggregateZero>(V)) {
1361 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1363 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1364 Constant *Zero = Constant::getNullValue(EltTy);
1365 Constant *Undef = UndefValue::get(EltTy);
1366 std::vector<Constant*> Elts;
1367 for (unsigned i = 0; i != VWidth; ++i)
1368 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1369 UndefElts = DemandedElts ^ EltMask;
1370 return ConstantVector::get(Elts);
1373 if (!V->hasOneUse()) { // Other users may use these bits.
1374 if (Depth != 0) { // Not at the root.
1375 // TODO: Just compute the UndefElts information recursively.
1379 } else if (Depth == 10) { // Limit search depth.
1383 Instruction *I = dyn_cast<Instruction>(V);
1384 if (!I) return false; // Only analyze instructions.
1386 bool MadeChange = false;
1387 uint64_t UndefElts2;
1389 switch (I->getOpcode()) {
1392 case Instruction::InsertElement: {
1393 // If this is a variable index, we don't know which element it overwrites.
1394 // demand exactly the same input as we produce.
1395 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1397 // Note that we can't propagate undef elt info, because we don't know
1398 // which elt is getting updated.
1399 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1400 UndefElts2, Depth+1);
1401 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1405 // If this is inserting an element that isn't demanded, remove this
1407 unsigned IdxNo = Idx->getZExtValue();
1408 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1409 return AddSoonDeadInstToWorklist(*I, 0);
1411 // Otherwise, the element inserted overwrites whatever was there, so the
1412 // input demanded set is simpler than the output set.
1413 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1414 DemandedElts & ~(1ULL << IdxNo),
1415 UndefElts, Depth+1);
1416 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1418 // The inserted element is defined.
1419 UndefElts |= 1ULL << IdxNo;
1423 case Instruction::And:
1424 case Instruction::Or:
1425 case Instruction::Xor:
1426 case Instruction::Add:
1427 case Instruction::Sub:
1428 case Instruction::Mul:
1429 // div/rem demand all inputs, because they don't want divide by zero.
1430 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1431 UndefElts, Depth+1);
1432 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1433 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1434 UndefElts2, Depth+1);
1435 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1437 // Output elements are undefined if both are undefined. Consider things
1438 // like undef&0. The result is known zero, not undef.
1439 UndefElts &= UndefElts2;
1442 case Instruction::Call: {
1443 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1445 switch (II->getIntrinsicID()) {
1448 // Binary vector operations that work column-wise. A dest element is a
1449 // function of the corresponding input elements from the two inputs.
1450 case Intrinsic::x86_sse_sub_ss:
1451 case Intrinsic::x86_sse_mul_ss:
1452 case Intrinsic::x86_sse_min_ss:
1453 case Intrinsic::x86_sse_max_ss:
1454 case Intrinsic::x86_sse2_sub_sd:
1455 case Intrinsic::x86_sse2_mul_sd:
1456 case Intrinsic::x86_sse2_min_sd:
1457 case Intrinsic::x86_sse2_max_sd:
1458 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1459 UndefElts, Depth+1);
1460 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1461 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1462 UndefElts2, Depth+1);
1463 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1465 // If only the low elt is demanded and this is a scalarizable intrinsic,
1466 // scalarize it now.
1467 if (DemandedElts == 1) {
1468 switch (II->getIntrinsicID()) {
1470 case Intrinsic::x86_sse_sub_ss:
1471 case Intrinsic::x86_sse_mul_ss:
1472 case Intrinsic::x86_sse2_sub_sd:
1473 case Intrinsic::x86_sse2_mul_sd:
1474 // TODO: Lower MIN/MAX/ABS/etc
1475 Value *LHS = II->getOperand(1);
1476 Value *RHS = II->getOperand(2);
1477 // Extract the element as scalars.
1478 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1479 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1481 switch (II->getIntrinsicID()) {
1482 default: assert(0 && "Case stmts out of sync!");
1483 case Intrinsic::x86_sse_sub_ss:
1484 case Intrinsic::x86_sse2_sub_sd:
1485 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1486 II->getName()), *II);
1488 case Intrinsic::x86_sse_mul_ss:
1489 case Intrinsic::x86_sse2_mul_sd:
1490 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1491 II->getName()), *II);
1496 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1498 InsertNewInstBefore(New, *II);
1499 AddSoonDeadInstToWorklist(*II, 0);
1504 // Output elements are undefined if both are undefined. Consider things
1505 // like undef&0. The result is known zero, not undef.
1506 UndefElts &= UndefElts2;
1512 return MadeChange ? I : 0;
1515 /// @returns true if the specified compare instruction is
1516 /// true when both operands are equal...
1517 /// @brief Determine if the ICmpInst returns true if both operands are equal
1518 static bool isTrueWhenEqual(ICmpInst &ICI) {
1519 ICmpInst::Predicate pred = ICI.getPredicate();
1520 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1521 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1522 pred == ICmpInst::ICMP_SLE;
1525 /// AssociativeOpt - Perform an optimization on an associative operator. This
1526 /// function is designed to check a chain of associative operators for a
1527 /// potential to apply a certain optimization. Since the optimization may be
1528 /// applicable if the expression was reassociated, this checks the chain, then
1529 /// reassociates the expression as necessary to expose the optimization
1530 /// opportunity. This makes use of a special Functor, which must define
1531 /// 'shouldApply' and 'apply' methods.
1533 template<typename Functor>
1534 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1535 unsigned Opcode = Root.getOpcode();
1536 Value *LHS = Root.getOperand(0);
1538 // Quick check, see if the immediate LHS matches...
1539 if (F.shouldApply(LHS))
1540 return F.apply(Root);
1542 // Otherwise, if the LHS is not of the same opcode as the root, return.
1543 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1544 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1545 // Should we apply this transform to the RHS?
1546 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1548 // If not to the RHS, check to see if we should apply to the LHS...
1549 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1550 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1554 // If the functor wants to apply the optimization to the RHS of LHSI,
1555 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1557 BasicBlock *BB = Root.getParent();
1559 // Now all of the instructions are in the current basic block, go ahead
1560 // and perform the reassociation.
1561 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1563 // First move the selected RHS to the LHS of the root...
1564 Root.setOperand(0, LHSI->getOperand(1));
1566 // Make what used to be the LHS of the root be the user of the root...
1567 Value *ExtraOperand = TmpLHSI->getOperand(1);
1568 if (&Root == TmpLHSI) {
1569 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1572 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1573 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1574 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1575 BasicBlock::iterator ARI = &Root; ++ARI;
1576 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1579 // Now propagate the ExtraOperand down the chain of instructions until we
1581 while (TmpLHSI != LHSI) {
1582 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1583 // Move the instruction to immediately before the chain we are
1584 // constructing to avoid breaking dominance properties.
1585 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1586 BB->getInstList().insert(ARI, NextLHSI);
1589 Value *NextOp = NextLHSI->getOperand(1);
1590 NextLHSI->setOperand(1, ExtraOperand);
1592 ExtraOperand = NextOp;
1595 // Now that the instructions are reassociated, have the functor perform
1596 // the transformation...
1597 return F.apply(Root);
1600 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1606 // AddRHS - Implements: X + X --> X << 1
1609 AddRHS(Value *rhs) : RHS(rhs) {}
1610 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1611 Instruction *apply(BinaryOperator &Add) const {
1612 return BinaryOperator::createShl(Add.getOperand(0),
1613 ConstantInt::get(Add.getType(), 1));
1617 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1619 struct AddMaskingAnd {
1621 AddMaskingAnd(Constant *c) : C2(c) {}
1622 bool shouldApply(Value *LHS) const {
1624 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1625 ConstantExpr::getAnd(C1, C2)->isNullValue();
1627 Instruction *apply(BinaryOperator &Add) const {
1628 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1632 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1634 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1635 if (Constant *SOC = dyn_cast<Constant>(SO))
1636 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1638 return IC->InsertNewInstBefore(CastInst::create(
1639 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1642 // Figure out if the constant is the left or the right argument.
1643 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1644 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1646 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1648 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1649 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1652 Value *Op0 = SO, *Op1 = ConstOperand;
1654 std::swap(Op0, Op1);
1656 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1657 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1658 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1659 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1660 SO->getName()+".cmp");
1662 assert(0 && "Unknown binary instruction type!");
1665 return IC->InsertNewInstBefore(New, I);
1668 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1669 // constant as the other operand, try to fold the binary operator into the
1670 // select arguments. This also works for Cast instructions, which obviously do
1671 // not have a second operand.
1672 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1674 // Don't modify shared select instructions
1675 if (!SI->hasOneUse()) return 0;
1676 Value *TV = SI->getOperand(1);
1677 Value *FV = SI->getOperand(2);
1679 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1680 // Bool selects with constant operands can be folded to logical ops.
1681 if (SI->getType() == Type::Int1Ty) return 0;
1683 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1684 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1686 return new SelectInst(SI->getCondition(), SelectTrueVal,
1693 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1694 /// node as operand #0, see if we can fold the instruction into the PHI (which
1695 /// is only possible if all operands to the PHI are constants).
1696 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1697 PHINode *PN = cast<PHINode>(I.getOperand(0));
1698 unsigned NumPHIValues = PN->getNumIncomingValues();
1699 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1701 // Check to see if all of the operands of the PHI are constants. If there is
1702 // one non-constant value, remember the BB it is. If there is more than one
1703 // or if *it* is a PHI, bail out.
1704 BasicBlock *NonConstBB = 0;
1705 for (unsigned i = 0; i != NumPHIValues; ++i)
1706 if (!isa<Constant>(PN->getIncomingValue(i))) {
1707 if (NonConstBB) return 0; // More than one non-const value.
1708 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1709 NonConstBB = PN->getIncomingBlock(i);
1711 // If the incoming non-constant value is in I's block, we have an infinite
1713 if (NonConstBB == I.getParent())
1717 // If there is exactly one non-constant value, we can insert a copy of the
1718 // operation in that block. However, if this is a critical edge, we would be
1719 // inserting the computation one some other paths (e.g. inside a loop). Only
1720 // do this if the pred block is unconditionally branching into the phi block.
1722 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1723 if (!BI || !BI->isUnconditional()) return 0;
1726 // Okay, we can do the transformation: create the new PHI node.
1727 PHINode *NewPN = new PHINode(I.getType(), "");
1728 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1729 InsertNewInstBefore(NewPN, *PN);
1730 NewPN->takeName(PN);
1732 // Next, add all of the operands to the PHI.
1733 if (I.getNumOperands() == 2) {
1734 Constant *C = cast<Constant>(I.getOperand(1));
1735 for (unsigned i = 0; i != NumPHIValues; ++i) {
1737 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1738 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1739 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1741 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1743 assert(PN->getIncomingBlock(i) == NonConstBB);
1744 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1745 InV = BinaryOperator::create(BO->getOpcode(),
1746 PN->getIncomingValue(i), C, "phitmp",
1747 NonConstBB->getTerminator());
1748 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1749 InV = CmpInst::create(CI->getOpcode(),
1751 PN->getIncomingValue(i), C, "phitmp",
1752 NonConstBB->getTerminator());
1754 assert(0 && "Unknown binop!");
1756 AddToWorkList(cast<Instruction>(InV));
1758 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1761 CastInst *CI = cast<CastInst>(&I);
1762 const Type *RetTy = CI->getType();
1763 for (unsigned i = 0; i != NumPHIValues; ++i) {
1765 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1766 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1768 assert(PN->getIncomingBlock(i) == NonConstBB);
1769 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1770 I.getType(), "phitmp",
1771 NonConstBB->getTerminator());
1772 AddToWorkList(cast<Instruction>(InV));
1774 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1777 return ReplaceInstUsesWith(I, NewPN);
1780 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1781 bool Changed = SimplifyCommutative(I);
1782 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1784 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1785 // X + undef -> undef
1786 if (isa<UndefValue>(RHS))
1787 return ReplaceInstUsesWith(I, RHS);
1790 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1791 if (RHSC->isNullValue())
1792 return ReplaceInstUsesWith(I, LHS);
1793 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1794 if (CFP->isExactlyValue(-0.0))
1795 return ReplaceInstUsesWith(I, LHS);
1798 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1799 // X + (signbit) --> X ^ signbit
1800 uint64_t Val = CI->getZExtValue();
1801 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1802 return BinaryOperator::createXor(LHS, RHS);
1804 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1805 // (X & 254)+1 -> (X&254)|1
1806 uint64_t KnownZero, KnownOne;
1807 if (!isa<VectorType>(I.getType()) &&
1808 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
1809 KnownZero, KnownOne))
1813 if (isa<PHINode>(LHS))
1814 if (Instruction *NV = FoldOpIntoPhi(I))
1817 ConstantInt *XorRHS = 0;
1819 if (isa<ConstantInt>(RHSC) &&
1820 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1821 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1822 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1823 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1825 uint64_t C0080Val = 1ULL << 31;
1826 int64_t CFF80Val = -C0080Val;
1829 if (TySizeBits > Size) {
1831 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1832 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1833 if (RHSSExt == CFF80Val) {
1834 if (XorRHS->getZExtValue() == C0080Val)
1836 } else if (RHSZExt == C0080Val) {
1837 if (XorRHS->getSExtValue() == CFF80Val)
1841 // This is a sign extend if the top bits are known zero.
1842 uint64_t Mask = ~0ULL;
1843 Mask <<= 64-(TySizeBits-Size);
1844 Mask &= cast<IntegerType>(XorLHS->getType())->getBitMask();
1845 if (!MaskedValueIsZero(XorLHS, Mask))
1846 Size = 0; // Not a sign ext, but can't be any others either.
1853 } while (Size >= 8);
1856 const Type *MiddleType = 0;
1859 case 32: MiddleType = Type::Int32Ty; break;
1860 case 16: MiddleType = Type::Int16Ty; break;
1861 case 8: MiddleType = Type::Int8Ty; break;
1864 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1865 InsertNewInstBefore(NewTrunc, I);
1866 return new SExtInst(NewTrunc, I.getType());
1872 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
1873 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1875 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1876 if (RHSI->getOpcode() == Instruction::Sub)
1877 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1878 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1880 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1881 if (LHSI->getOpcode() == Instruction::Sub)
1882 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1883 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1888 if (Value *V = dyn_castNegVal(LHS))
1889 return BinaryOperator::createSub(RHS, V);
1892 if (!isa<Constant>(RHS))
1893 if (Value *V = dyn_castNegVal(RHS))
1894 return BinaryOperator::createSub(LHS, V);
1898 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1899 if (X == RHS) // X*C + X --> X * (C+1)
1900 return BinaryOperator::createMul(RHS, AddOne(C2));
1902 // X*C1 + X*C2 --> X * (C1+C2)
1904 if (X == dyn_castFoldableMul(RHS, C1))
1905 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1908 // X + X*C --> X * (C+1)
1909 if (dyn_castFoldableMul(RHS, C2) == LHS)
1910 return BinaryOperator::createMul(LHS, AddOne(C2));
1912 // X + ~X --> -1 since ~X = -X-1
1913 if (dyn_castNotVal(LHS) == RHS ||
1914 dyn_castNotVal(RHS) == LHS)
1915 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1918 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1919 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1920 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
1923 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1925 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1926 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1927 return BinaryOperator::createSub(C, X);
1930 // (X & FF00) + xx00 -> (X+xx00) & FF00
1931 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1932 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1933 if (Anded == CRHS) {
1934 // See if all bits from the first bit set in the Add RHS up are included
1935 // in the mask. First, get the rightmost bit.
1936 uint64_t AddRHSV = CRHS->getZExtValue();
1938 // Form a mask of all bits from the lowest bit added through the top.
1939 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1940 AddRHSHighBits &= C2->getType()->getBitMask();
1942 // See if the and mask includes all of these bits.
1943 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1945 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1946 // Okay, the xform is safe. Insert the new add pronto.
1947 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1948 LHS->getName()), I);
1949 return BinaryOperator::createAnd(NewAdd, C2);
1954 // Try to fold constant add into select arguments.
1955 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1956 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1960 // add (cast *A to intptrtype) B ->
1961 // cast (GEP (cast *A to sbyte*) B) ->
1964 CastInst *CI = dyn_cast<CastInst>(LHS);
1967 CI = dyn_cast<CastInst>(RHS);
1970 if (CI && CI->getType()->isSized() &&
1971 (CI->getType()->getPrimitiveSizeInBits() ==
1972 TD->getIntPtrType()->getPrimitiveSizeInBits())
1973 && isa<PointerType>(CI->getOperand(0)->getType())) {
1974 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1975 PointerType::get(Type::Int8Ty), I);
1976 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1977 return new PtrToIntInst(I2, CI->getType());
1981 return Changed ? &I : 0;
1984 // isSignBit - Return true if the value represented by the constant only has the
1985 // highest order bit set.
1986 static bool isSignBit(ConstantInt *CI) {
1987 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1988 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1991 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1992 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1994 if (Op0 == Op1) // sub X, X -> 0
1995 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1997 // If this is a 'B = x-(-A)', change to B = x+A...
1998 if (Value *V = dyn_castNegVal(Op1))
1999 return BinaryOperator::createAdd(Op0, V);
2001 if (isa<UndefValue>(Op0))
2002 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2003 if (isa<UndefValue>(Op1))
2004 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2006 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2007 // Replace (-1 - A) with (~A)...
2008 if (C->isAllOnesValue())
2009 return BinaryOperator::createNot(Op1);
2011 // C - ~X == X + (1+C)
2013 if (match(Op1, m_Not(m_Value(X))))
2014 return BinaryOperator::createAdd(X,
2015 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
2016 // -(X >>u 31) -> (X >>s 31)
2017 // -(X >>s 31) -> (X >>u 31)
2018 if (C->isNullValue()) {
2019 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2020 if (SI->getOpcode() == Instruction::LShr) {
2021 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2022 // Check to see if we are shifting out everything but the sign bit.
2023 if (CU->getZExtValue() ==
2024 SI->getType()->getPrimitiveSizeInBits()-1) {
2025 // Ok, the transformation is safe. Insert AShr.
2026 return BinaryOperator::create(Instruction::AShr,
2027 SI->getOperand(0), CU, SI->getName());
2031 else if (SI->getOpcode() == Instruction::AShr) {
2032 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2033 // Check to see if we are shifting out everything but the sign bit.
2034 if (CU->getZExtValue() ==
2035 SI->getType()->getPrimitiveSizeInBits()-1) {
2036 // Ok, the transformation is safe. Insert LShr.
2037 return BinaryOperator::createLShr(
2038 SI->getOperand(0), CU, SI->getName());
2044 // Try to fold constant sub into select arguments.
2045 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2046 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2049 if (isa<PHINode>(Op0))
2050 if (Instruction *NV = FoldOpIntoPhi(I))
2054 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2055 if (Op1I->getOpcode() == Instruction::Add &&
2056 !Op0->getType()->isFPOrFPVector()) {
2057 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2058 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2059 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2060 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2061 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2062 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2063 // C1-(X+C2) --> (C1-C2)-X
2064 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2065 Op1I->getOperand(0));
2069 if (Op1I->hasOneUse()) {
2070 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2071 // is not used by anyone else...
2073 if (Op1I->getOpcode() == Instruction::Sub &&
2074 !Op1I->getType()->isFPOrFPVector()) {
2075 // Swap the two operands of the subexpr...
2076 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2077 Op1I->setOperand(0, IIOp1);
2078 Op1I->setOperand(1, IIOp0);
2080 // Create the new top level add instruction...
2081 return BinaryOperator::createAdd(Op0, Op1);
2084 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2086 if (Op1I->getOpcode() == Instruction::And &&
2087 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2088 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2091 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2092 return BinaryOperator::createAnd(Op0, NewNot);
2095 // 0 - (X sdiv C) -> (X sdiv -C)
2096 if (Op1I->getOpcode() == Instruction::SDiv)
2097 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2098 if (CSI->isNullValue())
2099 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2100 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2101 ConstantExpr::getNeg(DivRHS));
2103 // X - X*C --> X * (1-C)
2104 ConstantInt *C2 = 0;
2105 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2107 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2108 return BinaryOperator::createMul(Op0, CP1);
2113 if (!Op0->getType()->isFPOrFPVector())
2114 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2115 if (Op0I->getOpcode() == Instruction::Add) {
2116 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2117 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2118 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2119 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2120 } else if (Op0I->getOpcode() == Instruction::Sub) {
2121 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2122 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2126 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2127 if (X == Op1) { // X*C - X --> X * (C-1)
2128 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2129 return BinaryOperator::createMul(Op1, CP1);
2132 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2133 if (X == dyn_castFoldableMul(Op1, C2))
2134 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2139 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2140 /// really just returns true if the most significant (sign) bit is set.
2141 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2143 case ICmpInst::ICMP_SLT:
2144 // True if LHS s< RHS and RHS == 0
2145 return RHS->isNullValue();
2146 case ICmpInst::ICMP_SLE:
2147 // True if LHS s<= RHS and RHS == -1
2148 return RHS->isAllOnesValue();
2149 case ICmpInst::ICMP_UGE:
2150 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2151 return RHS->getZExtValue() == (1ULL <<
2152 (RHS->getType()->getPrimitiveSizeInBits()-1));
2153 case ICmpInst::ICMP_UGT:
2154 // True if LHS u> RHS and RHS == high-bit-mask - 1
2155 return RHS->getZExtValue() ==
2156 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2162 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2163 bool Changed = SimplifyCommutative(I);
2164 Value *Op0 = I.getOperand(0);
2166 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2167 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2169 // Simplify mul instructions with a constant RHS...
2170 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2171 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2173 // ((X << C1)*C2) == (X * (C2 << C1))
2174 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2175 if (SI->getOpcode() == Instruction::Shl)
2176 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2177 return BinaryOperator::createMul(SI->getOperand(0),
2178 ConstantExpr::getShl(CI, ShOp));
2180 if (CI->isNullValue())
2181 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2182 if (CI->equalsInt(1)) // X * 1 == X
2183 return ReplaceInstUsesWith(I, Op0);
2184 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2185 return BinaryOperator::createNeg(Op0, I.getName());
2187 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2188 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2189 uint64_t C = Log2_64(Val);
2190 return BinaryOperator::createShl(Op0,
2191 ConstantInt::get(Op0->getType(), C));
2193 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2194 if (Op1F->isNullValue())
2195 return ReplaceInstUsesWith(I, Op1);
2197 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2198 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2199 if (Op1F->getValue() == 1.0)
2200 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2203 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2204 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2205 isa<ConstantInt>(Op0I->getOperand(1))) {
2206 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2207 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2209 InsertNewInstBefore(Add, I);
2210 Value *C1C2 = ConstantExpr::getMul(Op1,
2211 cast<Constant>(Op0I->getOperand(1)));
2212 return BinaryOperator::createAdd(Add, C1C2);
2216 // Try to fold constant mul into select arguments.
2217 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2218 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2221 if (isa<PHINode>(Op0))
2222 if (Instruction *NV = FoldOpIntoPhi(I))
2226 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2227 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2228 return BinaryOperator::createMul(Op0v, Op1v);
2230 // If one of the operands of the multiply is a cast from a boolean value, then
2231 // we know the bool is either zero or one, so this is a 'masking' multiply.
2232 // See if we can simplify things based on how the boolean was originally
2234 CastInst *BoolCast = 0;
2235 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2236 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2239 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2240 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2243 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2244 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2245 const Type *SCOpTy = SCIOp0->getType();
2247 // If the icmp is true iff the sign bit of X is set, then convert this
2248 // multiply into a shift/and combination.
2249 if (isa<ConstantInt>(SCIOp1) &&
2250 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2251 // Shift the X value right to turn it into "all signbits".
2252 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2253 SCOpTy->getPrimitiveSizeInBits()-1);
2255 InsertNewInstBefore(
2256 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2257 BoolCast->getOperand(0)->getName()+
2260 // If the multiply type is not the same as the source type, sign extend
2261 // or truncate to the multiply type.
2262 if (I.getType() != V->getType()) {
2263 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2264 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2265 Instruction::CastOps opcode =
2266 (SrcBits == DstBits ? Instruction::BitCast :
2267 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2268 V = InsertCastBefore(opcode, V, I.getType(), I);
2271 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2272 return BinaryOperator::createAnd(V, OtherOp);
2277 return Changed ? &I : 0;
2280 /// This function implements the transforms on div instructions that work
2281 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2282 /// used by the visitors to those instructions.
2283 /// @brief Transforms common to all three div instructions
2284 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2285 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2288 if (isa<UndefValue>(Op0))
2289 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2291 // X / undef -> undef
2292 if (isa<UndefValue>(Op1))
2293 return ReplaceInstUsesWith(I, Op1);
2295 // Handle cases involving: div X, (select Cond, Y, Z)
2296 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2297 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2298 // same basic block, then we replace the select with Y, and the condition
2299 // of the select with false (if the cond value is in the same BB). If the
2300 // select has uses other than the div, this allows them to be simplified
2301 // also. Note that div X, Y is just as good as div X, 0 (undef)
2302 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2303 if (ST->isNullValue()) {
2304 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2305 if (CondI && CondI->getParent() == I.getParent())
2306 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2307 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2308 I.setOperand(1, SI->getOperand(2));
2310 UpdateValueUsesWith(SI, SI->getOperand(2));
2314 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2315 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2316 if (ST->isNullValue()) {
2317 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2318 if (CondI && CondI->getParent() == I.getParent())
2319 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2320 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2321 I.setOperand(1, SI->getOperand(1));
2323 UpdateValueUsesWith(SI, SI->getOperand(1));
2331 /// This function implements the transforms common to both integer division
2332 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2333 /// division instructions.
2334 /// @brief Common integer divide transforms
2335 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2336 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2338 if (Instruction *Common = commonDivTransforms(I))
2341 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2343 if (RHS->equalsInt(1))
2344 return ReplaceInstUsesWith(I, Op0);
2346 // (X / C1) / C2 -> X / (C1*C2)
2347 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2348 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2349 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2350 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2351 ConstantExpr::getMul(RHS, LHSRHS));
2354 if (!RHS->isNullValue()) { // avoid X udiv 0
2355 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2356 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2358 if (isa<PHINode>(Op0))
2359 if (Instruction *NV = FoldOpIntoPhi(I))
2364 // 0 / X == 0, we don't need to preserve faults!
2365 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2366 if (LHS->equalsInt(0))
2367 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2372 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2373 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2375 // Handle the integer div common cases
2376 if (Instruction *Common = commonIDivTransforms(I))
2379 // X udiv C^2 -> X >> C
2380 // Check to see if this is an unsigned division with an exact power of 2,
2381 // if so, convert to a right shift.
2382 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2383 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2384 if (isPowerOf2_64(Val)) {
2385 uint64_t ShiftAmt = Log2_64(Val);
2386 return BinaryOperator::createLShr(Op0,
2387 ConstantInt::get(Op0->getType(), ShiftAmt));
2391 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2392 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2393 if (RHSI->getOpcode() == Instruction::Shl &&
2394 isa<ConstantInt>(RHSI->getOperand(0))) {
2395 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2396 if (isPowerOf2_64(C1)) {
2397 Value *N = RHSI->getOperand(1);
2398 const Type *NTy = N->getType();
2399 if (uint64_t C2 = Log2_64(C1)) {
2400 Constant *C2V = ConstantInt::get(NTy, C2);
2401 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2403 return BinaryOperator::createLShr(Op0, N);
2408 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2409 // where C1&C2 are powers of two.
2410 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2411 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2412 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2413 if (!STO->isNullValue() && !STO->isNullValue()) {
2414 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2415 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2416 // Compute the shift amounts
2417 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2418 // Construct the "on true" case of the select
2419 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2420 Instruction *TSI = BinaryOperator::createLShr(
2421 Op0, TC, SI->getName()+".t");
2422 TSI = InsertNewInstBefore(TSI, I);
2424 // Construct the "on false" case of the select
2425 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2426 Instruction *FSI = BinaryOperator::createLShr(
2427 Op0, FC, SI->getName()+".f");
2428 FSI = InsertNewInstBefore(FSI, I);
2430 // construct the select instruction and return it.
2431 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2438 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2439 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2441 // Handle the integer div common cases
2442 if (Instruction *Common = commonIDivTransforms(I))
2445 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2447 if (RHS->isAllOnesValue())
2448 return BinaryOperator::createNeg(Op0);
2451 if (Value *LHSNeg = dyn_castNegVal(Op0))
2452 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2455 // If the sign bits of both operands are zero (i.e. we can prove they are
2456 // unsigned inputs), turn this into a udiv.
2457 if (I.getType()->isInteger()) {
2458 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2459 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2460 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2467 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2468 return commonDivTransforms(I);
2471 /// GetFactor - If we can prove that the specified value is at least a multiple
2472 /// of some factor, return that factor.
2473 static Constant *GetFactor(Value *V) {
2474 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2477 // Unless we can be tricky, we know this is a multiple of 1.
2478 Constant *Result = ConstantInt::get(V->getType(), 1);
2480 Instruction *I = dyn_cast<Instruction>(V);
2481 if (!I) return Result;
2483 if (I->getOpcode() == Instruction::Mul) {
2484 // Handle multiplies by a constant, etc.
2485 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2486 GetFactor(I->getOperand(1)));
2487 } else if (I->getOpcode() == Instruction::Shl) {
2488 // (X<<C) -> X * (1 << C)
2489 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2490 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2491 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2493 } else if (I->getOpcode() == Instruction::And) {
2494 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2495 // X & 0xFFF0 is known to be a multiple of 16.
2496 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2497 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2498 return ConstantExpr::getShl(Result,
2499 ConstantInt::get(Result->getType(), Zeros));
2501 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2502 // Only handle int->int casts.
2503 if (!CI->isIntegerCast())
2505 Value *Op = CI->getOperand(0);
2506 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2511 /// This function implements the transforms on rem instructions that work
2512 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2513 /// is used by the visitors to those instructions.
2514 /// @brief Transforms common to all three rem instructions
2515 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2516 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2518 // 0 % X == 0, we don't need to preserve faults!
2519 if (Constant *LHS = dyn_cast<Constant>(Op0))
2520 if (LHS->isNullValue())
2521 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2523 if (isa<UndefValue>(Op0)) // undef % X -> 0
2524 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2525 if (isa<UndefValue>(Op1))
2526 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2528 // Handle cases involving: rem X, (select Cond, Y, Z)
2529 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2530 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2531 // the same basic block, then we replace the select with Y, and the
2532 // condition of the select with false (if the cond value is in the same
2533 // BB). If the select has uses other than the div, this allows them to be
2535 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2536 if (ST->isNullValue()) {
2537 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2538 if (CondI && CondI->getParent() == I.getParent())
2539 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2540 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2541 I.setOperand(1, SI->getOperand(2));
2543 UpdateValueUsesWith(SI, SI->getOperand(2));
2546 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2547 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2548 if (ST->isNullValue()) {
2549 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2550 if (CondI && CondI->getParent() == I.getParent())
2551 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2552 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2553 I.setOperand(1, SI->getOperand(1));
2555 UpdateValueUsesWith(SI, SI->getOperand(1));
2563 /// This function implements the transforms common to both integer remainder
2564 /// instructions (urem and srem). It is called by the visitors to those integer
2565 /// remainder instructions.
2566 /// @brief Common integer remainder transforms
2567 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2568 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2570 if (Instruction *common = commonRemTransforms(I))
2573 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2574 // X % 0 == undef, we don't need to preserve faults!
2575 if (RHS->equalsInt(0))
2576 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2578 if (RHS->equalsInt(1)) // X % 1 == 0
2579 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2581 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2582 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2583 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2585 } else if (isa<PHINode>(Op0I)) {
2586 if (Instruction *NV = FoldOpIntoPhi(I))
2589 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2590 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2591 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2598 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2599 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2601 if (Instruction *common = commonIRemTransforms(I))
2604 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2605 // X urem C^2 -> X and C
2606 // Check to see if this is an unsigned remainder with an exact power of 2,
2607 // if so, convert to a bitwise and.
2608 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2609 if (isPowerOf2_64(C->getZExtValue()))
2610 return BinaryOperator::createAnd(Op0, SubOne(C));
2613 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2614 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2615 if (RHSI->getOpcode() == Instruction::Shl &&
2616 isa<ConstantInt>(RHSI->getOperand(0))) {
2617 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2618 if (isPowerOf2_64(C1)) {
2619 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2620 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2622 return BinaryOperator::createAnd(Op0, Add);
2627 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2628 // where C1&C2 are powers of two.
2629 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2630 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2631 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2632 // STO == 0 and SFO == 0 handled above.
2633 if (isPowerOf2_64(STO->getZExtValue()) &&
2634 isPowerOf2_64(SFO->getZExtValue())) {
2635 Value *TrueAnd = InsertNewInstBefore(
2636 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2637 Value *FalseAnd = InsertNewInstBefore(
2638 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2639 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2647 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2648 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2650 if (Instruction *common = commonIRemTransforms(I))
2653 if (Value *RHSNeg = dyn_castNegVal(Op1))
2654 if (!isa<ConstantInt>(RHSNeg) ||
2655 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2657 AddUsesToWorkList(I);
2658 I.setOperand(1, RHSNeg);
2662 // If the top bits of both operands are zero (i.e. we can prove they are
2663 // unsigned inputs), turn this into a urem.
2664 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2665 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2666 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2667 return BinaryOperator::createURem(Op0, Op1, I.getName());
2673 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2674 return commonRemTransforms(I);
2677 // isMaxValueMinusOne - return true if this is Max-1
2678 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2680 // Calculate 0111111111..11111
2681 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2682 int64_t Val = INT64_MAX; // All ones
2683 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2684 return C->getSExtValue() == Val-1;
2686 return C->getZExtValue() == C->getType()->getBitMask()-1;
2689 // isMinValuePlusOne - return true if this is Min+1
2690 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2692 // Calculate 1111111111000000000000
2693 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2694 int64_t Val = -1; // All ones
2695 Val <<= TypeBits-1; // Shift over to the right spot
2696 return C->getSExtValue() == Val+1;
2698 return C->getZExtValue() == 1; // unsigned
2701 // isOneBitSet - Return true if there is exactly one bit set in the specified
2703 static bool isOneBitSet(const ConstantInt *CI) {
2704 uint64_t V = CI->getZExtValue();
2705 return V && (V & (V-1)) == 0;
2708 #if 0 // Currently unused
2709 // isLowOnes - Return true if the constant is of the form 0+1+.
2710 static bool isLowOnes(const ConstantInt *CI) {
2711 uint64_t V = CI->getZExtValue();
2713 // There won't be bits set in parts that the type doesn't contain.
2714 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2716 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2717 return U && V && (U & V) == 0;
2721 // isHighOnes - Return true if the constant is of the form 1+0+.
2722 // This is the same as lowones(~X).
2723 static bool isHighOnes(const ConstantInt *CI) {
2724 uint64_t V = ~CI->getZExtValue();
2725 if (~V == 0) return false; // 0's does not match "1+"
2727 // There won't be bits set in parts that the type doesn't contain.
2728 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2730 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2731 return U && V && (U & V) == 0;
2734 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2735 /// are carefully arranged to allow folding of expressions such as:
2737 /// (A < B) | (A > B) --> (A != B)
2739 /// Note that this is only valid if the first and second predicates have the
2740 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2742 /// Three bits are used to represent the condition, as follows:
2747 /// <=> Value Definition
2748 /// 000 0 Always false
2755 /// 111 7 Always true
2757 static unsigned getICmpCode(const ICmpInst *ICI) {
2758 switch (ICI->getPredicate()) {
2760 case ICmpInst::ICMP_UGT: return 1; // 001
2761 case ICmpInst::ICMP_SGT: return 1; // 001
2762 case ICmpInst::ICMP_EQ: return 2; // 010
2763 case ICmpInst::ICMP_UGE: return 3; // 011
2764 case ICmpInst::ICMP_SGE: return 3; // 011
2765 case ICmpInst::ICMP_ULT: return 4; // 100
2766 case ICmpInst::ICMP_SLT: return 4; // 100
2767 case ICmpInst::ICMP_NE: return 5; // 101
2768 case ICmpInst::ICMP_ULE: return 6; // 110
2769 case ICmpInst::ICMP_SLE: return 6; // 110
2772 assert(0 && "Invalid ICmp predicate!");
2777 /// getICmpValue - This is the complement of getICmpCode, which turns an
2778 /// opcode and two operands into either a constant true or false, or a brand
2779 /// new /// ICmp instruction. The sign is passed in to determine which kind
2780 /// of predicate to use in new icmp instructions.
2781 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2783 default: assert(0 && "Illegal ICmp code!");
2784 case 0: return ConstantInt::getFalse();
2787 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2789 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2790 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2793 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2795 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2798 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2800 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2801 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2804 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2806 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2807 case 7: return ConstantInt::getTrue();
2811 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2812 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2813 (ICmpInst::isSignedPredicate(p1) &&
2814 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2815 (ICmpInst::isSignedPredicate(p2) &&
2816 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2820 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2821 struct FoldICmpLogical {
2824 ICmpInst::Predicate pred;
2825 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2826 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2827 pred(ICI->getPredicate()) {}
2828 bool shouldApply(Value *V) const {
2829 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2830 if (PredicatesFoldable(pred, ICI->getPredicate()))
2831 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2832 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2835 Instruction *apply(Instruction &Log) const {
2836 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2837 if (ICI->getOperand(0) != LHS) {
2838 assert(ICI->getOperand(1) == LHS);
2839 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2842 unsigned LHSCode = getICmpCode(ICI);
2843 unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
2845 switch (Log.getOpcode()) {
2846 case Instruction::And: Code = LHSCode & RHSCode; break;
2847 case Instruction::Or: Code = LHSCode | RHSCode; break;
2848 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2849 default: assert(0 && "Illegal logical opcode!"); return 0;
2852 Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
2853 if (Instruction *I = dyn_cast<Instruction>(RV))
2855 // Otherwise, it's a constant boolean value...
2856 return IC.ReplaceInstUsesWith(Log, RV);
2859 } // end anonymous namespace
2861 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2862 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2863 // guaranteed to be a binary operator.
2864 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2866 ConstantInt *AndRHS,
2867 BinaryOperator &TheAnd) {
2868 Value *X = Op->getOperand(0);
2869 Constant *Together = 0;
2871 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2873 switch (Op->getOpcode()) {
2874 case Instruction::Xor:
2875 if (Op->hasOneUse()) {
2876 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2877 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2878 InsertNewInstBefore(And, TheAnd);
2880 return BinaryOperator::createXor(And, Together);
2883 case Instruction::Or:
2884 if (Together == AndRHS) // (X | C) & C --> C
2885 return ReplaceInstUsesWith(TheAnd, AndRHS);
2887 if (Op->hasOneUse() && Together != OpRHS) {
2888 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2889 Instruction *Or = BinaryOperator::createOr(X, Together);
2890 InsertNewInstBefore(Or, TheAnd);
2892 return BinaryOperator::createAnd(Or, AndRHS);
2895 case Instruction::Add:
2896 if (Op->hasOneUse()) {
2897 // Adding a one to a single bit bit-field should be turned into an XOR
2898 // of the bit. First thing to check is to see if this AND is with a
2899 // single bit constant.
2900 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2902 // Clear bits that are not part of the constant.
2903 AndRHSV &= AndRHS->getType()->getBitMask();
2905 // If there is only one bit set...
2906 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2907 // Ok, at this point, we know that we are masking the result of the
2908 // ADD down to exactly one bit. If the constant we are adding has
2909 // no bits set below this bit, then we can eliminate the ADD.
2910 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2912 // Check to see if any bits below the one bit set in AndRHSV are set.
2913 if ((AddRHS & (AndRHSV-1)) == 0) {
2914 // If not, the only thing that can effect the output of the AND is
2915 // the bit specified by AndRHSV. If that bit is set, the effect of
2916 // the XOR is to toggle the bit. If it is clear, then the ADD has
2918 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2919 TheAnd.setOperand(0, X);
2922 // Pull the XOR out of the AND.
2923 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
2924 InsertNewInstBefore(NewAnd, TheAnd);
2925 NewAnd->takeName(Op);
2926 return BinaryOperator::createXor(NewAnd, AndRHS);
2933 case Instruction::Shl: {
2934 // We know that the AND will not produce any of the bits shifted in, so if
2935 // the anded constant includes them, clear them now!
2937 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2938 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2939 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2941 if (CI == ShlMask) { // Masking out bits that the shift already masks
2942 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2943 } else if (CI != AndRHS) { // Reducing bits set in and.
2944 TheAnd.setOperand(1, CI);
2949 case Instruction::LShr:
2951 // We know that the AND will not produce any of the bits shifted in, so if
2952 // the anded constant includes them, clear them now! This only applies to
2953 // unsigned shifts, because a signed shr may bring in set bits!
2955 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2956 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2957 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2959 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2960 return ReplaceInstUsesWith(TheAnd, Op);
2961 } else if (CI != AndRHS) {
2962 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2967 case Instruction::AShr:
2969 // See if this is shifting in some sign extension, then masking it out
2971 if (Op->hasOneUse()) {
2972 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2973 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2974 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2975 if (C == AndRHS) { // Masking out bits shifted in.
2976 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2977 // Make the argument unsigned.
2978 Value *ShVal = Op->getOperand(0);
2979 ShVal = InsertNewInstBefore(
2980 BinaryOperator::createLShr(ShVal, OpRHS,
2981 Op->getName()), TheAnd);
2982 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2991 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2992 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2993 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2994 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2995 /// insert new instructions.
2996 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2997 bool isSigned, bool Inside,
2999 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3000 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3001 "Lo is not <= Hi in range emission code!");
3004 if (Lo == Hi) // Trivially false.
3005 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3007 // V >= Min && V < Hi --> V < Hi
3008 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3009 ICmpInst::Predicate pred = (isSigned ?
3010 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3011 return new ICmpInst(pred, V, Hi);
3014 // Emit V-Lo <u Hi-Lo
3015 Constant *NegLo = ConstantExpr::getNeg(Lo);
3016 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3017 InsertNewInstBefore(Add, IB);
3018 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3019 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3022 if (Lo == Hi) // Trivially true.
3023 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3025 // V < Min || V >= Hi ->'V > Hi-1'
3026 Hi = SubOne(cast<ConstantInt>(Hi));
3027 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3028 ICmpInst::Predicate pred = (isSigned ?
3029 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3030 return new ICmpInst(pred, V, Hi);
3033 // Emit V-Lo > Hi-1-Lo
3034 Constant *NegLo = ConstantExpr::getNeg(Lo);
3035 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3036 InsertNewInstBefore(Add, IB);
3037 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3038 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3041 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3042 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3043 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3044 // not, since all 1s are not contiguous.
3045 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
3046 uint64_t V = Val->getZExtValue();
3047 if (!isShiftedMask_64(V)) return false;
3049 // look for the first zero bit after the run of ones
3050 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
3051 // look for the first non-zero bit
3052 ME = 64-CountLeadingZeros_64(V);
3058 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3059 /// where isSub determines whether the operator is a sub. If we can fold one of
3060 /// the following xforms:
3062 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3063 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3064 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3066 /// return (A +/- B).
3068 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3069 ConstantInt *Mask, bool isSub,
3071 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3072 if (!LHSI || LHSI->getNumOperands() != 2 ||
3073 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3075 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3077 switch (LHSI->getOpcode()) {
3079 case Instruction::And:
3080 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3081 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3082 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3085 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3086 // part, we don't need any explicit masks to take them out of A. If that
3087 // is all N is, ignore it.
3089 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3090 uint64_t Mask = cast<IntegerType>(RHS->getType())->getBitMask();
3092 if (MaskedValueIsZero(RHS, Mask))
3097 case Instruction::Or:
3098 case Instruction::Xor:
3099 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3100 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3101 ConstantExpr::getAnd(N, Mask)->isNullValue())
3108 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3110 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3111 return InsertNewInstBefore(New, I);
3114 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3115 bool Changed = SimplifyCommutative(I);
3116 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3118 if (isa<UndefValue>(Op1)) // X & undef -> 0
3119 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3123 return ReplaceInstUsesWith(I, Op1);
3125 // See if we can simplify any instructions used by the instruction whose sole
3126 // purpose is to compute bits we don't care about.
3127 uint64_t KnownZero, KnownOne;
3128 if (!isa<VectorType>(I.getType())) {
3129 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3130 KnownZero, KnownOne))
3133 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3134 if (CP->isAllOnesValue())
3135 return ReplaceInstUsesWith(I, I.getOperand(0));
3139 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3140 uint64_t AndRHSMask = AndRHS->getZExtValue();
3141 uint64_t TypeMask = cast<IntegerType>(Op0->getType())->getBitMask();
3142 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3144 // Optimize a variety of ((val OP C1) & C2) combinations...
3145 if (isa<BinaryOperator>(Op0)) {
3146 Instruction *Op0I = cast<Instruction>(Op0);
3147 Value *Op0LHS = Op0I->getOperand(0);
3148 Value *Op0RHS = Op0I->getOperand(1);
3149 switch (Op0I->getOpcode()) {
3150 case Instruction::Xor:
3151 case Instruction::Or:
3152 // If the mask is only needed on one incoming arm, push it up.
3153 if (Op0I->hasOneUse()) {
3154 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3155 // Not masking anything out for the LHS, move to RHS.
3156 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3157 Op0RHS->getName()+".masked");
3158 InsertNewInstBefore(NewRHS, I);
3159 return BinaryOperator::create(
3160 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3162 if (!isa<Constant>(Op0RHS) &&
3163 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3164 // Not masking anything out for the RHS, move to LHS.
3165 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3166 Op0LHS->getName()+".masked");
3167 InsertNewInstBefore(NewLHS, I);
3168 return BinaryOperator::create(
3169 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3174 case Instruction::Add:
3175 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3176 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3177 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3178 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3179 return BinaryOperator::createAnd(V, AndRHS);
3180 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3181 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3184 case Instruction::Sub:
3185 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3186 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3187 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3188 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3189 return BinaryOperator::createAnd(V, AndRHS);
3193 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3194 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3196 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3197 // If this is an integer truncation or change from signed-to-unsigned, and
3198 // if the source is an and/or with immediate, transform it. This
3199 // frequently occurs for bitfield accesses.
3200 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3201 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3202 CastOp->getNumOperands() == 2)
3203 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3204 if (CastOp->getOpcode() == Instruction::And) {
3205 // Change: and (cast (and X, C1) to T), C2
3206 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3207 // This will fold the two constants together, which may allow
3208 // other simplifications.
3209 Instruction *NewCast = CastInst::createTruncOrBitCast(
3210 CastOp->getOperand(0), I.getType(),
3211 CastOp->getName()+".shrunk");
3212 NewCast = InsertNewInstBefore(NewCast, I);
3213 // trunc_or_bitcast(C1)&C2
3214 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3215 C3 = ConstantExpr::getAnd(C3, AndRHS);
3216 return BinaryOperator::createAnd(NewCast, C3);
3217 } else if (CastOp->getOpcode() == Instruction::Or) {
3218 // Change: and (cast (or X, C1) to T), C2
3219 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3220 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3221 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3222 return ReplaceInstUsesWith(I, AndRHS);
3227 // Try to fold constant and into select arguments.
3228 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3229 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3231 if (isa<PHINode>(Op0))
3232 if (Instruction *NV = FoldOpIntoPhi(I))
3236 Value *Op0NotVal = dyn_castNotVal(Op0);
3237 Value *Op1NotVal = dyn_castNotVal(Op1);
3239 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3240 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3242 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3243 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3244 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3245 I.getName()+".demorgan");
3246 InsertNewInstBefore(Or, I);
3247 return BinaryOperator::createNot(Or);
3251 Value *A = 0, *B = 0;
3252 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3253 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3254 return ReplaceInstUsesWith(I, Op1);
3255 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3256 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3257 return ReplaceInstUsesWith(I, Op0);
3259 if (Op0->hasOneUse() &&
3260 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3261 if (A == Op1) { // (A^B)&A -> A&(A^B)
3262 I.swapOperands(); // Simplify below
3263 std::swap(Op0, Op1);
3264 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3265 cast<BinaryOperator>(Op0)->swapOperands();
3266 I.swapOperands(); // Simplify below
3267 std::swap(Op0, Op1);
3270 if (Op1->hasOneUse() &&
3271 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3272 if (B == Op0) { // B&(A^B) -> B&(B^A)
3273 cast<BinaryOperator>(Op1)->swapOperands();
3276 if (A == Op0) { // A&(A^B) -> A & ~B
3277 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3278 InsertNewInstBefore(NotB, I);
3279 return BinaryOperator::createAnd(A, NotB);
3284 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3285 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3286 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3289 Value *LHSVal, *RHSVal;
3290 ConstantInt *LHSCst, *RHSCst;
3291 ICmpInst::Predicate LHSCC, RHSCC;
3292 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3293 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3294 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3295 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3296 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3297 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3298 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3299 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3300 // Ensure that the larger constant is on the RHS.
3301 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3302 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3303 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3304 ICmpInst *LHS = cast<ICmpInst>(Op0);
3305 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3306 std::swap(LHS, RHS);
3307 std::swap(LHSCst, RHSCst);
3308 std::swap(LHSCC, RHSCC);
3311 // At this point, we know we have have two icmp instructions
3312 // comparing a value against two constants and and'ing the result
3313 // together. Because of the above check, we know that we only have
3314 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3315 // (from the FoldICmpLogical check above), that the two constants
3316 // are not equal and that the larger constant is on the RHS
3317 assert(LHSCst != RHSCst && "Compares not folded above?");
3320 default: assert(0 && "Unknown integer condition code!");
3321 case ICmpInst::ICMP_EQ:
3323 default: assert(0 && "Unknown integer condition code!");
3324 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3325 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3326 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3327 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3328 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3329 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3330 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3331 return ReplaceInstUsesWith(I, LHS);
3333 case ICmpInst::ICMP_NE:
3335 default: assert(0 && "Unknown integer condition code!");
3336 case ICmpInst::ICMP_ULT:
3337 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3338 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3339 break; // (X != 13 & X u< 15) -> no change
3340 case ICmpInst::ICMP_SLT:
3341 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3342 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3343 break; // (X != 13 & X s< 15) -> no change
3344 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3345 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3346 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3347 return ReplaceInstUsesWith(I, RHS);
3348 case ICmpInst::ICMP_NE:
3349 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3350 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3351 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3352 LHSVal->getName()+".off");
3353 InsertNewInstBefore(Add, I);
3354 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3355 ConstantInt::get(Add->getType(), 1));
3357 break; // (X != 13 & X != 15) -> no change
3360 case ICmpInst::ICMP_ULT:
3362 default: assert(0 && "Unknown integer condition code!");
3363 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3364 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3365 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3366 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3368 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3369 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3370 return ReplaceInstUsesWith(I, LHS);
3371 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3375 case ICmpInst::ICMP_SLT:
3377 default: assert(0 && "Unknown integer condition code!");
3378 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3379 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3380 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3381 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3383 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3384 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3385 return ReplaceInstUsesWith(I, LHS);
3386 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3390 case ICmpInst::ICMP_UGT:
3392 default: assert(0 && "Unknown integer condition code!");
3393 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3394 return ReplaceInstUsesWith(I, LHS);
3395 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3396 return ReplaceInstUsesWith(I, RHS);
3397 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3399 case ICmpInst::ICMP_NE:
3400 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3401 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3402 break; // (X u> 13 & X != 15) -> no change
3403 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3404 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3406 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3410 case ICmpInst::ICMP_SGT:
3412 default: assert(0 && "Unknown integer condition code!");
3413 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3414 return ReplaceInstUsesWith(I, LHS);
3415 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3416 return ReplaceInstUsesWith(I, RHS);
3417 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3419 case ICmpInst::ICMP_NE:
3420 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3421 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3422 break; // (X s> 13 & X != 15) -> no change
3423 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3424 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3426 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3434 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3435 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3436 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3437 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3438 const Type *SrcTy = Op0C->getOperand(0)->getType();
3439 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3440 // Only do this if the casts both really cause code to be generated.
3441 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3443 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3445 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3446 Op1C->getOperand(0),
3448 InsertNewInstBefore(NewOp, I);
3449 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3453 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3454 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3455 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3456 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3457 SI0->getOperand(1) == SI1->getOperand(1) &&
3458 (SI0->hasOneUse() || SI1->hasOneUse())) {
3459 Instruction *NewOp =
3460 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3462 SI0->getName()), I);
3463 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3464 SI1->getOperand(1));
3468 return Changed ? &I : 0;
3471 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3472 /// in the result. If it does, and if the specified byte hasn't been filled in
3473 /// yet, fill it in and return false.
3474 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3475 Instruction *I = dyn_cast<Instruction>(V);
3476 if (I == 0) return true;
3478 // If this is an or instruction, it is an inner node of the bswap.
3479 if (I->getOpcode() == Instruction::Or)
3480 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3481 CollectBSwapParts(I->getOperand(1), ByteValues);
3483 // If this is a shift by a constant int, and it is "24", then its operand
3484 // defines a byte. We only handle unsigned types here.
3485 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3486 // Not shifting the entire input by N-1 bytes?
3487 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3488 8*(ByteValues.size()-1))
3492 if (I->getOpcode() == Instruction::Shl) {
3493 // X << 24 defines the top byte with the lowest of the input bytes.
3494 DestNo = ByteValues.size()-1;
3496 // X >>u 24 defines the low byte with the highest of the input bytes.
3500 // If the destination byte value is already defined, the values are or'd
3501 // together, which isn't a bswap (unless it's an or of the same bits).
3502 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3504 ByteValues[DestNo] = I->getOperand(0);
3508 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3510 Value *Shift = 0, *ShiftLHS = 0;
3511 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3512 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3513 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3515 Instruction *SI = cast<Instruction>(Shift);
3517 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3518 if (ShiftAmt->getZExtValue() & 7 ||
3519 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3522 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3524 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3525 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3527 // Unknown mask for bswap.
3528 if (DestByte == ByteValues.size()) return true;
3530 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3532 if (SI->getOpcode() == Instruction::Shl)
3533 SrcByte = DestByte - ShiftBytes;
3535 SrcByte = DestByte + ShiftBytes;
3537 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3538 if (SrcByte != ByteValues.size()-DestByte-1)
3541 // If the destination byte value is already defined, the values are or'd
3542 // together, which isn't a bswap (unless it's an or of the same bits).
3543 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3545 ByteValues[DestByte] = SI->getOperand(0);
3549 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3550 /// If so, insert the new bswap intrinsic and return it.
3551 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3552 // We cannot bswap one byte.
3553 if (I.getType() == Type::Int8Ty)
3556 /// ByteValues - For each byte of the result, we keep track of which value
3557 /// defines each byte.
3558 SmallVector<Value*, 8> ByteValues;
3559 ByteValues.resize(TD->getTypeSize(I.getType()));
3561 // Try to find all the pieces corresponding to the bswap.
3562 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3563 CollectBSwapParts(I.getOperand(1), ByteValues))
3566 // Check to see if all of the bytes come from the same value.
3567 Value *V = ByteValues[0];
3568 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3570 // Check to make sure that all of the bytes come from the same value.
3571 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3572 if (ByteValues[i] != V)
3575 // If they do then *success* we can turn this into a bswap. Figure out what
3576 // bswap to make it into.
3577 Module *M = I.getParent()->getParent()->getParent();
3578 const char *FnName = 0;
3579 if (I.getType() == Type::Int16Ty)
3580 FnName = "llvm.bswap.i16";
3581 else if (I.getType() == Type::Int32Ty)
3582 FnName = "llvm.bswap.i32";
3583 else if (I.getType() == Type::Int64Ty)
3584 FnName = "llvm.bswap.i64";
3586 assert(0 && "Unknown integer type!");
3587 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3588 return new CallInst(F, V);
3592 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3593 bool Changed = SimplifyCommutative(I);
3594 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3596 if (isa<UndefValue>(Op1))
3597 return ReplaceInstUsesWith(I, // X | undef -> -1
3598 ConstantInt::getAllOnesValue(I.getType()));
3602 return ReplaceInstUsesWith(I, Op0);
3604 // See if we can simplify any instructions used by the instruction whose sole
3605 // purpose is to compute bits we don't care about.
3606 uint64_t KnownZero, KnownOne;
3607 if (!isa<VectorType>(I.getType()) &&
3608 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3609 KnownZero, KnownOne))
3613 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3614 ConstantInt *C1 = 0; Value *X = 0;
3615 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3616 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3617 Instruction *Or = BinaryOperator::createOr(X, RHS);
3618 InsertNewInstBefore(Or, I);
3620 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3623 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3624 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3625 Instruction *Or = BinaryOperator::createOr(X, RHS);
3626 InsertNewInstBefore(Or, I);
3628 return BinaryOperator::createXor(Or,
3629 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3632 // Try to fold constant and into select arguments.
3633 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3634 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3636 if (isa<PHINode>(Op0))
3637 if (Instruction *NV = FoldOpIntoPhi(I))
3641 Value *A = 0, *B = 0;
3642 ConstantInt *C1 = 0, *C2 = 0;
3644 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3645 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3646 return ReplaceInstUsesWith(I, Op1);
3647 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3648 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3649 return ReplaceInstUsesWith(I, Op0);
3651 // (A | B) | C and A | (B | C) -> bswap if possible.
3652 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3653 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3654 match(Op1, m_Or(m_Value(), m_Value())) ||
3655 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3656 match(Op1, m_Shift(m_Value(), m_Value())))) {
3657 if (Instruction *BSwap = MatchBSwap(I))
3661 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3662 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3663 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3664 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3665 InsertNewInstBefore(NOr, I);
3667 return BinaryOperator::createXor(NOr, C1);
3670 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3671 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3672 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3673 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3674 InsertNewInstBefore(NOr, I);
3676 return BinaryOperator::createXor(NOr, C1);
3679 // (A & C1)|(B & C2)
3680 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3681 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3683 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3684 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3687 // If we have: ((V + N) & C1) | (V & C2)
3688 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3689 // replace with V+N.
3690 if (C1 == ConstantExpr::getNot(C2)) {
3691 Value *V1 = 0, *V2 = 0;
3692 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3693 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3694 // Add commutes, try both ways.
3695 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3696 return ReplaceInstUsesWith(I, A);
3697 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3698 return ReplaceInstUsesWith(I, A);
3700 // Or commutes, try both ways.
3701 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3702 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3703 // Add commutes, try both ways.
3704 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3705 return ReplaceInstUsesWith(I, B);
3706 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3707 return ReplaceInstUsesWith(I, B);
3712 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3713 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3714 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3715 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3716 SI0->getOperand(1) == SI1->getOperand(1) &&
3717 (SI0->hasOneUse() || SI1->hasOneUse())) {
3718 Instruction *NewOp =
3719 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3721 SI0->getName()), I);
3722 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3723 SI1->getOperand(1));
3727 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3728 if (A == Op1) // ~A | A == -1
3729 return ReplaceInstUsesWith(I,
3730 ConstantInt::getAllOnesValue(I.getType()));
3734 // Note, A is still live here!
3735 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3737 return ReplaceInstUsesWith(I,
3738 ConstantInt::getAllOnesValue(I.getType()));
3740 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3741 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3742 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3743 I.getName()+".demorgan"), I);
3744 return BinaryOperator::createNot(And);
3748 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3749 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3750 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3753 Value *LHSVal, *RHSVal;
3754 ConstantInt *LHSCst, *RHSCst;
3755 ICmpInst::Predicate LHSCC, RHSCC;
3756 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3757 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3758 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3759 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3760 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3761 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3762 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3763 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3764 // Ensure that the larger constant is on the RHS.
3765 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3766 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3767 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3768 ICmpInst *LHS = cast<ICmpInst>(Op0);
3769 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3770 std::swap(LHS, RHS);
3771 std::swap(LHSCst, RHSCst);
3772 std::swap(LHSCC, RHSCC);
3775 // At this point, we know we have have two icmp instructions
3776 // comparing a value against two constants and or'ing the result
3777 // together. Because of the above check, we know that we only have
3778 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3779 // FoldICmpLogical check above), that the two constants are not
3781 assert(LHSCst != RHSCst && "Compares not folded above?");
3784 default: assert(0 && "Unknown integer condition code!");
3785 case ICmpInst::ICMP_EQ:
3787 default: assert(0 && "Unknown integer condition code!");
3788 case ICmpInst::ICMP_EQ:
3789 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3790 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3791 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3792 LHSVal->getName()+".off");
3793 InsertNewInstBefore(Add, I);
3794 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3795 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3797 break; // (X == 13 | X == 15) -> no change
3798 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3799 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3801 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3802 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3803 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3804 return ReplaceInstUsesWith(I, RHS);
3807 case ICmpInst::ICMP_NE:
3809 default: assert(0 && "Unknown integer condition code!");
3810 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3811 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3812 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3813 return ReplaceInstUsesWith(I, LHS);
3814 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3815 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3816 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3817 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3820 case ICmpInst::ICMP_ULT:
3822 default: assert(0 && "Unknown integer condition code!");
3823 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3825 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3826 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3828 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3830 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3831 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3832 return ReplaceInstUsesWith(I, RHS);
3833 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3837 case ICmpInst::ICMP_SLT:
3839 default: assert(0 && "Unknown integer condition code!");
3840 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3842 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3843 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3845 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3847 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3848 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3849 return ReplaceInstUsesWith(I, RHS);
3850 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3854 case ICmpInst::ICMP_UGT:
3856 default: assert(0 && "Unknown integer condition code!");
3857 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3858 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3859 return ReplaceInstUsesWith(I, LHS);
3860 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3862 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3863 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3864 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3865 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3869 case ICmpInst::ICMP_SGT:
3871 default: assert(0 && "Unknown integer condition code!");
3872 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3873 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3874 return ReplaceInstUsesWith(I, LHS);
3875 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3877 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3878 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3879 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3880 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3888 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3889 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3890 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3891 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3892 const Type *SrcTy = Op0C->getOperand(0)->getType();
3893 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3894 // Only do this if the casts both really cause code to be generated.
3895 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3897 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3899 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3900 Op1C->getOperand(0),
3902 InsertNewInstBefore(NewOp, I);
3903 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3908 return Changed ? &I : 0;
3911 // XorSelf - Implements: X ^ X --> 0
3914 XorSelf(Value *rhs) : RHS(rhs) {}
3915 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3916 Instruction *apply(BinaryOperator &Xor) const {
3922 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3923 bool Changed = SimplifyCommutative(I);
3924 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3926 if (isa<UndefValue>(Op1))
3927 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3929 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3930 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3931 assert(Result == &I && "AssociativeOpt didn't work?");
3932 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3935 // See if we can simplify any instructions used by the instruction whose sole
3936 // purpose is to compute bits we don't care about.
3937 uint64_t KnownZero, KnownOne;
3938 if (!isa<VectorType>(I.getType()) &&
3939 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3940 KnownZero, KnownOne))
3943 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3944 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3945 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3946 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
3947 return new ICmpInst(ICI->getInversePredicate(),
3948 ICI->getOperand(0), ICI->getOperand(1));
3950 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3951 // ~(c-X) == X-c-1 == X+(-c-1)
3952 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3953 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3954 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3955 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3956 ConstantInt::get(I.getType(), 1));
3957 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3960 // ~(~X & Y) --> (X | ~Y)
3961 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3962 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3963 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3965 BinaryOperator::createNot(Op0I->getOperand(1),
3966 Op0I->getOperand(1)->getName()+".not");
3967 InsertNewInstBefore(NotY, I);
3968 return BinaryOperator::createOr(Op0NotVal, NotY);
3972 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3973 if (Op0I->getOpcode() == Instruction::Add) {
3974 // ~(X-c) --> (-c-1)-X
3975 if (RHS->isAllOnesValue()) {
3976 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3977 return BinaryOperator::createSub(
3978 ConstantExpr::getSub(NegOp0CI,
3979 ConstantInt::get(I.getType(), 1)),
3980 Op0I->getOperand(0));
3982 } else if (Op0I->getOpcode() == Instruction::Or) {
3983 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3984 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3985 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3986 // Anything in both C1 and C2 is known to be zero, remove it from
3988 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3989 NewRHS = ConstantExpr::getAnd(NewRHS,
3990 ConstantExpr::getNot(CommonBits));
3991 AddToWorkList(Op0I);
3992 I.setOperand(0, Op0I->getOperand(0));
3993 I.setOperand(1, NewRHS);
3999 // Try to fold constant and into select arguments.
4000 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4001 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4003 if (isa<PHINode>(Op0))
4004 if (Instruction *NV = FoldOpIntoPhi(I))
4008 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4010 return ReplaceInstUsesWith(I,
4011 ConstantInt::getAllOnesValue(I.getType()));
4013 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4015 return ReplaceInstUsesWith(I,
4016 ConstantInt::getAllOnesValue(I.getType()));
4018 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
4019 if (Op1I->getOpcode() == Instruction::Or) {
4020 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
4021 Op1I->swapOperands();
4023 std::swap(Op0, Op1);
4024 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
4025 I.swapOperands(); // Simplified below.
4026 std::swap(Op0, Op1);
4028 } else if (Op1I->getOpcode() == Instruction::Xor) {
4029 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
4030 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
4031 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
4032 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
4033 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
4034 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
4035 Op1I->swapOperands();
4036 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
4037 I.swapOperands(); // Simplified below.
4038 std::swap(Op0, Op1);
4042 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
4043 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
4044 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
4045 Op0I->swapOperands();
4046 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
4047 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
4048 InsertNewInstBefore(NotB, I);
4049 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
4051 } else if (Op0I->getOpcode() == Instruction::Xor) {
4052 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
4053 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
4054 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
4055 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
4056 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
4057 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
4058 Op0I->swapOperands();
4059 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
4060 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4061 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
4062 InsertNewInstBefore(N, I);
4063 return BinaryOperator::createAnd(N, Op1);
4067 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4068 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4069 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4072 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4073 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4074 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4075 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4076 const Type *SrcTy = Op0C->getOperand(0)->getType();
4077 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4078 // Only do this if the casts both really cause code to be generated.
4079 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4081 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4083 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4084 Op1C->getOperand(0),
4086 InsertNewInstBefore(NewOp, I);
4087 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4091 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4092 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4093 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4094 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4095 SI0->getOperand(1) == SI1->getOperand(1) &&
4096 (SI0->hasOneUse() || SI1->hasOneUse())) {
4097 Instruction *NewOp =
4098 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
4100 SI0->getName()), I);
4101 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4102 SI1->getOperand(1));
4106 return Changed ? &I : 0;
4109 static bool isPositive(ConstantInt *C) {
4110 return C->getSExtValue() >= 0;
4113 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4114 /// overflowed for this type.
4115 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4117 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4119 return cast<ConstantInt>(Result)->getZExtValue() <
4120 cast<ConstantInt>(In1)->getZExtValue();
4123 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4124 /// code necessary to compute the offset from the base pointer (without adding
4125 /// in the base pointer). Return the result as a signed integer of intptr size.
4126 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4127 TargetData &TD = IC.getTargetData();
4128 gep_type_iterator GTI = gep_type_begin(GEP);
4129 const Type *IntPtrTy = TD.getIntPtrType();
4130 Value *Result = Constant::getNullValue(IntPtrTy);
4132 // Build a mask for high order bits.
4133 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4135 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4136 Value *Op = GEP->getOperand(i);
4137 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4138 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4139 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4140 if (!OpC->isNullValue()) {
4141 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4142 Scale = ConstantExpr::getMul(OpC, Scale);
4143 if (Constant *RC = dyn_cast<Constant>(Result))
4144 Result = ConstantExpr::getAdd(RC, Scale);
4146 // Emit an add instruction.
4147 Result = IC.InsertNewInstBefore(
4148 BinaryOperator::createAdd(Result, Scale,
4149 GEP->getName()+".offs"), I);
4153 // Convert to correct type.
4154 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4155 Op->getName()+".c"), I);
4157 // We'll let instcombine(mul) convert this to a shl if possible.
4158 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4159 GEP->getName()+".idx"), I);
4161 // Emit an add instruction.
4162 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4163 GEP->getName()+".offs"), I);
4169 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4170 /// else. At this point we know that the GEP is on the LHS of the comparison.
4171 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4172 ICmpInst::Predicate Cond,
4174 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4176 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4177 if (isa<PointerType>(CI->getOperand(0)->getType()))
4178 RHS = CI->getOperand(0);
4180 Value *PtrBase = GEPLHS->getOperand(0);
4181 if (PtrBase == RHS) {
4182 // As an optimization, we don't actually have to compute the actual value of
4183 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4184 // each index is zero or not.
4185 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4186 Instruction *InVal = 0;
4187 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4188 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4190 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4191 if (isa<UndefValue>(C)) // undef index -> undef.
4192 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4193 if (C->isNullValue())
4195 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4196 EmitIt = false; // This is indexing into a zero sized array?
4197 } else if (isa<ConstantInt>(C))
4198 return ReplaceInstUsesWith(I, // No comparison is needed here.
4199 ConstantInt::get(Type::Int1Ty,
4200 Cond == ICmpInst::ICMP_NE));
4205 new ICmpInst(Cond, GEPLHS->getOperand(i),
4206 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4210 InVal = InsertNewInstBefore(InVal, I);
4211 InsertNewInstBefore(Comp, I);
4212 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4213 InVal = BinaryOperator::createOr(InVal, Comp);
4214 else // True if all are equal
4215 InVal = BinaryOperator::createAnd(InVal, Comp);
4223 // No comparison is needed here, all indexes = 0
4224 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4225 Cond == ICmpInst::ICMP_EQ));
4228 // Only lower this if the icmp is the only user of the GEP or if we expect
4229 // the result to fold to a constant!
4230 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4231 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4232 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4233 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4234 Constant::getNullValue(Offset->getType()));
4236 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4237 // If the base pointers are different, but the indices are the same, just
4238 // compare the base pointer.
4239 if (PtrBase != GEPRHS->getOperand(0)) {
4240 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4241 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4242 GEPRHS->getOperand(0)->getType();
4244 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4245 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4246 IndicesTheSame = false;
4250 // If all indices are the same, just compare the base pointers.
4252 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4253 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4255 // Otherwise, the base pointers are different and the indices are
4256 // different, bail out.
4260 // If one of the GEPs has all zero indices, recurse.
4261 bool AllZeros = true;
4262 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4263 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4264 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4269 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4270 ICmpInst::getSwappedPredicate(Cond), I);
4272 // If the other GEP has all zero indices, recurse.
4274 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4275 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4276 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4281 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4283 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4284 // If the GEPs only differ by one index, compare it.
4285 unsigned NumDifferences = 0; // Keep track of # differences.
4286 unsigned DiffOperand = 0; // The operand that differs.
4287 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4288 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4289 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4290 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4291 // Irreconcilable differences.
4295 if (NumDifferences++) break;
4300 if (NumDifferences == 0) // SAME GEP?
4301 return ReplaceInstUsesWith(I, // No comparison is needed here.
4302 ConstantInt::get(Type::Int1Ty,
4303 Cond == ICmpInst::ICMP_EQ));
4304 else if (NumDifferences == 1) {
4305 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4306 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4307 // Make sure we do a signed comparison here.
4308 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4312 // Only lower this if the icmp is the only user of the GEP or if we expect
4313 // the result to fold to a constant!
4314 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4315 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4316 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4317 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4318 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4319 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4325 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4326 bool Changed = SimplifyCompare(I);
4327 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4329 // Fold trivial predicates.
4330 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4331 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4332 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4333 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4335 // Simplify 'fcmp pred X, X'
4337 switch (I.getPredicate()) {
4338 default: assert(0 && "Unknown predicate!");
4339 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4340 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4341 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4342 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4343 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4344 case FCmpInst::FCMP_OLT: // True if ordered and less than
4345 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4346 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4348 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4349 case FCmpInst::FCMP_ULT: // True if unordered or less than
4350 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4351 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4352 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4353 I.setPredicate(FCmpInst::FCMP_UNO);
4354 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4357 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4358 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4359 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4360 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4361 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4362 I.setPredicate(FCmpInst::FCMP_ORD);
4363 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4368 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4369 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4371 // Handle fcmp with constant RHS
4372 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4373 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4374 switch (LHSI->getOpcode()) {
4375 case Instruction::PHI:
4376 if (Instruction *NV = FoldOpIntoPhi(I))
4379 case Instruction::Select:
4380 // If either operand of the select is a constant, we can fold the
4381 // comparison into the select arms, which will cause one to be
4382 // constant folded and the select turned into a bitwise or.
4383 Value *Op1 = 0, *Op2 = 0;
4384 if (LHSI->hasOneUse()) {
4385 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4386 // Fold the known value into the constant operand.
4387 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4388 // Insert a new FCmp of the other select operand.
4389 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4390 LHSI->getOperand(2), RHSC,
4392 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4393 // Fold the known value into the constant operand.
4394 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4395 // Insert a new FCmp of the other select operand.
4396 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4397 LHSI->getOperand(1), RHSC,
4403 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4408 return Changed ? &I : 0;
4411 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4412 bool Changed = SimplifyCompare(I);
4413 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4414 const Type *Ty = Op0->getType();
4418 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4419 isTrueWhenEqual(I)));
4421 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4422 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4424 // icmp of GlobalValues can never equal each other as long as they aren't
4425 // external weak linkage type.
4426 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4427 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4428 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4429 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4430 !isTrueWhenEqual(I)));
4432 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4433 // addresses never equal each other! We already know that Op0 != Op1.
4434 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4435 isa<ConstantPointerNull>(Op0)) &&
4436 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4437 isa<ConstantPointerNull>(Op1)))
4438 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4439 !isTrueWhenEqual(I)));
4441 // icmp's with boolean values can always be turned into bitwise operations
4442 if (Ty == Type::Int1Ty) {
4443 switch (I.getPredicate()) {
4444 default: assert(0 && "Invalid icmp instruction!");
4445 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4446 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4447 InsertNewInstBefore(Xor, I);
4448 return BinaryOperator::createNot(Xor);
4450 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4451 return BinaryOperator::createXor(Op0, Op1);
4453 case ICmpInst::ICMP_UGT:
4454 case ICmpInst::ICMP_SGT:
4455 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4457 case ICmpInst::ICMP_ULT:
4458 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4459 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4460 InsertNewInstBefore(Not, I);
4461 return BinaryOperator::createAnd(Not, Op1);
4463 case ICmpInst::ICMP_UGE:
4464 case ICmpInst::ICMP_SGE:
4465 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4467 case ICmpInst::ICMP_ULE:
4468 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4469 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4470 InsertNewInstBefore(Not, I);
4471 return BinaryOperator::createOr(Not, Op1);
4476 // See if we are doing a comparison between a constant and an instruction that
4477 // can be folded into the comparison.
4478 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4479 switch (I.getPredicate()) {
4481 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4482 if (CI->isMinValue(false))
4483 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4484 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4485 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4486 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4487 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4490 case ICmpInst::ICMP_SLT:
4491 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4492 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4493 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4494 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4495 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4496 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4499 case ICmpInst::ICMP_UGT:
4500 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4501 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4502 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4503 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4504 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4505 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4508 case ICmpInst::ICMP_SGT:
4509 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4510 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4511 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4512 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4513 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4514 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4517 case ICmpInst::ICMP_ULE:
4518 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4519 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4520 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4521 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4522 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4523 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4526 case ICmpInst::ICMP_SLE:
4527 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4528 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4529 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4530 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4531 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4532 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4535 case ICmpInst::ICMP_UGE:
4536 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4537 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4538 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4539 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4540 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4541 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4544 case ICmpInst::ICMP_SGE:
4545 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4546 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4547 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4548 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4549 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4550 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4554 // If we still have a icmp le or icmp ge instruction, turn it into the
4555 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4556 // already been handled above, this requires little checking.
4558 if (I.getPredicate() == ICmpInst::ICMP_ULE)
4559 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4560 if (I.getPredicate() == ICmpInst::ICMP_SLE)
4561 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4562 if (I.getPredicate() == ICmpInst::ICMP_UGE)
4563 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4564 if (I.getPredicate() == ICmpInst::ICMP_SGE)
4565 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4567 // See if we can fold the comparison based on bits known to be zero or one
4569 uint64_t KnownZero, KnownOne;
4570 if (SimplifyDemandedBits(Op0, cast<IntegerType>(Ty)->getBitMask(),
4571 KnownZero, KnownOne, 0))
4574 // Given the known and unknown bits, compute a range that the LHS could be
4576 if (KnownOne | KnownZero) {
4577 // Compute the Min, Max and RHS values based on the known bits. For the
4578 // EQ and NE we use unsigned values.
4579 uint64_t UMin = 0, UMax = 0, URHSVal = 0;
4580 int64_t SMin = 0, SMax = 0, SRHSVal = 0;
4581 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4582 SRHSVal = CI->getSExtValue();
4583 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
4586 URHSVal = CI->getZExtValue();
4587 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
4590 switch (I.getPredicate()) { // LE/GE have been folded already.
4591 default: assert(0 && "Unknown icmp opcode!");
4592 case ICmpInst::ICMP_EQ:
4593 if (UMax < URHSVal || UMin > URHSVal)
4594 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4596 case ICmpInst::ICMP_NE:
4597 if (UMax < URHSVal || UMin > URHSVal)
4598 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4600 case ICmpInst::ICMP_ULT:
4602 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4604 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4606 case ICmpInst::ICMP_UGT:
4608 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4610 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4612 case ICmpInst::ICMP_SLT:
4614 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4616 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4618 case ICmpInst::ICMP_SGT:
4620 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4622 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4627 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4628 // instruction, see if that instruction also has constants so that the
4629 // instruction can be folded into the icmp
4630 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4631 switch (LHSI->getOpcode()) {
4632 case Instruction::And:
4633 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4634 LHSI->getOperand(0)->hasOneUse()) {
4635 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4637 // If the LHS is an AND of a truncating cast, we can widen the
4638 // and/compare to be the input width without changing the value
4639 // produced, eliminating a cast.
4640 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4641 // We can do this transformation if either the AND constant does not
4642 // have its sign bit set or if it is an equality comparison.
4643 // Extending a relational comparison when we're checking the sign
4644 // bit would not work.
4645 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4647 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4648 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4649 ConstantInt *NewCST;
4651 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4652 AndCST->getZExtValue());
4653 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4654 CI->getZExtValue());
4655 Instruction *NewAnd =
4656 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4658 InsertNewInstBefore(NewAnd, I);
4659 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4663 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4664 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4665 // happens a LOT in code produced by the C front-end, for bitfield
4667 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
4668 if (Shift && !Shift->isShift())
4672 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4673 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4674 const Type *AndTy = AndCST->getType(); // Type of the and.
4676 // We can fold this as long as we can't shift unknown bits
4677 // into the mask. This can only happen with signed shift
4678 // rights, as they sign-extend.
4680 bool CanFold = Shift->isLogicalShift();
4682 // To test for the bad case of the signed shr, see if any
4683 // of the bits shifted in could be tested after the mask.
4684 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4685 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4687 Constant *OShAmt = ConstantInt::get(AndTy, ShAmtVal);
4689 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4691 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4697 if (Shift->getOpcode() == Instruction::Shl)
4698 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4700 NewCst = ConstantExpr::getShl(CI, ShAmt);
4702 // Check to see if we are shifting out any of the bits being
4704 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4705 // If we shifted bits out, the fold is not going to work out.
4706 // As a special case, check to see if this means that the
4707 // result is always true or false now.
4708 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4709 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4710 if (I.getPredicate() == ICmpInst::ICMP_NE)
4711 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4713 I.setOperand(1, NewCst);
4714 Constant *NewAndCST;
4715 if (Shift->getOpcode() == Instruction::Shl)
4716 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4718 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4719 LHSI->setOperand(1, NewAndCST);
4720 LHSI->setOperand(0, Shift->getOperand(0));
4721 AddToWorkList(Shift); // Shift is dead.
4722 AddUsesToWorkList(I);
4728 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4729 // preferable because it allows the C<<Y expression to be hoisted out
4730 // of a loop if Y is invariant and X is not.
4731 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4732 I.isEquality() && !Shift->isArithmeticShift() &&
4733 isa<Instruction>(Shift->getOperand(0))) {
4736 if (Shift->getOpcode() == Instruction::LShr) {
4737 NS = BinaryOperator::createShl(AndCST,
4738 Shift->getOperand(1), "tmp");
4740 // Insert a logical shift.
4741 NS = BinaryOperator::createLShr(AndCST,
4742 Shift->getOperand(1), "tmp");
4744 InsertNewInstBefore(cast<Instruction>(NS), I);
4746 // Compute X & (C << Y).
4747 Instruction *NewAnd = BinaryOperator::createAnd(
4748 Shift->getOperand(0), NS, LHSI->getName());
4749 InsertNewInstBefore(NewAnd, I);
4751 I.setOperand(0, NewAnd);
4757 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4758 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4759 if (I.isEquality()) {
4760 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4762 // Check that the shift amount is in range. If not, don't perform
4763 // undefined shifts. When the shift is visited it will be
4765 if (ShAmt->getZExtValue() >= TypeBits)
4768 // If we are comparing against bits always shifted out, the
4769 // comparison cannot succeed.
4771 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4772 if (Comp != CI) {// Comparing against a bit that we know is zero.
4773 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4774 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4775 return ReplaceInstUsesWith(I, Cst);
4778 if (LHSI->hasOneUse()) {
4779 // Otherwise strength reduce the shift into an and.
4780 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4781 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4782 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4785 BinaryOperator::createAnd(LHSI->getOperand(0),
4786 Mask, LHSI->getName()+".mask");
4787 Value *And = InsertNewInstBefore(AndI, I);
4788 return new ICmpInst(I.getPredicate(), And,
4789 ConstantExpr::getLShr(CI, ShAmt));
4795 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4796 case Instruction::AShr:
4797 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4798 if (I.isEquality()) {
4799 // Check that the shift amount is in range. If not, don't perform
4800 // undefined shifts. When the shift is visited it will be
4802 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4803 if (ShAmt->getZExtValue() >= TypeBits)
4806 // If we are comparing against bits always shifted out, the
4807 // comparison cannot succeed.
4809 if (LHSI->getOpcode() == Instruction::LShr)
4810 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4813 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4816 if (Comp != CI) {// Comparing against a bit that we know is zero.
4817 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4818 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4819 return ReplaceInstUsesWith(I, Cst);
4822 if (LHSI->hasOneUse() || CI->isNullValue()) {
4823 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4825 // Otherwise strength reduce the shift into an and.
4826 uint64_t Val = ~0ULL; // All ones.
4827 Val <<= ShAmtVal; // Shift over to the right spot.
4828 Val &= ~0ULL >> (64-TypeBits);
4829 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4832 BinaryOperator::createAnd(LHSI->getOperand(0),
4833 Mask, LHSI->getName()+".mask");
4834 Value *And = InsertNewInstBefore(AndI, I);
4835 return new ICmpInst(I.getPredicate(), And,
4836 ConstantExpr::getShl(CI, ShAmt));
4842 case Instruction::SDiv:
4843 case Instruction::UDiv:
4844 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4845 // Fold this div into the comparison, producing a range check.
4846 // Determine, based on the divide type, what the range is being
4847 // checked. If there is an overflow on the low or high side, remember
4848 // it, otherwise compute the range [low, hi) bounding the new value.
4849 // See: InsertRangeTest above for the kinds of replacements possible.
4850 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4851 // FIXME: If the operand types don't match the type of the divide
4852 // then don't attempt this transform. The code below doesn't have the
4853 // logic to deal with a signed divide and an unsigned compare (and
4854 // vice versa). This is because (x /s C1) <s C2 produces different
4855 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4856 // (x /u C1) <u C2. Simply casting the operands and result won't
4857 // work. :( The if statement below tests that condition and bails
4859 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4860 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4863 // Initialize the variables that will indicate the nature of the
4865 bool LoOverflow = false, HiOverflow = false;
4866 ConstantInt *LoBound = 0, *HiBound = 0;
4868 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4869 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4870 // C2 (CI). By solving for X we can turn this into a range check
4871 // instead of computing a divide.
4873 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4875 // Determine if the product overflows by seeing if the product is
4876 // not equal to the divide. Make sure we do the same kind of divide
4877 // as in the LHS instruction that we're folding.
4878 bool ProdOV = !DivRHS->isNullValue() &&
4879 (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4880 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4882 // Get the ICmp opcode
4883 ICmpInst::Predicate predicate = I.getPredicate();
4885 if (DivRHS->isNullValue()) {
4886 // Don't hack on divide by zeros!
4887 } else if (!DivIsSigned) { // udiv
4889 LoOverflow = ProdOV;
4890 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4891 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4892 if (CI->isNullValue()) { // (X / pos) op 0
4894 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4896 } else if (isPositive(CI)) { // (X / pos) op pos
4898 LoOverflow = ProdOV;
4899 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4900 } else { // (X / pos) op neg
4901 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4902 LoOverflow = AddWithOverflow(LoBound, Prod,
4903 cast<ConstantInt>(DivRHSH));
4905 HiOverflow = ProdOV;
4907 } else { // Divisor is < 0.
4908 if (CI->isNullValue()) { // (X / neg) op 0
4909 LoBound = AddOne(DivRHS);
4910 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4911 if (HiBound == DivRHS)
4912 LoBound = 0; // - INTMIN = INTMIN
4913 } else if (isPositive(CI)) { // (X / neg) op pos
4914 HiOverflow = LoOverflow = ProdOV;
4916 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4917 HiBound = AddOne(Prod);
4918 } else { // (X / neg) op neg
4920 LoOverflow = HiOverflow = ProdOV;
4921 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4924 // Dividing by a negate swaps the condition.
4925 predicate = ICmpInst::getSwappedPredicate(predicate);
4929 Value *X = LHSI->getOperand(0);
4930 switch (predicate) {
4931 default: assert(0 && "Unhandled icmp opcode!");
4932 case ICmpInst::ICMP_EQ:
4933 if (LoOverflow && HiOverflow)
4934 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4935 else if (HiOverflow)
4936 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4937 ICmpInst::ICMP_UGE, X, LoBound);
4938 else if (LoOverflow)
4939 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4940 ICmpInst::ICMP_ULT, X, HiBound);
4942 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4944 case ICmpInst::ICMP_NE:
4945 if (LoOverflow && HiOverflow)
4946 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4947 else if (HiOverflow)
4948 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4949 ICmpInst::ICMP_ULT, X, LoBound);
4950 else if (LoOverflow)
4951 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4952 ICmpInst::ICMP_UGE, X, HiBound);
4954 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4956 case ICmpInst::ICMP_ULT:
4957 case ICmpInst::ICMP_SLT:
4959 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4960 return new ICmpInst(predicate, X, LoBound);
4961 case ICmpInst::ICMP_UGT:
4962 case ICmpInst::ICMP_SGT:
4964 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4965 if (predicate == ICmpInst::ICMP_UGT)
4966 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
4968 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
4975 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
4976 if (I.isEquality()) {
4977 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4979 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4980 // the second operand is a constant, simplify a bit.
4981 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4982 switch (BO->getOpcode()) {
4983 case Instruction::SRem:
4984 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4985 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4987 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4988 if (V > 1 && isPowerOf2_64(V)) {
4989 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4990 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4991 return new ICmpInst(I.getPredicate(), NewRem,
4992 Constant::getNullValue(BO->getType()));
4996 case Instruction::Add:
4997 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4998 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4999 if (BO->hasOneUse())
5000 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5001 ConstantExpr::getSub(CI, BOp1C));
5002 } else if (CI->isNullValue()) {
5003 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5004 // efficiently invertible, or if the add has just this one use.
5005 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5007 if (Value *NegVal = dyn_castNegVal(BOp1))
5008 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
5009 else if (Value *NegVal = dyn_castNegVal(BOp0))
5010 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
5011 else if (BO->hasOneUse()) {
5012 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5013 InsertNewInstBefore(Neg, I);
5015 return new ICmpInst(I.getPredicate(), BOp0, Neg);
5019 case Instruction::Xor:
5020 // For the xor case, we can xor two constants together, eliminating
5021 // the explicit xor.
5022 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5023 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5024 ConstantExpr::getXor(CI, BOC));
5027 case Instruction::Sub:
5028 // Replace (([sub|xor] A, B) != 0) with (A != B)
5029 if (CI->isNullValue())
5030 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5034 case Instruction::Or:
5035 // If bits are being or'd in that are not present in the constant we
5036 // are comparing against, then the comparison could never succeed!
5037 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5038 Constant *NotCI = ConstantExpr::getNot(CI);
5039 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5040 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5045 case Instruction::And:
5046 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5047 // If bits are being compared against that are and'd out, then the
5048 // comparison can never succeed!
5049 if (!ConstantExpr::getAnd(CI,
5050 ConstantExpr::getNot(BOC))->isNullValue())
5051 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5054 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5055 if (CI == BOC && isOneBitSet(CI))
5056 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5057 ICmpInst::ICMP_NE, Op0,
5058 Constant::getNullValue(CI->getType()));
5060 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5061 if (isSignBit(BOC)) {
5062 Value *X = BO->getOperand(0);
5063 Constant *Zero = Constant::getNullValue(X->getType());
5064 ICmpInst::Predicate pred = isICMP_NE ?
5065 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5066 return new ICmpInst(pred, X, Zero);
5069 // ((X & ~7) == 0) --> X < 8
5070 if (CI->isNullValue() && isHighOnes(BOC)) {
5071 Value *X = BO->getOperand(0);
5072 Constant *NegX = ConstantExpr::getNeg(BOC);
5073 ICmpInst::Predicate pred = isICMP_NE ?
5074 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5075 return new ICmpInst(pred, X, NegX);
5081 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5082 // Handle set{eq|ne} <intrinsic>, intcst.
5083 switch (II->getIntrinsicID()) {
5085 case Intrinsic::bswap_i16:
5086 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5087 AddToWorkList(II); // Dead?
5088 I.setOperand(0, II->getOperand(1));
5089 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5090 ByteSwap_16(CI->getZExtValue())));
5092 case Intrinsic::bswap_i32:
5093 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5094 AddToWorkList(II); // Dead?
5095 I.setOperand(0, II->getOperand(1));
5096 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5097 ByteSwap_32(CI->getZExtValue())));
5099 case Intrinsic::bswap_i64:
5100 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5101 AddToWorkList(II); // Dead?
5102 I.setOperand(0, II->getOperand(1));
5103 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5104 ByteSwap_64(CI->getZExtValue())));
5108 } else { // Not a ICMP_EQ/ICMP_NE
5109 // If the LHS is a cast from an integral value of the same size, then
5110 // since we know the RHS is a constant, try to simlify.
5111 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5112 Value *CastOp = Cast->getOperand(0);
5113 const Type *SrcTy = CastOp->getType();
5114 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5115 if (SrcTy->isInteger() &&
5116 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5117 // If this is an unsigned comparison, try to make the comparison use
5118 // smaller constant values.
5119 switch (I.getPredicate()) {
5121 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5122 ConstantInt *CUI = cast<ConstantInt>(CI);
5123 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5124 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5125 ConstantInt::get(SrcTy, -1ULL));
5128 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5129 ConstantInt *CUI = cast<ConstantInt>(CI);
5130 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5131 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5132 Constant::getNullValue(SrcTy));
5142 // Handle icmp with constant RHS
5143 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5144 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5145 switch (LHSI->getOpcode()) {
5146 case Instruction::GetElementPtr:
5147 if (RHSC->isNullValue()) {
5148 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5149 bool isAllZeros = true;
5150 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5151 if (!isa<Constant>(LHSI->getOperand(i)) ||
5152 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5157 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5158 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5162 case Instruction::PHI:
5163 if (Instruction *NV = FoldOpIntoPhi(I))
5166 case Instruction::Select:
5167 // If either operand of the select is a constant, we can fold the
5168 // comparison into the select arms, which will cause one to be
5169 // constant folded and the select turned into a bitwise or.
5170 Value *Op1 = 0, *Op2 = 0;
5171 if (LHSI->hasOneUse()) {
5172 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5173 // Fold the known value into the constant operand.
5174 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5175 // Insert a new ICmp of the other select operand.
5176 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5177 LHSI->getOperand(2), RHSC,
5179 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5180 // Fold the known value into the constant operand.
5181 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5182 // Insert a new ICmp of the other select operand.
5183 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5184 LHSI->getOperand(1), RHSC,
5190 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5195 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5196 if (User *GEP = dyn_castGetElementPtr(Op0))
5197 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5199 if (User *GEP = dyn_castGetElementPtr(Op1))
5200 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5201 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5204 // Test to see if the operands of the icmp are casted versions of other
5205 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5207 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5208 if (isa<PointerType>(Op0->getType()) &&
5209 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5210 // We keep moving the cast from the left operand over to the right
5211 // operand, where it can often be eliminated completely.
5212 Op0 = CI->getOperand(0);
5214 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5215 // so eliminate it as well.
5216 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5217 Op1 = CI2->getOperand(0);
5219 // If Op1 is a constant, we can fold the cast into the constant.
5220 if (Op0->getType() != Op1->getType())
5221 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5222 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5224 // Otherwise, cast the RHS right before the icmp
5225 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5227 return new ICmpInst(I.getPredicate(), Op0, Op1);
5231 if (isa<CastInst>(Op0)) {
5232 // Handle the special case of: icmp (cast bool to X), <cst>
5233 // This comes up when you have code like
5236 // For generality, we handle any zero-extension of any operand comparison
5237 // with a constant or another cast from the same type.
5238 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5239 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5243 if (I.isEquality()) {
5244 Value *A, *B, *C, *D;
5245 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5246 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5247 Value *OtherVal = A == Op1 ? B : A;
5248 return new ICmpInst(I.getPredicate(), OtherVal,
5249 Constant::getNullValue(A->getType()));
5252 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5253 // A^c1 == C^c2 --> A == C^(c1^c2)
5254 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5255 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5256 if (Op1->hasOneUse()) {
5257 Constant *NC = ConstantExpr::getXor(C1, C2);
5258 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5259 return new ICmpInst(I.getPredicate(), A,
5260 InsertNewInstBefore(Xor, I));
5263 // A^B == A^D -> B == D
5264 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5265 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5266 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5267 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5271 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5272 (A == Op0 || B == Op0)) {
5273 // A == (A^B) -> B == 0
5274 Value *OtherVal = A == Op0 ? B : A;
5275 return new ICmpInst(I.getPredicate(), OtherVal,
5276 Constant::getNullValue(A->getType()));
5278 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5279 // (A-B) == A -> B == 0
5280 return new ICmpInst(I.getPredicate(), B,
5281 Constant::getNullValue(B->getType()));
5283 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5284 // A == (A-B) -> B == 0
5285 return new ICmpInst(I.getPredicate(), B,
5286 Constant::getNullValue(B->getType()));
5289 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5290 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5291 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5292 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5293 Value *X = 0, *Y = 0, *Z = 0;
5296 X = B; Y = D; Z = A;
5297 } else if (A == D) {
5298 X = B; Y = C; Z = A;
5299 } else if (B == C) {
5300 X = A; Y = D; Z = B;
5301 } else if (B == D) {
5302 X = A; Y = C; Z = B;
5305 if (X) { // Build (X^Y) & Z
5306 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5307 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5308 I.setOperand(0, Op1);
5309 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5314 return Changed ? &I : 0;
5317 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5318 // We only handle extending casts so far.
5320 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5321 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5322 Value *LHSCIOp = LHSCI->getOperand(0);
5323 const Type *SrcTy = LHSCIOp->getType();
5324 const Type *DestTy = LHSCI->getType();
5327 // We only handle extension cast instructions, so far. Enforce this.
5328 if (LHSCI->getOpcode() != Instruction::ZExt &&
5329 LHSCI->getOpcode() != Instruction::SExt)
5332 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5333 bool isSignedCmp = ICI.isSignedPredicate();
5335 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5336 // Not an extension from the same type?
5337 RHSCIOp = CI->getOperand(0);
5338 if (RHSCIOp->getType() != LHSCIOp->getType())
5341 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5342 // and the other is a zext), then we can't handle this.
5343 if (CI->getOpcode() != LHSCI->getOpcode())
5346 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5347 // then we can't handle this.
5348 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5351 // Okay, just insert a compare of the reduced operands now!
5352 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5355 // If we aren't dealing with a constant on the RHS, exit early
5356 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5360 // Compute the constant that would happen if we truncated to SrcTy then
5361 // reextended to DestTy.
5362 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5363 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5365 // If the re-extended constant didn't change...
5367 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5368 // For example, we might have:
5369 // %A = sext short %X to uint
5370 // %B = icmp ugt uint %A, 1330
5371 // It is incorrect to transform this into
5372 // %B = icmp ugt short %X, 1330
5373 // because %A may have negative value.
5375 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5376 // OR operation is EQ/NE.
5377 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5378 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5383 // The re-extended constant changed so the constant cannot be represented
5384 // in the shorter type. Consequently, we cannot emit a simple comparison.
5386 // First, handle some easy cases. We know the result cannot be equal at this
5387 // point so handle the ICI.isEquality() cases
5388 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5389 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5390 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5391 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5393 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5394 // should have been folded away previously and not enter in here.
5397 // We're performing a signed comparison.
5398 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5399 Result = ConstantInt::getFalse(); // X < (small) --> false
5401 Result = ConstantInt::getTrue(); // X < (large) --> true
5403 // We're performing an unsigned comparison.
5405 // We're performing an unsigned comp with a sign extended value.
5406 // This is true if the input is >= 0. [aka >s -1]
5407 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5408 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5409 NegOne, ICI.getName()), ICI);
5411 // Unsigned extend & unsigned compare -> always true.
5412 Result = ConstantInt::getTrue();
5416 // Finally, return the value computed.
5417 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5418 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5419 return ReplaceInstUsesWith(ICI, Result);
5421 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5422 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5423 "ICmp should be folded!");
5424 if (Constant *CI = dyn_cast<Constant>(Result))
5425 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5427 return BinaryOperator::createNot(Result);
5431 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5432 return commonShiftTransforms(I);
5435 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5436 return commonShiftTransforms(I);
5439 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5440 return commonShiftTransforms(I);
5443 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5444 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5445 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5447 // shl X, 0 == X and shr X, 0 == X
5448 // shl 0, X == 0 and shr 0, X == 0
5449 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5450 Op0 == Constant::getNullValue(Op0->getType()))
5451 return ReplaceInstUsesWith(I, Op0);
5453 if (isa<UndefValue>(Op0)) {
5454 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5455 return ReplaceInstUsesWith(I, Op0);
5456 else // undef << X -> 0, undef >>u X -> 0
5457 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5459 if (isa<UndefValue>(Op1)) {
5460 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5461 return ReplaceInstUsesWith(I, Op0);
5462 else // X << undef, X >>u undef -> 0
5463 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5466 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5467 if (I.getOpcode() == Instruction::AShr)
5468 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5469 if (CSI->isAllOnesValue())
5470 return ReplaceInstUsesWith(I, CSI);
5472 // Try to fold constant and into select arguments.
5473 if (isa<Constant>(Op0))
5474 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5475 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5478 // See if we can turn a signed shr into an unsigned shr.
5479 if (I.isArithmeticShift()) {
5480 if (MaskedValueIsZero(Op0,
5481 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5482 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5486 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5487 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5492 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5493 BinaryOperator &I) {
5494 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5496 // See if we can simplify any instructions used by the instruction whose sole
5497 // purpose is to compute bits we don't care about.
5498 uint64_t KnownZero, KnownOne;
5499 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
5500 KnownZero, KnownOne))
5503 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5504 // of a signed value.
5506 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5507 if (Op1->getZExtValue() >= TypeBits) {
5508 if (I.getOpcode() != Instruction::AShr)
5509 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5511 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5516 // ((X*C1) << C2) == (X * (C1 << C2))
5517 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5518 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5519 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5520 return BinaryOperator::createMul(BO->getOperand(0),
5521 ConstantExpr::getShl(BOOp, Op1));
5523 // Try to fold constant and into select arguments.
5524 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5525 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5527 if (isa<PHINode>(Op0))
5528 if (Instruction *NV = FoldOpIntoPhi(I))
5531 if (Op0->hasOneUse()) {
5532 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5533 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5536 switch (Op0BO->getOpcode()) {
5538 case Instruction::Add:
5539 case Instruction::And:
5540 case Instruction::Or:
5541 case Instruction::Xor: {
5542 // These operators commute.
5543 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5544 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5545 match(Op0BO->getOperand(1),
5546 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5547 Instruction *YS = BinaryOperator::createShl(
5548 Op0BO->getOperand(0), Op1,
5550 InsertNewInstBefore(YS, I); // (Y << C)
5552 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5553 Op0BO->getOperand(1)->getName());
5554 InsertNewInstBefore(X, I); // (X + (Y << C))
5555 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5556 C2 = ConstantExpr::getShl(C2, Op1);
5557 return BinaryOperator::createAnd(X, C2);
5560 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5561 Value *Op0BOOp1 = Op0BO->getOperand(1);
5562 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5564 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5565 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5567 Instruction *YS = BinaryOperator::createShl(
5568 Op0BO->getOperand(0), Op1,
5570 InsertNewInstBefore(YS, I); // (Y << C)
5572 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5573 V1->getName()+".mask");
5574 InsertNewInstBefore(XM, I); // X & (CC << C)
5576 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5581 case Instruction::Sub: {
5582 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5583 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5584 match(Op0BO->getOperand(0),
5585 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5586 Instruction *YS = BinaryOperator::createShl(
5587 Op0BO->getOperand(1), Op1,
5589 InsertNewInstBefore(YS, I); // (Y << C)
5591 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5592 Op0BO->getOperand(0)->getName());
5593 InsertNewInstBefore(X, I); // (X + (Y << C))
5594 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5595 C2 = ConstantExpr::getShl(C2, Op1);
5596 return BinaryOperator::createAnd(X, C2);
5599 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5600 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5601 match(Op0BO->getOperand(0),
5602 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5603 m_ConstantInt(CC))) && V2 == Op1 &&
5604 cast<BinaryOperator>(Op0BO->getOperand(0))
5605 ->getOperand(0)->hasOneUse()) {
5606 Instruction *YS = BinaryOperator::createShl(
5607 Op0BO->getOperand(1), Op1,
5609 InsertNewInstBefore(YS, I); // (Y << C)
5611 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5612 V1->getName()+".mask");
5613 InsertNewInstBefore(XM, I); // X & (CC << C)
5615 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5623 // If the operand is an bitwise operator with a constant RHS, and the
5624 // shift is the only use, we can pull it out of the shift.
5625 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5626 bool isValid = true; // Valid only for And, Or, Xor
5627 bool highBitSet = false; // Transform if high bit of constant set?
5629 switch (Op0BO->getOpcode()) {
5630 default: isValid = false; break; // Do not perform transform!
5631 case Instruction::Add:
5632 isValid = isLeftShift;
5634 case Instruction::Or:
5635 case Instruction::Xor:
5638 case Instruction::And:
5643 // If this is a signed shift right, and the high bit is modified
5644 // by the logical operation, do not perform the transformation.
5645 // The highBitSet boolean indicates the value of the high bit of
5646 // the constant which would cause it to be modified for this
5649 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
5650 uint64_t Val = Op0C->getZExtValue();
5651 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5655 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5657 Instruction *NewShift =
5658 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
5659 InsertNewInstBefore(NewShift, I);
5660 NewShift->takeName(Op0BO);
5662 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5669 // Find out if this is a shift of a shift by a constant.
5670 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
5671 if (ShiftOp && !ShiftOp->isShift())
5674 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5675 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5676 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5677 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5678 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
5679 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
5680 Value *X = ShiftOp->getOperand(0);
5682 unsigned AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5683 if (AmtSum > I.getType()->getPrimitiveSizeInBits())
5684 AmtSum = I.getType()->getPrimitiveSizeInBits();
5686 const IntegerType *Ty = cast<IntegerType>(I.getType());
5688 // Check for (X << c1) << c2 and (X >> c1) >> c2
5689 if (I.getOpcode() == ShiftOp->getOpcode()) {
5690 return BinaryOperator::create(I.getOpcode(), X,
5691 ConstantInt::get(Ty, AmtSum));
5692 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
5693 I.getOpcode() == Instruction::AShr) {
5694 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
5695 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
5696 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
5697 I.getOpcode() == Instruction::LShr) {
5698 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
5699 Instruction *Shift =
5700 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
5701 InsertNewInstBefore(Shift, I);
5703 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5704 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5707 // Okay, if we get here, one shift must be left, and the other shift must be
5708 // right. See if the amounts are equal.
5709 if (ShiftAmt1 == ShiftAmt2) {
5710 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
5711 if (I.getOpcode() == Instruction::Shl) {
5712 uint64_t Mask = Ty->getBitMask() << ShiftAmt1;
5713 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
5715 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
5716 if (I.getOpcode() == Instruction::LShr) {
5717 uint64_t Mask = Ty->getBitMask() >> ShiftAmt1;
5718 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
5720 // We can simplify ((X << C) >>s C) into a trunc + sext.
5721 // NOTE: we could do this for any C, but that would make 'unusual' integer
5722 // types. For now, just stick to ones well-supported by the code
5724 const Type *SExtType = 0;
5725 switch (Ty->getBitWidth() - ShiftAmt1) {
5726 case 8 : SExtType = Type::Int8Ty; break;
5727 case 16: SExtType = Type::Int16Ty; break;
5728 case 32: SExtType = Type::Int32Ty; break;
5732 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
5733 InsertNewInstBefore(NewTrunc, I);
5734 return new SExtInst(NewTrunc, Ty);
5736 // Otherwise, we can't handle it yet.
5737 } else if (ShiftAmt1 < ShiftAmt2) {
5738 unsigned ShiftDiff = ShiftAmt2-ShiftAmt1;
5740 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
5741 if (I.getOpcode() == Instruction::Shl) {
5742 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5743 ShiftOp->getOpcode() == Instruction::AShr);
5744 Instruction *Shift =
5745 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5746 InsertNewInstBefore(Shift, I);
5748 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
5749 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5752 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
5753 if (I.getOpcode() == Instruction::LShr) {
5754 assert(ShiftOp->getOpcode() == Instruction::Shl);
5755 Instruction *Shift =
5756 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
5757 InsertNewInstBefore(Shift, I);
5759 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5760 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5763 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
5765 assert(ShiftAmt2 < ShiftAmt1);
5766 unsigned ShiftDiff = ShiftAmt1-ShiftAmt2;
5768 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
5769 if (I.getOpcode() == Instruction::Shl) {
5770 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5771 ShiftOp->getOpcode() == Instruction::AShr);
5772 Instruction *Shift =
5773 BinaryOperator::create(ShiftOp->getOpcode(), X,
5774 ConstantInt::get(Ty, ShiftDiff));
5775 InsertNewInstBefore(Shift, I);
5777 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
5778 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5781 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
5782 if (I.getOpcode() == Instruction::LShr) {
5783 assert(ShiftOp->getOpcode() == Instruction::Shl);
5784 Instruction *Shift =
5785 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5786 InsertNewInstBefore(Shift, I);
5788 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5789 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5792 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
5799 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5800 /// expression. If so, decompose it, returning some value X, such that Val is
5803 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5805 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5806 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5807 Offset = CI->getZExtValue();
5809 return ConstantInt::get(Type::Int32Ty, 0);
5810 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5811 if (I->getNumOperands() == 2) {
5812 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5813 if (I->getOpcode() == Instruction::Shl) {
5814 // This is a value scaled by '1 << the shift amt'.
5815 Scale = 1U << CUI->getZExtValue();
5817 return I->getOperand(0);
5818 } else if (I->getOpcode() == Instruction::Mul) {
5819 // This value is scaled by 'CUI'.
5820 Scale = CUI->getZExtValue();
5822 return I->getOperand(0);
5823 } else if (I->getOpcode() == Instruction::Add) {
5824 // We have X+C. Check to see if we really have (X*C2)+C1,
5825 // where C1 is divisible by C2.
5828 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5829 Offset += CUI->getZExtValue();
5830 if (SubScale > 1 && (Offset % SubScale == 0)) {
5839 // Otherwise, we can't look past this.
5846 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5847 /// try to eliminate the cast by moving the type information into the alloc.
5848 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5849 AllocationInst &AI) {
5850 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5851 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5853 // Remove any uses of AI that are dead.
5854 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5856 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5857 Instruction *User = cast<Instruction>(*UI++);
5858 if (isInstructionTriviallyDead(User)) {
5859 while (UI != E && *UI == User)
5860 ++UI; // If this instruction uses AI more than once, don't break UI.
5863 DOUT << "IC: DCE: " << *User;
5864 EraseInstFromFunction(*User);
5868 // Get the type really allocated and the type casted to.
5869 const Type *AllocElTy = AI.getAllocatedType();
5870 const Type *CastElTy = PTy->getElementType();
5871 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5873 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
5874 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
5875 if (CastElTyAlign < AllocElTyAlign) return 0;
5877 // If the allocation has multiple uses, only promote it if we are strictly
5878 // increasing the alignment of the resultant allocation. If we keep it the
5879 // same, we open the door to infinite loops of various kinds.
5880 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5882 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5883 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5884 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5886 // See if we can satisfy the modulus by pulling a scale out of the array
5888 unsigned ArraySizeScale, ArrayOffset;
5889 Value *NumElements = // See if the array size is a decomposable linear expr.
5890 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5892 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5894 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5895 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5897 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5902 // If the allocation size is constant, form a constant mul expression
5903 Amt = ConstantInt::get(Type::Int32Ty, Scale);
5904 if (isa<ConstantInt>(NumElements))
5905 Amt = ConstantExpr::getMul(
5906 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5907 // otherwise multiply the amount and the number of elements
5908 else if (Scale != 1) {
5909 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5910 Amt = InsertNewInstBefore(Tmp, AI);
5914 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5915 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
5916 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5917 Amt = InsertNewInstBefore(Tmp, AI);
5920 AllocationInst *New;
5921 if (isa<MallocInst>(AI))
5922 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
5924 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
5925 InsertNewInstBefore(New, AI);
5928 // If the allocation has multiple uses, insert a cast and change all things
5929 // that used it to use the new cast. This will also hack on CI, but it will
5931 if (!AI.hasOneUse()) {
5932 AddUsesToWorkList(AI);
5933 // New is the allocation instruction, pointer typed. AI is the original
5934 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5935 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5936 InsertNewInstBefore(NewCast, AI);
5937 AI.replaceAllUsesWith(NewCast);
5939 return ReplaceInstUsesWith(CI, New);
5942 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5943 /// and return it as type Ty without inserting any new casts and without
5944 /// changing the computed value. This is used by code that tries to decide
5945 /// whether promoting or shrinking integer operations to wider or smaller types
5946 /// will allow us to eliminate a truncate or extend.
5948 /// This is a truncation operation if Ty is smaller than V->getType(), or an
5949 /// extension operation if Ty is larger.
5950 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
5951 int &NumCastsRemoved) {
5952 // We can always evaluate constants in another type.
5953 if (isa<ConstantInt>(V))
5956 Instruction *I = dyn_cast<Instruction>(V);
5957 if (!I) return false;
5959 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
5961 switch (I->getOpcode()) {
5962 case Instruction::Add:
5963 case Instruction::Sub:
5964 case Instruction::And:
5965 case Instruction::Or:
5966 case Instruction::Xor:
5967 if (!I->hasOneUse()) return false;
5968 // These operators can all arbitrarily be extended or truncated.
5969 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5970 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5972 case Instruction::Shl:
5973 if (!I->hasOneUse()) return false;
5974 // If we are truncating the result of this SHL, and if it's a shift of a
5975 // constant amount, we can always perform a SHL in a smaller type.
5976 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5977 if (Ty->getBitWidth() < OrigTy->getBitWidth() &&
5978 CI->getZExtValue() < Ty->getBitWidth())
5979 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
5982 case Instruction::LShr:
5983 if (!I->hasOneUse()) return false;
5984 // If this is a truncate of a logical shr, we can truncate it to a smaller
5985 // lshr iff we know that the bits we would otherwise be shifting in are
5987 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5988 if (Ty->getBitWidth() < OrigTy->getBitWidth() &&
5989 MaskedValueIsZero(I->getOperand(0),
5990 OrigTy->getBitMask() & ~Ty->getBitMask()) &&
5991 CI->getZExtValue() < Ty->getBitWidth()) {
5992 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5996 case Instruction::Trunc:
5997 case Instruction::ZExt:
5998 case Instruction::SExt:
5999 // If this is a cast from the destination type, we can trivially eliminate
6000 // it, and this will remove a cast overall.
6001 if (I->getOperand(0)->getType() == Ty) {
6002 // If the first operand is itself a cast, and is eliminable, do not count
6003 // this as an eliminable cast. We would prefer to eliminate those two
6005 if (isa<CastInst>(I->getOperand(0)))
6013 // TODO: Can handle more cases here.
6020 /// EvaluateInDifferentType - Given an expression that
6021 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6022 /// evaluate the expression.
6023 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6025 if (Constant *C = dyn_cast<Constant>(V))
6026 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6028 // Otherwise, it must be an instruction.
6029 Instruction *I = cast<Instruction>(V);
6030 Instruction *Res = 0;
6031 switch (I->getOpcode()) {
6032 case Instruction::Add:
6033 case Instruction::Sub:
6034 case Instruction::And:
6035 case Instruction::Or:
6036 case Instruction::Xor:
6037 case Instruction::AShr:
6038 case Instruction::LShr:
6039 case Instruction::Shl: {
6040 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6041 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6042 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6043 LHS, RHS, I->getName());
6046 case Instruction::Trunc:
6047 case Instruction::ZExt:
6048 case Instruction::SExt:
6049 case Instruction::BitCast:
6050 // If the source type of the cast is the type we're trying for then we can
6051 // just return the source. There's no need to insert it because its not new.
6052 if (I->getOperand(0)->getType() == Ty)
6053 return I->getOperand(0);
6055 // Some other kind of cast, which shouldn't happen, so just ..
6058 // TODO: Can handle more cases here.
6059 assert(0 && "Unreachable!");
6063 return InsertNewInstBefore(Res, *I);
6066 /// @brief Implement the transforms common to all CastInst visitors.
6067 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6068 Value *Src = CI.getOperand(0);
6070 // Casting undef to anything results in undef so might as just replace it and
6071 // get rid of the cast.
6072 if (isa<UndefValue>(Src)) // cast undef -> undef
6073 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6075 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
6076 // eliminate it now.
6077 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6078 if (Instruction::CastOps opc =
6079 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6080 // The first cast (CSrc) is eliminable so we need to fix up or replace
6081 // the second cast (CI). CSrc will then have a good chance of being dead.
6082 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6086 // If casting the result of a getelementptr instruction with no offset, turn
6087 // this into a cast of the original pointer!
6089 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6090 bool AllZeroOperands = true;
6091 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6092 if (!isa<Constant>(GEP->getOperand(i)) ||
6093 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6094 AllZeroOperands = false;
6097 if (AllZeroOperands) {
6098 // Changing the cast operand is usually not a good idea but it is safe
6099 // here because the pointer operand is being replaced with another
6100 // pointer operand so the opcode doesn't need to change.
6101 CI.setOperand(0, GEP->getOperand(0));
6106 // If we are casting a malloc or alloca to a pointer to a type of the same
6107 // size, rewrite the allocation instruction to allocate the "right" type.
6108 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6109 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6112 // If we are casting a select then fold the cast into the select
6113 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6114 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6117 // If we are casting a PHI then fold the cast into the PHI
6118 if (isa<PHINode>(Src))
6119 if (Instruction *NV = FoldOpIntoPhi(CI))
6125 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6126 /// integer types. This function implements the common transforms for all those
6128 /// @brief Implement the transforms common to CastInst with integer operands
6129 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6130 if (Instruction *Result = commonCastTransforms(CI))
6133 Value *Src = CI.getOperand(0);
6134 const Type *SrcTy = Src->getType();
6135 const Type *DestTy = CI.getType();
6136 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6137 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6139 // See if we can simplify any instructions used by the LHS whose sole
6140 // purpose is to compute bits we don't care about.
6141 uint64_t KnownZero = 0, KnownOne = 0;
6142 if (SimplifyDemandedBits(&CI, cast<IntegerType>(DestTy)->getBitMask(),
6143 KnownZero, KnownOne))
6146 // If the source isn't an instruction or has more than one use then we
6147 // can't do anything more.
6148 Instruction *SrcI = dyn_cast<Instruction>(Src);
6149 if (!SrcI || !Src->hasOneUse())
6152 // Attempt to propagate the cast into the instruction for int->int casts.
6153 int NumCastsRemoved = 0;
6154 if (!isa<BitCastInst>(CI) &&
6155 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6157 // If this cast is a truncate, evaluting in a different type always
6158 // eliminates the cast, so it is always a win. If this is a noop-cast
6159 // this just removes a noop cast which isn't pointful, but simplifies
6160 // the code. If this is a zero-extension, we need to do an AND to
6161 // maintain the clear top-part of the computation, so we require that
6162 // the input have eliminated at least one cast. If this is a sign
6163 // extension, we insert two new casts (to do the extension) so we
6164 // require that two casts have been eliminated.
6166 switch (CI.getOpcode()) {
6168 // All the others use floating point so we shouldn't actually
6169 // get here because of the check above.
6170 assert(0 && "Unknown cast type");
6171 case Instruction::Trunc:
6174 case Instruction::ZExt:
6175 DoXForm = NumCastsRemoved >= 1;
6177 case Instruction::SExt:
6178 DoXForm = NumCastsRemoved >= 2;
6180 case Instruction::BitCast:
6186 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6187 CI.getOpcode() == Instruction::SExt);
6188 assert(Res->getType() == DestTy);
6189 switch (CI.getOpcode()) {
6190 default: assert(0 && "Unknown cast type!");
6191 case Instruction::Trunc:
6192 case Instruction::BitCast:
6193 // Just replace this cast with the result.
6194 return ReplaceInstUsesWith(CI, Res);
6195 case Instruction::ZExt: {
6196 // We need to emit an AND to clear the high bits.
6197 assert(SrcBitSize < DestBitSize && "Not a zext?");
6199 ConstantInt::get(Type::Int64Ty, (1ULL << SrcBitSize)-1);
6200 if (DestBitSize < 64)
6201 C = ConstantExpr::getTrunc(C, DestTy);
6202 return BinaryOperator::createAnd(Res, C);
6204 case Instruction::SExt:
6205 // We need to emit a cast to truncate, then a cast to sext.
6206 return CastInst::create(Instruction::SExt,
6207 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6213 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6214 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6216 switch (SrcI->getOpcode()) {
6217 case Instruction::Add:
6218 case Instruction::Mul:
6219 case Instruction::And:
6220 case Instruction::Or:
6221 case Instruction::Xor:
6222 // If we are discarding information, or just changing the sign,
6224 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6225 // Don't insert two casts if they cannot be eliminated. We allow
6226 // two casts to be inserted if the sizes are the same. This could
6227 // only be converting signedness, which is a noop.
6228 if (DestBitSize == SrcBitSize ||
6229 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6230 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6231 Instruction::CastOps opcode = CI.getOpcode();
6232 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6233 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6234 return BinaryOperator::create(
6235 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6239 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6240 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6241 SrcI->getOpcode() == Instruction::Xor &&
6242 Op1 == ConstantInt::getTrue() &&
6243 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6244 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6245 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6248 case Instruction::SDiv:
6249 case Instruction::UDiv:
6250 case Instruction::SRem:
6251 case Instruction::URem:
6252 // If we are just changing the sign, rewrite.
6253 if (DestBitSize == SrcBitSize) {
6254 // Don't insert two casts if they cannot be eliminated. We allow
6255 // two casts to be inserted if the sizes are the same. This could
6256 // only be converting signedness, which is a noop.
6257 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6258 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6259 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6261 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6263 return BinaryOperator::create(
6264 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6269 case Instruction::Shl:
6270 // Allow changing the sign of the source operand. Do not allow
6271 // changing the size of the shift, UNLESS the shift amount is a
6272 // constant. We must not change variable sized shifts to a smaller
6273 // size, because it is undefined to shift more bits out than exist
6275 if (DestBitSize == SrcBitSize ||
6276 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6277 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6278 Instruction::BitCast : Instruction::Trunc);
6279 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6280 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6281 return BinaryOperator::createShl(Op0c, Op1c);
6284 case Instruction::AShr:
6285 // If this is a signed shr, and if all bits shifted in are about to be
6286 // truncated off, turn it into an unsigned shr to allow greater
6288 if (DestBitSize < SrcBitSize &&
6289 isa<ConstantInt>(Op1)) {
6290 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6291 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6292 // Insert the new logical shift right.
6293 return BinaryOperator::createLShr(Op0, Op1);
6298 case Instruction::ICmp:
6299 // If we are just checking for a icmp eq of a single bit and casting it
6300 // to an integer, then shift the bit to the appropriate place and then
6301 // cast to integer to avoid the comparison.
6302 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6303 uint64_t Op1CV = Op1C->getZExtValue();
6304 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6305 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6306 // cast (X == 1) to int --> X iff X has only the low bit set.
6307 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6308 // cast (X != 0) to int --> X iff X has only the low bit set.
6309 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6310 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6311 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6312 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
6313 // If Op1C some other power of two, convert:
6314 uint64_t KnownZero, KnownOne;
6315 uint64_t TypeMask = Op1C->getType()->getBitMask();
6316 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6318 // This only works for EQ and NE
6319 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6320 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6323 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
6324 bool isNE = pred == ICmpInst::ICMP_NE;
6325 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
6326 // (X&4) == 2 --> false
6327 // (X&4) != 2 --> true
6328 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6329 Res = ConstantExpr::getZExt(Res, CI.getType());
6330 return ReplaceInstUsesWith(CI, Res);
6333 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
6336 // Perform a logical shr by shiftamt.
6337 // Insert the shift to put the result in the low bit.
6338 In = InsertNewInstBefore(
6339 BinaryOperator::createLShr(In,
6340 ConstantInt::get(In->getType(), ShiftAmt),
6341 In->getName()+".lobit"), CI);
6344 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6345 Constant *One = ConstantInt::get(In->getType(), 1);
6346 In = BinaryOperator::createXor(In, One, "tmp");
6347 InsertNewInstBefore(cast<Instruction>(In), CI);
6350 if (CI.getType() == In->getType())
6351 return ReplaceInstUsesWith(CI, In);
6353 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6362 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6363 if (Instruction *Result = commonIntCastTransforms(CI))
6366 Value *Src = CI.getOperand(0);
6367 const Type *Ty = CI.getType();
6368 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6370 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6371 switch (SrcI->getOpcode()) {
6373 case Instruction::LShr:
6374 // We can shrink lshr to something smaller if we know the bits shifted in
6375 // are already zeros.
6376 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6377 unsigned ShAmt = ShAmtV->getZExtValue();
6379 // Get a mask for the bits shifting in.
6380 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6381 Value* SrcIOp0 = SrcI->getOperand(0);
6382 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6383 if (ShAmt >= DestBitWidth) // All zeros.
6384 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6386 // Okay, we can shrink this. Truncate the input, then return a new
6388 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6389 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6391 return BinaryOperator::createLShr(V1, V2);
6393 } else { // This is a variable shr.
6395 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6396 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6397 // loop-invariant and CSE'd.
6398 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6399 Value *One = ConstantInt::get(SrcI->getType(), 1);
6401 Value *V = InsertNewInstBefore(
6402 BinaryOperator::createShl(One, SrcI->getOperand(1),
6404 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6405 SrcI->getOperand(0),
6407 Value *Zero = Constant::getNullValue(V->getType());
6408 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6418 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6419 // If one of the common conversion will work ..
6420 if (Instruction *Result = commonIntCastTransforms(CI))
6423 Value *Src = CI.getOperand(0);
6425 // If this is a cast of a cast
6426 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6427 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6428 // types and if the sizes are just right we can convert this into a logical
6429 // 'and' which will be much cheaper than the pair of casts.
6430 if (isa<TruncInst>(CSrc)) {
6431 // Get the sizes of the types involved
6432 Value *A = CSrc->getOperand(0);
6433 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6434 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6435 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6436 // If we're actually extending zero bits and the trunc is a no-op
6437 if (MidSize < DstSize && SrcSize == DstSize) {
6438 // Replace both of the casts with an And of the type mask.
6439 uint64_t AndValue = cast<IntegerType>(CSrc->getType())->getBitMask();
6440 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6442 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6443 // Unfortunately, if the type changed, we need to cast it back.
6444 if (And->getType() != CI.getType()) {
6445 And->setName(CSrc->getName()+".mask");
6446 InsertNewInstBefore(And, CI);
6447 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6457 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6458 return commonIntCastTransforms(CI);
6461 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6462 return commonCastTransforms(CI);
6465 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6466 return commonCastTransforms(CI);
6469 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6470 return commonCastTransforms(CI);
6473 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6474 return commonCastTransforms(CI);
6477 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6478 return commonCastTransforms(CI);
6481 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6482 return commonCastTransforms(CI);
6485 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6486 return commonCastTransforms(CI);
6489 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6490 return commonCastTransforms(CI);
6493 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6495 // If the operands are integer typed then apply the integer transforms,
6496 // otherwise just apply the common ones.
6497 Value *Src = CI.getOperand(0);
6498 const Type *SrcTy = Src->getType();
6499 const Type *DestTy = CI.getType();
6501 if (SrcTy->isInteger() && DestTy->isInteger()) {
6502 if (Instruction *Result = commonIntCastTransforms(CI))
6505 if (Instruction *Result = commonCastTransforms(CI))
6510 // Get rid of casts from one type to the same type. These are useless and can
6511 // be replaced by the operand.
6512 if (DestTy == Src->getType())
6513 return ReplaceInstUsesWith(CI, Src);
6515 // If the source and destination are pointers, and this cast is equivalent to
6516 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6517 // This can enhance SROA and other transforms that want type-safe pointers.
6518 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6519 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6520 const Type *DstElTy = DstPTy->getElementType();
6521 const Type *SrcElTy = SrcPTy->getElementType();
6523 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6524 unsigned NumZeros = 0;
6525 while (SrcElTy != DstElTy &&
6526 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6527 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6528 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6532 // If we found a path from the src to dest, create the getelementptr now.
6533 if (SrcElTy == DstElTy) {
6534 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
6535 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
6540 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6541 if (SVI->hasOneUse()) {
6542 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6543 // a bitconvert to a vector with the same # elts.
6544 if (isa<VectorType>(DestTy) &&
6545 cast<VectorType>(DestTy)->getNumElements() ==
6546 SVI->getType()->getNumElements()) {
6548 // If either of the operands is a cast from CI.getType(), then
6549 // evaluating the shuffle in the casted destination's type will allow
6550 // us to eliminate at least one cast.
6551 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6552 Tmp->getOperand(0)->getType() == DestTy) ||
6553 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6554 Tmp->getOperand(0)->getType() == DestTy)) {
6555 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6556 SVI->getOperand(0), DestTy, &CI);
6557 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6558 SVI->getOperand(1), DestTy, &CI);
6559 // Return a new shuffle vector. Use the same element ID's, as we
6560 // know the vector types match #elts.
6561 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6569 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6571 /// %D = select %cond, %C, %A
6573 /// %C = select %cond, %B, 0
6576 /// Assuming that the specified instruction is an operand to the select, return
6577 /// a bitmask indicating which operands of this instruction are foldable if they
6578 /// equal the other incoming value of the select.
6580 static unsigned GetSelectFoldableOperands(Instruction *I) {
6581 switch (I->getOpcode()) {
6582 case Instruction::Add:
6583 case Instruction::Mul:
6584 case Instruction::And:
6585 case Instruction::Or:
6586 case Instruction::Xor:
6587 return 3; // Can fold through either operand.
6588 case Instruction::Sub: // Can only fold on the amount subtracted.
6589 case Instruction::Shl: // Can only fold on the shift amount.
6590 case Instruction::LShr:
6591 case Instruction::AShr:
6594 return 0; // Cannot fold
6598 /// GetSelectFoldableConstant - For the same transformation as the previous
6599 /// function, return the identity constant that goes into the select.
6600 static Constant *GetSelectFoldableConstant(Instruction *I) {
6601 switch (I->getOpcode()) {
6602 default: assert(0 && "This cannot happen!"); abort();
6603 case Instruction::Add:
6604 case Instruction::Sub:
6605 case Instruction::Or:
6606 case Instruction::Xor:
6607 case Instruction::Shl:
6608 case Instruction::LShr:
6609 case Instruction::AShr:
6610 return Constant::getNullValue(I->getType());
6611 case Instruction::And:
6612 return ConstantInt::getAllOnesValue(I->getType());
6613 case Instruction::Mul:
6614 return ConstantInt::get(I->getType(), 1);
6618 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6619 /// have the same opcode and only one use each. Try to simplify this.
6620 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6622 if (TI->getNumOperands() == 1) {
6623 // If this is a non-volatile load or a cast from the same type,
6626 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6629 return 0; // unknown unary op.
6632 // Fold this by inserting a select from the input values.
6633 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6634 FI->getOperand(0), SI.getName()+".v");
6635 InsertNewInstBefore(NewSI, SI);
6636 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6640 // Only handle binary operators here.
6641 if (!isa<BinaryOperator>(TI))
6644 // Figure out if the operations have any operands in common.
6645 Value *MatchOp, *OtherOpT, *OtherOpF;
6647 if (TI->getOperand(0) == FI->getOperand(0)) {
6648 MatchOp = TI->getOperand(0);
6649 OtherOpT = TI->getOperand(1);
6650 OtherOpF = FI->getOperand(1);
6651 MatchIsOpZero = true;
6652 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6653 MatchOp = TI->getOperand(1);
6654 OtherOpT = TI->getOperand(0);
6655 OtherOpF = FI->getOperand(0);
6656 MatchIsOpZero = false;
6657 } else if (!TI->isCommutative()) {
6659 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6660 MatchOp = TI->getOperand(0);
6661 OtherOpT = TI->getOperand(1);
6662 OtherOpF = FI->getOperand(0);
6663 MatchIsOpZero = true;
6664 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6665 MatchOp = TI->getOperand(1);
6666 OtherOpT = TI->getOperand(0);
6667 OtherOpF = FI->getOperand(1);
6668 MatchIsOpZero = true;
6673 // If we reach here, they do have operations in common.
6674 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6675 OtherOpF, SI.getName()+".v");
6676 InsertNewInstBefore(NewSI, SI);
6678 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6680 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6682 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6684 assert(0 && "Shouldn't get here");
6688 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6689 Value *CondVal = SI.getCondition();
6690 Value *TrueVal = SI.getTrueValue();
6691 Value *FalseVal = SI.getFalseValue();
6693 // select true, X, Y -> X
6694 // select false, X, Y -> Y
6695 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6696 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6698 // select C, X, X -> X
6699 if (TrueVal == FalseVal)
6700 return ReplaceInstUsesWith(SI, TrueVal);
6702 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6703 return ReplaceInstUsesWith(SI, FalseVal);
6704 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6705 return ReplaceInstUsesWith(SI, TrueVal);
6706 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6707 if (isa<Constant>(TrueVal))
6708 return ReplaceInstUsesWith(SI, TrueVal);
6710 return ReplaceInstUsesWith(SI, FalseVal);
6713 if (SI.getType() == Type::Int1Ty) {
6714 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6715 if (C->getZExtValue()) {
6716 // Change: A = select B, true, C --> A = or B, C
6717 return BinaryOperator::createOr(CondVal, FalseVal);
6719 // Change: A = select B, false, C --> A = and !B, C
6721 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6722 "not."+CondVal->getName()), SI);
6723 return BinaryOperator::createAnd(NotCond, FalseVal);
6725 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6726 if (C->getZExtValue() == false) {
6727 // Change: A = select B, C, false --> A = and B, C
6728 return BinaryOperator::createAnd(CondVal, TrueVal);
6730 // Change: A = select B, C, true --> A = or !B, C
6732 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6733 "not."+CondVal->getName()), SI);
6734 return BinaryOperator::createOr(NotCond, TrueVal);
6739 // Selecting between two integer constants?
6740 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6741 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6742 // select C, 1, 0 -> cast C to int
6743 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6744 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6745 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6746 // select C, 0, 1 -> cast !C to int
6748 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6749 "not."+CondVal->getName()), SI);
6750 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6753 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6755 // (x <s 0) ? -1 : 0 -> ashr x, 31
6756 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6757 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6758 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6759 bool CanXForm = false;
6760 if (IC->isSignedPredicate())
6761 CanXForm = CmpCst->isNullValue() &&
6762 IC->getPredicate() == ICmpInst::ICMP_SLT;
6764 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6765 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6766 IC->getPredicate() == ICmpInst::ICMP_UGT;
6770 // The comparison constant and the result are not neccessarily the
6771 // same width. Make an all-ones value by inserting a AShr.
6772 Value *X = IC->getOperand(0);
6773 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6774 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
6775 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
6777 InsertNewInstBefore(SRA, SI);
6779 // Finally, convert to the type of the select RHS. We figure out
6780 // if this requires a SExt, Trunc or BitCast based on the sizes.
6781 Instruction::CastOps opc = Instruction::BitCast;
6782 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6783 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6784 if (SRASize < SISize)
6785 opc = Instruction::SExt;
6786 else if (SRASize > SISize)
6787 opc = Instruction::Trunc;
6788 return CastInst::create(opc, SRA, SI.getType());
6793 // If one of the constants is zero (we know they can't both be) and we
6794 // have a fcmp instruction with zero, and we have an 'and' with the
6795 // non-constant value, eliminate this whole mess. This corresponds to
6796 // cases like this: ((X & 27) ? 27 : 0)
6797 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6798 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6799 cast<Constant>(IC->getOperand(1))->isNullValue())
6800 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6801 if (ICA->getOpcode() == Instruction::And &&
6802 isa<ConstantInt>(ICA->getOperand(1)) &&
6803 (ICA->getOperand(1) == TrueValC ||
6804 ICA->getOperand(1) == FalseValC) &&
6805 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6806 // Okay, now we know that everything is set up, we just don't
6807 // know whether we have a icmp_ne or icmp_eq and whether the
6808 // true or false val is the zero.
6809 bool ShouldNotVal = !TrueValC->isNullValue();
6810 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6813 V = InsertNewInstBefore(BinaryOperator::create(
6814 Instruction::Xor, V, ICA->getOperand(1)), SI);
6815 return ReplaceInstUsesWith(SI, V);
6820 // See if we are selecting two values based on a comparison of the two values.
6821 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6822 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6823 // Transform (X == Y) ? X : Y -> Y
6824 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6825 return ReplaceInstUsesWith(SI, FalseVal);
6826 // Transform (X != Y) ? X : Y -> X
6827 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6828 return ReplaceInstUsesWith(SI, TrueVal);
6829 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6831 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6832 // Transform (X == Y) ? Y : X -> X
6833 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6834 return ReplaceInstUsesWith(SI, FalseVal);
6835 // Transform (X != Y) ? Y : X -> Y
6836 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6837 return ReplaceInstUsesWith(SI, TrueVal);
6838 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6842 // See if we are selecting two values based on a comparison of the two values.
6843 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6844 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6845 // Transform (X == Y) ? X : Y -> Y
6846 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6847 return ReplaceInstUsesWith(SI, FalseVal);
6848 // Transform (X != Y) ? X : Y -> X
6849 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6850 return ReplaceInstUsesWith(SI, TrueVal);
6851 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6853 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6854 // Transform (X == Y) ? Y : X -> X
6855 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6856 return ReplaceInstUsesWith(SI, FalseVal);
6857 // Transform (X != Y) ? Y : X -> Y
6858 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6859 return ReplaceInstUsesWith(SI, TrueVal);
6860 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6864 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6865 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6866 if (TI->hasOneUse() && FI->hasOneUse()) {
6867 Instruction *AddOp = 0, *SubOp = 0;
6869 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6870 if (TI->getOpcode() == FI->getOpcode())
6871 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6874 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6875 // even legal for FP.
6876 if (TI->getOpcode() == Instruction::Sub &&
6877 FI->getOpcode() == Instruction::Add) {
6878 AddOp = FI; SubOp = TI;
6879 } else if (FI->getOpcode() == Instruction::Sub &&
6880 TI->getOpcode() == Instruction::Add) {
6881 AddOp = TI; SubOp = FI;
6885 Value *OtherAddOp = 0;
6886 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6887 OtherAddOp = AddOp->getOperand(1);
6888 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6889 OtherAddOp = AddOp->getOperand(0);
6893 // So at this point we know we have (Y -> OtherAddOp):
6894 // select C, (add X, Y), (sub X, Z)
6895 Value *NegVal; // Compute -Z
6896 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6897 NegVal = ConstantExpr::getNeg(C);
6899 NegVal = InsertNewInstBefore(
6900 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6903 Value *NewTrueOp = OtherAddOp;
6904 Value *NewFalseOp = NegVal;
6906 std::swap(NewTrueOp, NewFalseOp);
6907 Instruction *NewSel =
6908 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6910 NewSel = InsertNewInstBefore(NewSel, SI);
6911 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6916 // See if we can fold the select into one of our operands.
6917 if (SI.getType()->isInteger()) {
6918 // See the comment above GetSelectFoldableOperands for a description of the
6919 // transformation we are doing here.
6920 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6921 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6922 !isa<Constant>(FalseVal))
6923 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6924 unsigned OpToFold = 0;
6925 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6927 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6932 Constant *C = GetSelectFoldableConstant(TVI);
6933 Instruction *NewSel =
6934 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
6935 InsertNewInstBefore(NewSel, SI);
6936 NewSel->takeName(TVI);
6937 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6938 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6940 assert(0 && "Unknown instruction!!");
6945 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6946 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6947 !isa<Constant>(TrueVal))
6948 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6949 unsigned OpToFold = 0;
6950 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6952 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6957 Constant *C = GetSelectFoldableConstant(FVI);
6958 Instruction *NewSel =
6959 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
6960 InsertNewInstBefore(NewSel, SI);
6961 NewSel->takeName(FVI);
6962 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6963 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6965 assert(0 && "Unknown instruction!!");
6970 if (BinaryOperator::isNot(CondVal)) {
6971 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6972 SI.setOperand(1, FalseVal);
6973 SI.setOperand(2, TrueVal);
6980 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6981 /// determine, return it, otherwise return 0.
6982 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6983 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6984 unsigned Align = GV->getAlignment();
6985 if (Align == 0 && TD)
6986 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
6988 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6989 unsigned Align = AI->getAlignment();
6990 if (Align == 0 && TD) {
6991 if (isa<AllocaInst>(AI))
6992 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
6993 else if (isa<MallocInst>(AI)) {
6994 // Malloc returns maximally aligned memory.
6995 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
6998 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7001 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7005 } else if (isa<BitCastInst>(V) ||
7006 (isa<ConstantExpr>(V) &&
7007 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7008 User *CI = cast<User>(V);
7009 if (isa<PointerType>(CI->getOperand(0)->getType()))
7010 return GetKnownAlignment(CI->getOperand(0), TD);
7012 } else if (isa<GetElementPtrInst>(V) ||
7013 (isa<ConstantExpr>(V) &&
7014 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
7015 User *GEPI = cast<User>(V);
7016 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7017 if (BaseAlignment == 0) return 0;
7019 // If all indexes are zero, it is just the alignment of the base pointer.
7020 bool AllZeroOperands = true;
7021 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7022 if (!isa<Constant>(GEPI->getOperand(i)) ||
7023 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7024 AllZeroOperands = false;
7027 if (AllZeroOperands)
7028 return BaseAlignment;
7030 // Otherwise, if the base alignment is >= the alignment we expect for the
7031 // base pointer type, then we know that the resultant pointer is aligned at
7032 // least as much as its type requires.
7035 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7036 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7037 if (TD->getABITypeAlignment(PtrTy->getElementType())
7039 const Type *GEPTy = GEPI->getType();
7040 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7041 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7049 /// visitCallInst - CallInst simplification. This mostly only handles folding
7050 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7051 /// the heavy lifting.
7053 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7054 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7055 if (!II) return visitCallSite(&CI);
7057 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7059 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7060 bool Changed = false;
7062 // memmove/cpy/set of zero bytes is a noop.
7063 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7064 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7066 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7067 if (CI->getZExtValue() == 1) {
7068 // Replace the instruction with just byte operations. We would
7069 // transform other cases to loads/stores, but we don't know if
7070 // alignment is sufficient.
7074 // If we have a memmove and the source operation is a constant global,
7075 // then the source and dest pointers can't alias, so we can change this
7076 // into a call to memcpy.
7077 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7078 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7079 if (GVSrc->isConstant()) {
7080 Module *M = CI.getParent()->getParent()->getParent();
7082 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7084 Name = "llvm.memcpy.i32";
7086 Name = "llvm.memcpy.i64";
7087 Constant *MemCpy = M->getOrInsertFunction(Name,
7088 CI.getCalledFunction()->getFunctionType());
7089 CI.setOperand(0, MemCpy);
7094 // If we can determine a pointer alignment that is bigger than currently
7095 // set, update the alignment.
7096 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7097 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7098 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7099 unsigned Align = std::min(Alignment1, Alignment2);
7100 if (MI->getAlignment()->getZExtValue() < Align) {
7101 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7104 } else if (isa<MemSetInst>(MI)) {
7105 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7106 if (MI->getAlignment()->getZExtValue() < Alignment) {
7107 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7112 if (Changed) return II;
7114 switch (II->getIntrinsicID()) {
7116 case Intrinsic::ppc_altivec_lvx:
7117 case Intrinsic::ppc_altivec_lvxl:
7118 case Intrinsic::x86_sse_loadu_ps:
7119 case Intrinsic::x86_sse2_loadu_pd:
7120 case Intrinsic::x86_sse2_loadu_dq:
7121 // Turn PPC lvx -> load if the pointer is known aligned.
7122 // Turn X86 loadups -> load if the pointer is known aligned.
7123 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7124 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7125 PointerType::get(II->getType()), CI);
7126 return new LoadInst(Ptr);
7129 case Intrinsic::ppc_altivec_stvx:
7130 case Intrinsic::ppc_altivec_stvxl:
7131 // Turn stvx -> store if the pointer is known aligned.
7132 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7133 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7134 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7136 return new StoreInst(II->getOperand(1), Ptr);
7139 case Intrinsic::x86_sse_storeu_ps:
7140 case Intrinsic::x86_sse2_storeu_pd:
7141 case Intrinsic::x86_sse2_storeu_dq:
7142 case Intrinsic::x86_sse2_storel_dq:
7143 // Turn X86 storeu -> store if the pointer is known aligned.
7144 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7145 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7146 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7148 return new StoreInst(II->getOperand(2), Ptr);
7152 case Intrinsic::x86_sse_cvttss2si: {
7153 // These intrinsics only demands the 0th element of its input vector. If
7154 // we can simplify the input based on that, do so now.
7156 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7158 II->setOperand(1, V);
7164 case Intrinsic::ppc_altivec_vperm:
7165 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7166 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7167 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7169 // Check that all of the elements are integer constants or undefs.
7170 bool AllEltsOk = true;
7171 for (unsigned i = 0; i != 16; ++i) {
7172 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7173 !isa<UndefValue>(Mask->getOperand(i))) {
7180 // Cast the input vectors to byte vectors.
7181 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7182 II->getOperand(1), Mask->getType(), CI);
7183 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7184 II->getOperand(2), Mask->getType(), CI);
7185 Value *Result = UndefValue::get(Op0->getType());
7187 // Only extract each element once.
7188 Value *ExtractedElts[32];
7189 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7191 for (unsigned i = 0; i != 16; ++i) {
7192 if (isa<UndefValue>(Mask->getOperand(i)))
7194 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7195 Idx &= 31; // Match the hardware behavior.
7197 if (ExtractedElts[Idx] == 0) {
7199 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7200 InsertNewInstBefore(Elt, CI);
7201 ExtractedElts[Idx] = Elt;
7204 // Insert this value into the result vector.
7205 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7206 InsertNewInstBefore(cast<Instruction>(Result), CI);
7208 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7213 case Intrinsic::stackrestore: {
7214 // If the save is right next to the restore, remove the restore. This can
7215 // happen when variable allocas are DCE'd.
7216 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7217 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7218 BasicBlock::iterator BI = SS;
7220 return EraseInstFromFunction(CI);
7224 // If the stack restore is in a return/unwind block and if there are no
7225 // allocas or calls between the restore and the return, nuke the restore.
7226 TerminatorInst *TI = II->getParent()->getTerminator();
7227 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7228 BasicBlock::iterator BI = II;
7229 bool CannotRemove = false;
7230 for (++BI; &*BI != TI; ++BI) {
7231 if (isa<AllocaInst>(BI) ||
7232 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7233 CannotRemove = true;
7238 return EraseInstFromFunction(CI);
7245 return visitCallSite(II);
7248 // InvokeInst simplification
7250 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7251 return visitCallSite(&II);
7254 // visitCallSite - Improvements for call and invoke instructions.
7256 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7257 bool Changed = false;
7259 // If the callee is a constexpr cast of a function, attempt to move the cast
7260 // to the arguments of the call/invoke.
7261 if (transformConstExprCastCall(CS)) return 0;
7263 Value *Callee = CS.getCalledValue();
7265 if (Function *CalleeF = dyn_cast<Function>(Callee))
7266 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7267 Instruction *OldCall = CS.getInstruction();
7268 // If the call and callee calling conventions don't match, this call must
7269 // be unreachable, as the call is undefined.
7270 new StoreInst(ConstantInt::getTrue(),
7271 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7272 if (!OldCall->use_empty())
7273 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7274 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7275 return EraseInstFromFunction(*OldCall);
7279 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7280 // This instruction is not reachable, just remove it. We insert a store to
7281 // undef so that we know that this code is not reachable, despite the fact
7282 // that we can't modify the CFG here.
7283 new StoreInst(ConstantInt::getTrue(),
7284 UndefValue::get(PointerType::get(Type::Int1Ty)),
7285 CS.getInstruction());
7287 if (!CS.getInstruction()->use_empty())
7288 CS.getInstruction()->
7289 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7291 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7292 // Don't break the CFG, insert a dummy cond branch.
7293 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7294 ConstantInt::getTrue(), II);
7296 return EraseInstFromFunction(*CS.getInstruction());
7299 const PointerType *PTy = cast<PointerType>(Callee->getType());
7300 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7301 if (FTy->isVarArg()) {
7302 // See if we can optimize any arguments passed through the varargs area of
7304 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7305 E = CS.arg_end(); I != E; ++I)
7306 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7307 // If this cast does not effect the value passed through the varargs
7308 // area, we can eliminate the use of the cast.
7309 Value *Op = CI->getOperand(0);
7310 if (CI->isLosslessCast()) {
7317 return Changed ? CS.getInstruction() : 0;
7320 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7321 // attempt to move the cast to the arguments of the call/invoke.
7323 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7324 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7325 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7326 if (CE->getOpcode() != Instruction::BitCast ||
7327 !isa<Function>(CE->getOperand(0)))
7329 Function *Callee = cast<Function>(CE->getOperand(0));
7330 Instruction *Caller = CS.getInstruction();
7332 // Okay, this is a cast from a function to a different type. Unless doing so
7333 // would cause a type conversion of one of our arguments, change this call to
7334 // be a direct call with arguments casted to the appropriate types.
7336 const FunctionType *FT = Callee->getFunctionType();
7337 const Type *OldRetTy = Caller->getType();
7339 // Check to see if we are changing the return type...
7340 if (OldRetTy != FT->getReturnType()) {
7341 if (Callee->isDeclaration() && !Caller->use_empty() &&
7342 OldRetTy != FT->getReturnType() &&
7343 // Conversion is ok if changing from pointer to int of same size.
7344 !(isa<PointerType>(FT->getReturnType()) &&
7345 TD->getIntPtrType() == OldRetTy))
7346 return false; // Cannot transform this return value.
7348 // If the callsite is an invoke instruction, and the return value is used by
7349 // a PHI node in a successor, we cannot change the return type of the call
7350 // because there is no place to put the cast instruction (without breaking
7351 // the critical edge). Bail out in this case.
7352 if (!Caller->use_empty())
7353 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7354 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7356 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7357 if (PN->getParent() == II->getNormalDest() ||
7358 PN->getParent() == II->getUnwindDest())
7362 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7363 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7365 CallSite::arg_iterator AI = CS.arg_begin();
7366 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7367 const Type *ParamTy = FT->getParamType(i);
7368 const Type *ActTy = (*AI)->getType();
7369 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7370 //Either we can cast directly, or we can upconvert the argument
7371 bool isConvertible = ActTy == ParamTy ||
7372 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7373 (ParamTy->isInteger() && ActTy->isInteger() &&
7374 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7375 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7376 && c->getSExtValue() > 0);
7377 if (Callee->isDeclaration() && !isConvertible) return false;
7380 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7381 Callee->isDeclaration())
7382 return false; // Do not delete arguments unless we have a function body...
7384 // Okay, we decided that this is a safe thing to do: go ahead and start
7385 // inserting cast instructions as necessary...
7386 std::vector<Value*> Args;
7387 Args.reserve(NumActualArgs);
7389 AI = CS.arg_begin();
7390 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7391 const Type *ParamTy = FT->getParamType(i);
7392 if ((*AI)->getType() == ParamTy) {
7393 Args.push_back(*AI);
7395 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7396 false, ParamTy, false);
7397 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7398 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7402 // If the function takes more arguments than the call was taking, add them
7404 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7405 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7407 // If we are removing arguments to the function, emit an obnoxious warning...
7408 if (FT->getNumParams() < NumActualArgs)
7409 if (!FT->isVarArg()) {
7410 cerr << "WARNING: While resolving call to function '"
7411 << Callee->getName() << "' arguments were dropped!\n";
7413 // Add all of the arguments in their promoted form to the arg list...
7414 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7415 const Type *PTy = getPromotedType((*AI)->getType());
7416 if (PTy != (*AI)->getType()) {
7417 // Must promote to pass through va_arg area!
7418 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7420 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7421 InsertNewInstBefore(Cast, *Caller);
7422 Args.push_back(Cast);
7424 Args.push_back(*AI);
7429 if (FT->getReturnType() == Type::VoidTy)
7430 Caller->setName(""); // Void type should not have a name.
7433 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7434 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7435 &Args[0], Args.size(), Caller->getName(), Caller);
7436 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7438 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7439 if (cast<CallInst>(Caller)->isTailCall())
7440 cast<CallInst>(NC)->setTailCall();
7441 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7444 // Insert a cast of the return type as necessary.
7446 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7447 if (NV->getType() != Type::VoidTy) {
7448 const Type *CallerTy = Caller->getType();
7449 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7451 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7453 // If this is an invoke instruction, we should insert it after the first
7454 // non-phi, instruction in the normal successor block.
7455 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7456 BasicBlock::iterator I = II->getNormalDest()->begin();
7457 while (isa<PHINode>(I)) ++I;
7458 InsertNewInstBefore(NC, *I);
7460 // Otherwise, it's a call, just insert cast right after the call instr
7461 InsertNewInstBefore(NC, *Caller);
7463 AddUsersToWorkList(*Caller);
7465 NV = UndefValue::get(Caller->getType());
7469 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7470 Caller->replaceAllUsesWith(NV);
7471 Caller->eraseFromParent();
7472 RemoveFromWorkList(Caller);
7476 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7477 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7478 /// and a single binop.
7479 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7480 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7481 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
7482 isa<CmpInst>(FirstInst));
7483 unsigned Opc = FirstInst->getOpcode();
7484 Value *LHSVal = FirstInst->getOperand(0);
7485 Value *RHSVal = FirstInst->getOperand(1);
7487 const Type *LHSType = LHSVal->getType();
7488 const Type *RHSType = RHSVal->getType();
7490 // Scan to see if all operands are the same opcode, all have one use, and all
7491 // kill their operands (i.e. the operands have one use).
7492 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7493 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7494 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7495 // Verify type of the LHS matches so we don't fold cmp's of different
7496 // types or GEP's with different index types.
7497 I->getOperand(0)->getType() != LHSType ||
7498 I->getOperand(1)->getType() != RHSType)
7501 // If they are CmpInst instructions, check their predicates
7502 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7503 if (cast<CmpInst>(I)->getPredicate() !=
7504 cast<CmpInst>(FirstInst)->getPredicate())
7507 // Keep track of which operand needs a phi node.
7508 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7509 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7512 // Otherwise, this is safe to transform, determine if it is profitable.
7514 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7515 // Indexes are often folded into load/store instructions, so we don't want to
7516 // hide them behind a phi.
7517 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7520 Value *InLHS = FirstInst->getOperand(0);
7521 Value *InRHS = FirstInst->getOperand(1);
7522 PHINode *NewLHS = 0, *NewRHS = 0;
7524 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7525 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7526 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7527 InsertNewInstBefore(NewLHS, PN);
7532 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7533 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7534 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7535 InsertNewInstBefore(NewRHS, PN);
7539 // Add all operands to the new PHIs.
7540 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7542 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7543 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7546 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7547 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7551 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7552 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7553 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7554 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7557 assert(isa<GetElementPtrInst>(FirstInst));
7558 return new GetElementPtrInst(LHSVal, RHSVal);
7562 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7563 /// of the block that defines it. This means that it must be obvious the value
7564 /// of the load is not changed from the point of the load to the end of the
7567 /// Finally, it is safe, but not profitable, to sink a load targetting a
7568 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
7570 static bool isSafeToSinkLoad(LoadInst *L) {
7571 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7573 for (++BBI; BBI != E; ++BBI)
7574 if (BBI->mayWriteToMemory())
7577 // Check for non-address taken alloca. If not address-taken already, it isn't
7578 // profitable to do this xform.
7579 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
7580 bool isAddressTaken = false;
7581 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
7583 if (isa<LoadInst>(UI)) continue;
7584 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
7585 // If storing TO the alloca, then the address isn't taken.
7586 if (SI->getOperand(1) == AI) continue;
7588 isAddressTaken = true;
7592 if (!isAddressTaken)
7600 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7601 // operator and they all are only used by the PHI, PHI together their
7602 // inputs, and do the operation once, to the result of the PHI.
7603 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7604 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7606 // Scan the instruction, looking for input operations that can be folded away.
7607 // If all input operands to the phi are the same instruction (e.g. a cast from
7608 // the same type or "+42") we can pull the operation through the PHI, reducing
7609 // code size and simplifying code.
7610 Constant *ConstantOp = 0;
7611 const Type *CastSrcTy = 0;
7612 bool isVolatile = false;
7613 if (isa<CastInst>(FirstInst)) {
7614 CastSrcTy = FirstInst->getOperand(0)->getType();
7615 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
7616 // Can fold binop, compare or shift here if the RHS is a constant,
7617 // otherwise call FoldPHIArgBinOpIntoPHI.
7618 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7619 if (ConstantOp == 0)
7620 return FoldPHIArgBinOpIntoPHI(PN);
7621 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7622 isVolatile = LI->isVolatile();
7623 // We can't sink the load if the loaded value could be modified between the
7624 // load and the PHI.
7625 if (LI->getParent() != PN.getIncomingBlock(0) ||
7626 !isSafeToSinkLoad(LI))
7628 } else if (isa<GetElementPtrInst>(FirstInst)) {
7629 if (FirstInst->getNumOperands() == 2)
7630 return FoldPHIArgBinOpIntoPHI(PN);
7631 // Can't handle general GEPs yet.
7634 return 0; // Cannot fold this operation.
7637 // Check to see if all arguments are the same operation.
7638 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7639 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7640 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7641 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7644 if (I->getOperand(0)->getType() != CastSrcTy)
7645 return 0; // Cast operation must match.
7646 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7647 // We can't sink the load if the loaded value could be modified between
7648 // the load and the PHI.
7649 if (LI->isVolatile() != isVolatile ||
7650 LI->getParent() != PN.getIncomingBlock(i) ||
7651 !isSafeToSinkLoad(LI))
7653 } else if (I->getOperand(1) != ConstantOp) {
7658 // Okay, they are all the same operation. Create a new PHI node of the
7659 // correct type, and PHI together all of the LHS's of the instructions.
7660 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7661 PN.getName()+".in");
7662 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7664 Value *InVal = FirstInst->getOperand(0);
7665 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7667 // Add all operands to the new PHI.
7668 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7669 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7670 if (NewInVal != InVal)
7672 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7677 // The new PHI unions all of the same values together. This is really
7678 // common, so we handle it intelligently here for compile-time speed.
7682 InsertNewInstBefore(NewPN, PN);
7686 // Insert and return the new operation.
7687 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7688 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7689 else if (isa<LoadInst>(FirstInst))
7690 return new LoadInst(PhiVal, "", isVolatile);
7691 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7692 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7693 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7694 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7695 PhiVal, ConstantOp);
7697 assert(0 && "Unknown operation");
7701 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7703 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7704 if (PN->use_empty()) return true;
7705 if (!PN->hasOneUse()) return false;
7707 // Remember this node, and if we find the cycle, return.
7708 if (!PotentiallyDeadPHIs.insert(PN).second)
7711 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7712 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7717 // PHINode simplification
7719 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7720 // If LCSSA is around, don't mess with Phi nodes
7721 if (MustPreserveLCSSA) return 0;
7723 if (Value *V = PN.hasConstantValue())
7724 return ReplaceInstUsesWith(PN, V);
7726 // If all PHI operands are the same operation, pull them through the PHI,
7727 // reducing code size.
7728 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7729 PN.getIncomingValue(0)->hasOneUse())
7730 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7733 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7734 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7735 // PHI)... break the cycle.
7736 if (PN.hasOneUse()) {
7737 Instruction *PHIUser = cast<Instruction>(PN.use_back());
7738 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
7739 std::set<PHINode*> PotentiallyDeadPHIs;
7740 PotentiallyDeadPHIs.insert(&PN);
7741 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7742 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7745 // If this phi has a single use, and if that use just computes a value for
7746 // the next iteration of a loop, delete the phi. This occurs with unused
7747 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
7748 // common case here is good because the only other things that catch this
7749 // are induction variable analysis (sometimes) and ADCE, which is only run
7751 if (PHIUser->hasOneUse() &&
7752 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
7753 PHIUser->use_back() == &PN) {
7754 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7761 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7762 Instruction *InsertPoint,
7764 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7765 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7766 // We must cast correctly to the pointer type. Ensure that we
7767 // sign extend the integer value if it is smaller as this is
7768 // used for address computation.
7769 Instruction::CastOps opcode =
7770 (VTySize < PtrSize ? Instruction::SExt :
7771 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7772 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7776 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7777 Value *PtrOp = GEP.getOperand(0);
7778 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7779 // If so, eliminate the noop.
7780 if (GEP.getNumOperands() == 1)
7781 return ReplaceInstUsesWith(GEP, PtrOp);
7783 if (isa<UndefValue>(GEP.getOperand(0)))
7784 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7786 bool HasZeroPointerIndex = false;
7787 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7788 HasZeroPointerIndex = C->isNullValue();
7790 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7791 return ReplaceInstUsesWith(GEP, PtrOp);
7793 // Eliminate unneeded casts for indices.
7794 bool MadeChange = false;
7795 gep_type_iterator GTI = gep_type_begin(GEP);
7796 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7797 if (isa<SequentialType>(*GTI)) {
7798 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7799 if (CI->getOpcode() == Instruction::ZExt ||
7800 CI->getOpcode() == Instruction::SExt) {
7801 const Type *SrcTy = CI->getOperand(0)->getType();
7802 // We can eliminate a cast from i32 to i64 iff the target
7803 // is a 32-bit pointer target.
7804 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7806 GEP.setOperand(i, CI->getOperand(0));
7810 // If we are using a wider index than needed for this platform, shrink it
7811 // to what we need. If the incoming value needs a cast instruction,
7812 // insert it. This explicit cast can make subsequent optimizations more
7814 Value *Op = GEP.getOperand(i);
7815 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7816 if (Constant *C = dyn_cast<Constant>(Op)) {
7817 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7820 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7822 GEP.setOperand(i, Op);
7826 if (MadeChange) return &GEP;
7828 // Combine Indices - If the source pointer to this getelementptr instruction
7829 // is a getelementptr instruction, combine the indices of the two
7830 // getelementptr instructions into a single instruction.
7832 SmallVector<Value*, 8> SrcGEPOperands;
7833 if (User *Src = dyn_castGetElementPtr(PtrOp))
7834 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
7836 if (!SrcGEPOperands.empty()) {
7837 // Note that if our source is a gep chain itself that we wait for that
7838 // chain to be resolved before we perform this transformation. This
7839 // avoids us creating a TON of code in some cases.
7841 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7842 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7843 return 0; // Wait until our source is folded to completion.
7845 SmallVector<Value*, 8> Indices;
7847 // Find out whether the last index in the source GEP is a sequential idx.
7848 bool EndsWithSequential = false;
7849 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7850 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7851 EndsWithSequential = !isa<StructType>(*I);
7853 // Can we combine the two pointer arithmetics offsets?
7854 if (EndsWithSequential) {
7855 // Replace: gep (gep %P, long B), long A, ...
7856 // With: T = long A+B; gep %P, T, ...
7858 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7859 if (SO1 == Constant::getNullValue(SO1->getType())) {
7861 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7864 // If they aren't the same type, convert both to an integer of the
7865 // target's pointer size.
7866 if (SO1->getType() != GO1->getType()) {
7867 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7868 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7869 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7870 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7872 unsigned PS = TD->getPointerSize();
7873 if (TD->getTypeSize(SO1->getType()) == PS) {
7874 // Convert GO1 to SO1's type.
7875 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7877 } else if (TD->getTypeSize(GO1->getType()) == PS) {
7878 // Convert SO1 to GO1's type.
7879 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7881 const Type *PT = TD->getIntPtrType();
7882 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7883 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7887 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7888 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7890 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7891 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7895 // Recycle the GEP we already have if possible.
7896 if (SrcGEPOperands.size() == 2) {
7897 GEP.setOperand(0, SrcGEPOperands[0]);
7898 GEP.setOperand(1, Sum);
7901 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7902 SrcGEPOperands.end()-1);
7903 Indices.push_back(Sum);
7904 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7906 } else if (isa<Constant>(*GEP.idx_begin()) &&
7907 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7908 SrcGEPOperands.size() != 1) {
7909 // Otherwise we can do the fold if the first index of the GEP is a zero
7910 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7911 SrcGEPOperands.end());
7912 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7915 if (!Indices.empty())
7916 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
7917 Indices.size(), GEP.getName());
7919 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7920 // GEP of global variable. If all of the indices for this GEP are
7921 // constants, we can promote this to a constexpr instead of an instruction.
7923 // Scan for nonconstants...
7924 SmallVector<Constant*, 8> Indices;
7925 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7926 for (; I != E && isa<Constant>(*I); ++I)
7927 Indices.push_back(cast<Constant>(*I));
7929 if (I == E) { // If they are all constants...
7930 Constant *CE = ConstantExpr::getGetElementPtr(GV,
7931 &Indices[0],Indices.size());
7933 // Replace all uses of the GEP with the new constexpr...
7934 return ReplaceInstUsesWith(GEP, CE);
7936 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7937 if (!isa<PointerType>(X->getType())) {
7938 // Not interesting. Source pointer must be a cast from pointer.
7939 } else if (HasZeroPointerIndex) {
7940 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7941 // into : GEP [10 x ubyte]* X, long 0, ...
7943 // This occurs when the program declares an array extern like "int X[];"
7945 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7946 const PointerType *XTy = cast<PointerType>(X->getType());
7947 if (const ArrayType *XATy =
7948 dyn_cast<ArrayType>(XTy->getElementType()))
7949 if (const ArrayType *CATy =
7950 dyn_cast<ArrayType>(CPTy->getElementType()))
7951 if (CATy->getElementType() == XATy->getElementType()) {
7952 // At this point, we know that the cast source type is a pointer
7953 // to an array of the same type as the destination pointer
7954 // array. Because the array type is never stepped over (there
7955 // is a leading zero) we can fold the cast into this GEP.
7956 GEP.setOperand(0, X);
7959 } else if (GEP.getNumOperands() == 2) {
7960 // Transform things like:
7961 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7962 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7963 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7964 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7965 if (isa<ArrayType>(SrcElTy) &&
7966 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7967 TD->getTypeSize(ResElTy)) {
7968 Value *V = InsertNewInstBefore(
7969 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7970 GEP.getOperand(1), GEP.getName()), GEP);
7971 // V and GEP are both pointer types --> BitCast
7972 return new BitCastInst(V, GEP.getType());
7975 // Transform things like:
7976 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7977 // (where tmp = 8*tmp2) into:
7978 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7980 if (isa<ArrayType>(SrcElTy) &&
7981 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
7982 uint64_t ArrayEltSize =
7983 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7985 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7986 // allow either a mul, shift, or constant here.
7988 ConstantInt *Scale = 0;
7989 if (ArrayEltSize == 1) {
7990 NewIdx = GEP.getOperand(1);
7991 Scale = ConstantInt::get(NewIdx->getType(), 1);
7992 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7993 NewIdx = ConstantInt::get(CI->getType(), 1);
7995 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7996 if (Inst->getOpcode() == Instruction::Shl &&
7997 isa<ConstantInt>(Inst->getOperand(1))) {
7999 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
8000 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
8001 NewIdx = Inst->getOperand(0);
8002 } else if (Inst->getOpcode() == Instruction::Mul &&
8003 isa<ConstantInt>(Inst->getOperand(1))) {
8004 Scale = cast<ConstantInt>(Inst->getOperand(1));
8005 NewIdx = Inst->getOperand(0);
8009 // If the index will be to exactly the right offset with the scale taken
8010 // out, perform the transformation.
8011 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8012 if (isa<ConstantInt>(Scale))
8013 Scale = ConstantInt::get(Scale->getType(),
8014 Scale->getZExtValue() / ArrayEltSize);
8015 if (Scale->getZExtValue() != 1) {
8016 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8018 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8019 NewIdx = InsertNewInstBefore(Sc, GEP);
8022 // Insert the new GEP instruction.
8023 Instruction *NewGEP =
8024 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8025 NewIdx, GEP.getName());
8026 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8027 // The NewGEP must be pointer typed, so must the old one -> BitCast
8028 return new BitCastInst(NewGEP, GEP.getType());
8037 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8038 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8039 if (AI.isArrayAllocation()) // Check C != 1
8040 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8042 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8043 AllocationInst *New = 0;
8045 // Create and insert the replacement instruction...
8046 if (isa<MallocInst>(AI))
8047 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8049 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8050 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8053 InsertNewInstBefore(New, AI);
8055 // Scan to the end of the allocation instructions, to skip over a block of
8056 // allocas if possible...
8058 BasicBlock::iterator It = New;
8059 while (isa<AllocationInst>(*It)) ++It;
8061 // Now that I is pointing to the first non-allocation-inst in the block,
8062 // insert our getelementptr instruction...
8064 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8065 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8066 New->getName()+".sub", It);
8068 // Now make everything use the getelementptr instead of the original
8070 return ReplaceInstUsesWith(AI, V);
8071 } else if (isa<UndefValue>(AI.getArraySize())) {
8072 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8075 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8076 // Note that we only do this for alloca's, because malloc should allocate and
8077 // return a unique pointer, even for a zero byte allocation.
8078 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8079 TD->getTypeSize(AI.getAllocatedType()) == 0)
8080 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8085 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8086 Value *Op = FI.getOperand(0);
8088 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8089 if (CastInst *CI = dyn_cast<CastInst>(Op))
8090 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8091 FI.setOperand(0, CI->getOperand(0));
8095 // free undef -> unreachable.
8096 if (isa<UndefValue>(Op)) {
8097 // Insert a new store to null because we cannot modify the CFG here.
8098 new StoreInst(ConstantInt::getTrue(),
8099 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8100 return EraseInstFromFunction(FI);
8103 // If we have 'free null' delete the instruction. This can happen in stl code
8104 // when lots of inlining happens.
8105 if (isa<ConstantPointerNull>(Op))
8106 return EraseInstFromFunction(FI);
8112 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8113 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8114 User *CI = cast<User>(LI.getOperand(0));
8115 Value *CastOp = CI->getOperand(0);
8117 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8118 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8119 const Type *SrcPTy = SrcTy->getElementType();
8121 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8122 isa<VectorType>(DestPTy)) {
8123 // If the source is an array, the code below will not succeed. Check to
8124 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8126 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8127 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8128 if (ASrcTy->getNumElements() != 0) {
8130 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8131 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8132 SrcTy = cast<PointerType>(CastOp->getType());
8133 SrcPTy = SrcTy->getElementType();
8136 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8137 isa<VectorType>(SrcPTy)) &&
8138 // Do not allow turning this into a load of an integer, which is then
8139 // casted to a pointer, this pessimizes pointer analysis a lot.
8140 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8141 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8142 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8144 // Okay, we are casting from one integer or pointer type to another of
8145 // the same size. Instead of casting the pointer before the load, cast
8146 // the result of the loaded value.
8147 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8149 LI.isVolatile()),LI);
8150 // Now cast the result of the load.
8151 return new BitCastInst(NewLoad, LI.getType());
8158 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8159 /// from this value cannot trap. If it is not obviously safe to load from the
8160 /// specified pointer, we do a quick local scan of the basic block containing
8161 /// ScanFrom, to determine if the address is already accessed.
8162 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8163 // If it is an alloca or global variable, it is always safe to load from.
8164 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8166 // Otherwise, be a little bit agressive by scanning the local block where we
8167 // want to check to see if the pointer is already being loaded or stored
8168 // from/to. If so, the previous load or store would have already trapped,
8169 // so there is no harm doing an extra load (also, CSE will later eliminate
8170 // the load entirely).
8171 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8176 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8177 if (LI->getOperand(0) == V) return true;
8178 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8179 if (SI->getOperand(1) == V) return true;
8185 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8186 Value *Op = LI.getOperand(0);
8188 // load (cast X) --> cast (load X) iff safe
8189 if (isa<CastInst>(Op))
8190 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8193 // None of the following transforms are legal for volatile loads.
8194 if (LI.isVolatile()) return 0;
8196 if (&LI.getParent()->front() != &LI) {
8197 BasicBlock::iterator BBI = &LI; --BBI;
8198 // If the instruction immediately before this is a store to the same
8199 // address, do a simple form of store->load forwarding.
8200 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8201 if (SI->getOperand(1) == LI.getOperand(0))
8202 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8203 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8204 if (LIB->getOperand(0) == LI.getOperand(0))
8205 return ReplaceInstUsesWith(LI, LIB);
8208 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8209 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8210 isa<UndefValue>(GEPI->getOperand(0))) {
8211 // Insert a new store to null instruction before the load to indicate
8212 // that this code is not reachable. We do this instead of inserting
8213 // an unreachable instruction directly because we cannot modify the
8215 new StoreInst(UndefValue::get(LI.getType()),
8216 Constant::getNullValue(Op->getType()), &LI);
8217 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8220 if (Constant *C = dyn_cast<Constant>(Op)) {
8221 // load null/undef -> undef
8222 if ((C->isNullValue() || isa<UndefValue>(C))) {
8223 // Insert a new store to null instruction before the load to indicate that
8224 // this code is not reachable. We do this instead of inserting an
8225 // unreachable instruction directly because we cannot modify the CFG.
8226 new StoreInst(UndefValue::get(LI.getType()),
8227 Constant::getNullValue(Op->getType()), &LI);
8228 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8231 // Instcombine load (constant global) into the value loaded.
8232 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8233 if (GV->isConstant() && !GV->isDeclaration())
8234 return ReplaceInstUsesWith(LI, GV->getInitializer());
8236 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8237 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8238 if (CE->getOpcode() == Instruction::GetElementPtr) {
8239 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8240 if (GV->isConstant() && !GV->isDeclaration())
8242 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8243 return ReplaceInstUsesWith(LI, V);
8244 if (CE->getOperand(0)->isNullValue()) {
8245 // Insert a new store to null instruction before the load to indicate
8246 // that this code is not reachable. We do this instead of inserting
8247 // an unreachable instruction directly because we cannot modify the
8249 new StoreInst(UndefValue::get(LI.getType()),
8250 Constant::getNullValue(Op->getType()), &LI);
8251 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8254 } else if (CE->isCast()) {
8255 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8260 if (Op->hasOneUse()) {
8261 // Change select and PHI nodes to select values instead of addresses: this
8262 // helps alias analysis out a lot, allows many others simplifications, and
8263 // exposes redundancy in the code.
8265 // Note that we cannot do the transformation unless we know that the
8266 // introduced loads cannot trap! Something like this is valid as long as
8267 // the condition is always false: load (select bool %C, int* null, int* %G),
8268 // but it would not be valid if we transformed it to load from null
8271 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8272 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8273 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8274 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8275 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8276 SI->getOperand(1)->getName()+".val"), LI);
8277 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8278 SI->getOperand(2)->getName()+".val"), LI);
8279 return new SelectInst(SI->getCondition(), V1, V2);
8282 // load (select (cond, null, P)) -> load P
8283 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8284 if (C->isNullValue()) {
8285 LI.setOperand(0, SI->getOperand(2));
8289 // load (select (cond, P, null)) -> load P
8290 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8291 if (C->isNullValue()) {
8292 LI.setOperand(0, SI->getOperand(1));
8300 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8302 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8303 User *CI = cast<User>(SI.getOperand(1));
8304 Value *CastOp = CI->getOperand(0);
8306 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8307 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8308 const Type *SrcPTy = SrcTy->getElementType();
8310 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8311 // If the source is an array, the code below will not succeed. Check to
8312 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8314 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8315 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8316 if (ASrcTy->getNumElements() != 0) {
8318 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8319 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8320 SrcTy = cast<PointerType>(CastOp->getType());
8321 SrcPTy = SrcTy->getElementType();
8324 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8325 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8326 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8328 // Okay, we are casting from one integer or pointer type to another of
8329 // the same size. Instead of casting the pointer before
8330 // the store, cast the value to be stored.
8332 Value *SIOp0 = SI.getOperand(0);
8333 Instruction::CastOps opcode = Instruction::BitCast;
8334 const Type* CastSrcTy = SIOp0->getType();
8335 const Type* CastDstTy = SrcPTy;
8336 if (isa<PointerType>(CastDstTy)) {
8337 if (CastSrcTy->isInteger())
8338 opcode = Instruction::IntToPtr;
8339 } else if (isa<IntegerType>(CastDstTy)) {
8340 if (isa<PointerType>(SIOp0->getType()))
8341 opcode = Instruction::PtrToInt;
8343 if (Constant *C = dyn_cast<Constant>(SIOp0))
8344 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8346 NewCast = IC.InsertNewInstBefore(
8347 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8349 return new StoreInst(NewCast, CastOp);
8356 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8357 Value *Val = SI.getOperand(0);
8358 Value *Ptr = SI.getOperand(1);
8360 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8361 EraseInstFromFunction(SI);
8366 // If the RHS is an alloca with a single use, zapify the store, making the
8368 if (Ptr->hasOneUse()) {
8369 if (isa<AllocaInst>(Ptr)) {
8370 EraseInstFromFunction(SI);
8375 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8376 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8377 GEP->getOperand(0)->hasOneUse()) {
8378 EraseInstFromFunction(SI);
8384 // Do really simple DSE, to catch cases where there are several consequtive
8385 // stores to the same location, separated by a few arithmetic operations. This
8386 // situation often occurs with bitfield accesses.
8387 BasicBlock::iterator BBI = &SI;
8388 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8392 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8393 // Prev store isn't volatile, and stores to the same location?
8394 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8397 EraseInstFromFunction(*PrevSI);
8403 // If this is a load, we have to stop. However, if the loaded value is from
8404 // the pointer we're loading and is producing the pointer we're storing,
8405 // then *this* store is dead (X = load P; store X -> P).
8406 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8407 if (LI == Val && LI->getOperand(0) == Ptr) {
8408 EraseInstFromFunction(SI);
8412 // Otherwise, this is a load from some other location. Stores before it
8417 // Don't skip over loads or things that can modify memory.
8418 if (BBI->mayWriteToMemory())
8423 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8425 // store X, null -> turns into 'unreachable' in SimplifyCFG
8426 if (isa<ConstantPointerNull>(Ptr)) {
8427 if (!isa<UndefValue>(Val)) {
8428 SI.setOperand(0, UndefValue::get(Val->getType()));
8429 if (Instruction *U = dyn_cast<Instruction>(Val))
8430 AddToWorkList(U); // Dropped a use.
8433 return 0; // Do not modify these!
8436 // store undef, Ptr -> noop
8437 if (isa<UndefValue>(Val)) {
8438 EraseInstFromFunction(SI);
8443 // If the pointer destination is a cast, see if we can fold the cast into the
8445 if (isa<CastInst>(Ptr))
8446 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8448 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8450 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8454 // If this store is the last instruction in the basic block, and if the block
8455 // ends with an unconditional branch, try to move it to the successor block.
8457 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8458 if (BI->isUnconditional()) {
8459 // Check to see if the successor block has exactly two incoming edges. If
8460 // so, see if the other predecessor contains a store to the same location.
8461 // if so, insert a PHI node (if needed) and move the stores down.
8462 BasicBlock *Dest = BI->getSuccessor(0);
8464 pred_iterator PI = pred_begin(Dest);
8465 BasicBlock *Other = 0;
8466 if (*PI != BI->getParent())
8469 if (PI != pred_end(Dest)) {
8470 if (*PI != BI->getParent())
8475 if (++PI != pred_end(Dest))
8478 if (Other) { // If only one other pred...
8479 BBI = Other->getTerminator();
8480 // Make sure this other block ends in an unconditional branch and that
8481 // there is an instruction before the branch.
8482 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8483 BBI != Other->begin()) {
8485 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8487 // If this instruction is a store to the same location.
8488 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8489 // Okay, we know we can perform this transformation. Insert a PHI
8490 // node now if we need it.
8491 Value *MergedVal = OtherStore->getOperand(0);
8492 if (MergedVal != SI.getOperand(0)) {
8493 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8494 PN->reserveOperandSpace(2);
8495 PN->addIncoming(SI.getOperand(0), SI.getParent());
8496 PN->addIncoming(OtherStore->getOperand(0), Other);
8497 MergedVal = InsertNewInstBefore(PN, Dest->front());
8500 // Advance to a place where it is safe to insert the new store and
8502 BBI = Dest->begin();
8503 while (isa<PHINode>(BBI)) ++BBI;
8504 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8505 OtherStore->isVolatile()), *BBI);
8507 // Nuke the old stores.
8508 EraseInstFromFunction(SI);
8509 EraseInstFromFunction(*OtherStore);
8521 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8522 // Change br (not X), label True, label False to: br X, label False, True
8524 BasicBlock *TrueDest;
8525 BasicBlock *FalseDest;
8526 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8527 !isa<Constant>(X)) {
8528 // Swap Destinations and condition...
8530 BI.setSuccessor(0, FalseDest);
8531 BI.setSuccessor(1, TrueDest);
8535 // Cannonicalize fcmp_one -> fcmp_oeq
8536 FCmpInst::Predicate FPred; Value *Y;
8537 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8538 TrueDest, FalseDest)))
8539 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8540 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8541 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8542 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8543 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
8544 NewSCC->takeName(I);
8545 // Swap Destinations and condition...
8546 BI.setCondition(NewSCC);
8547 BI.setSuccessor(0, FalseDest);
8548 BI.setSuccessor(1, TrueDest);
8549 RemoveFromWorkList(I);
8550 I->eraseFromParent();
8551 AddToWorkList(NewSCC);
8555 // Cannonicalize icmp_ne -> icmp_eq
8556 ICmpInst::Predicate IPred;
8557 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8558 TrueDest, FalseDest)))
8559 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8560 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8561 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8562 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8563 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8564 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
8565 NewSCC->takeName(I);
8566 // Swap Destinations and condition...
8567 BI.setCondition(NewSCC);
8568 BI.setSuccessor(0, FalseDest);
8569 BI.setSuccessor(1, TrueDest);
8570 RemoveFromWorkList(I);
8571 I->eraseFromParent();;
8572 AddToWorkList(NewSCC);
8579 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8580 Value *Cond = SI.getCondition();
8581 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8582 if (I->getOpcode() == Instruction::Add)
8583 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8584 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8585 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8586 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8588 SI.setOperand(0, I->getOperand(0));
8596 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8597 /// is to leave as a vector operation.
8598 static bool CheapToScalarize(Value *V, bool isConstant) {
8599 if (isa<ConstantAggregateZero>(V))
8601 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
8602 if (isConstant) return true;
8603 // If all elts are the same, we can extract.
8604 Constant *Op0 = C->getOperand(0);
8605 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8606 if (C->getOperand(i) != Op0)
8610 Instruction *I = dyn_cast<Instruction>(V);
8611 if (!I) return false;
8613 // Insert element gets simplified to the inserted element or is deleted if
8614 // this is constant idx extract element and its a constant idx insertelt.
8615 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8616 isa<ConstantInt>(I->getOperand(2)))
8618 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8620 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8621 if (BO->hasOneUse() &&
8622 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8623 CheapToScalarize(BO->getOperand(1), isConstant)))
8625 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8626 if (CI->hasOneUse() &&
8627 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8628 CheapToScalarize(CI->getOperand(1), isConstant)))
8634 /// Read and decode a shufflevector mask.
8636 /// It turns undef elements into values that are larger than the number of
8637 /// elements in the input.
8638 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8639 unsigned NElts = SVI->getType()->getNumElements();
8640 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8641 return std::vector<unsigned>(NElts, 0);
8642 if (isa<UndefValue>(SVI->getOperand(2)))
8643 return std::vector<unsigned>(NElts, 2*NElts);
8645 std::vector<unsigned> Result;
8646 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
8647 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8648 if (isa<UndefValue>(CP->getOperand(i)))
8649 Result.push_back(NElts*2); // undef -> 8
8651 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8655 /// FindScalarElement - Given a vector and an element number, see if the scalar
8656 /// value is already around as a register, for example if it were inserted then
8657 /// extracted from the vector.
8658 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8659 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
8660 const VectorType *PTy = cast<VectorType>(V->getType());
8661 unsigned Width = PTy->getNumElements();
8662 if (EltNo >= Width) // Out of range access.
8663 return UndefValue::get(PTy->getElementType());
8665 if (isa<UndefValue>(V))
8666 return UndefValue::get(PTy->getElementType());
8667 else if (isa<ConstantAggregateZero>(V))
8668 return Constant::getNullValue(PTy->getElementType());
8669 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
8670 return CP->getOperand(EltNo);
8671 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8672 // If this is an insert to a variable element, we don't know what it is.
8673 if (!isa<ConstantInt>(III->getOperand(2)))
8675 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8677 // If this is an insert to the element we are looking for, return the
8680 return III->getOperand(1);
8682 // Otherwise, the insertelement doesn't modify the value, recurse on its
8684 return FindScalarElement(III->getOperand(0), EltNo);
8685 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8686 unsigned InEl = getShuffleMask(SVI)[EltNo];
8688 return FindScalarElement(SVI->getOperand(0), InEl);
8689 else if (InEl < Width*2)
8690 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8692 return UndefValue::get(PTy->getElementType());
8695 // Otherwise, we don't know.
8699 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8701 // If packed val is undef, replace extract with scalar undef.
8702 if (isa<UndefValue>(EI.getOperand(0)))
8703 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8705 // If packed val is constant 0, replace extract with scalar 0.
8706 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8707 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8709 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
8710 // If packed val is constant with uniform operands, replace EI
8711 // with that operand
8712 Constant *op0 = C->getOperand(0);
8713 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8714 if (C->getOperand(i) != op0) {
8719 return ReplaceInstUsesWith(EI, op0);
8722 // If extracting a specified index from the vector, see if we can recursively
8723 // find a previously computed scalar that was inserted into the vector.
8724 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8725 // This instruction only demands the single element from the input vector.
8726 // If the input vector has a single use, simplify it based on this use
8728 uint64_t IndexVal = IdxC->getZExtValue();
8729 if (EI.getOperand(0)->hasOneUse()) {
8731 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8734 EI.setOperand(0, V);
8739 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8740 return ReplaceInstUsesWith(EI, Elt);
8743 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8744 if (I->hasOneUse()) {
8745 // Push extractelement into predecessor operation if legal and
8746 // profitable to do so
8747 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8748 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8749 if (CheapToScalarize(BO, isConstantElt)) {
8750 ExtractElementInst *newEI0 =
8751 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8752 EI.getName()+".lhs");
8753 ExtractElementInst *newEI1 =
8754 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8755 EI.getName()+".rhs");
8756 InsertNewInstBefore(newEI0, EI);
8757 InsertNewInstBefore(newEI1, EI);
8758 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8760 } else if (isa<LoadInst>(I)) {
8761 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8762 PointerType::get(EI.getType()), EI);
8763 GetElementPtrInst *GEP =
8764 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8765 InsertNewInstBefore(GEP, EI);
8766 return new LoadInst(GEP);
8769 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8770 // Extracting the inserted element?
8771 if (IE->getOperand(2) == EI.getOperand(1))
8772 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8773 // If the inserted and extracted elements are constants, they must not
8774 // be the same value, extract from the pre-inserted value instead.
8775 if (isa<Constant>(IE->getOperand(2)) &&
8776 isa<Constant>(EI.getOperand(1))) {
8777 AddUsesToWorkList(EI);
8778 EI.setOperand(0, IE->getOperand(0));
8781 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8782 // If this is extracting an element from a shufflevector, figure out where
8783 // it came from and extract from the appropriate input element instead.
8784 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8785 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8787 if (SrcIdx < SVI->getType()->getNumElements())
8788 Src = SVI->getOperand(0);
8789 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8790 SrcIdx -= SVI->getType()->getNumElements();
8791 Src = SVI->getOperand(1);
8793 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8795 return new ExtractElementInst(Src, SrcIdx);
8802 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8803 /// elements from either LHS or RHS, return the shuffle mask and true.
8804 /// Otherwise, return false.
8805 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8806 std::vector<Constant*> &Mask) {
8807 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8808 "Invalid CollectSingleShuffleElements");
8809 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8811 if (isa<UndefValue>(V)) {
8812 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8814 } else if (V == LHS) {
8815 for (unsigned i = 0; i != NumElts; ++i)
8816 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8818 } else if (V == RHS) {
8819 for (unsigned i = 0; i != NumElts; ++i)
8820 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8822 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8823 // If this is an insert of an extract from some other vector, include it.
8824 Value *VecOp = IEI->getOperand(0);
8825 Value *ScalarOp = IEI->getOperand(1);
8826 Value *IdxOp = IEI->getOperand(2);
8828 if (!isa<ConstantInt>(IdxOp))
8830 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8832 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8833 // Okay, we can handle this if the vector we are insertinting into is
8835 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8836 // If so, update the mask to reflect the inserted undef.
8837 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8840 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8841 if (isa<ConstantInt>(EI->getOperand(1)) &&
8842 EI->getOperand(0)->getType() == V->getType()) {
8843 unsigned ExtractedIdx =
8844 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8846 // This must be extracting from either LHS or RHS.
8847 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8848 // Okay, we can handle this if the vector we are insertinting into is
8850 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8851 // If so, update the mask to reflect the inserted value.
8852 if (EI->getOperand(0) == LHS) {
8853 Mask[InsertedIdx & (NumElts-1)] =
8854 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8856 assert(EI->getOperand(0) == RHS);
8857 Mask[InsertedIdx & (NumElts-1)] =
8858 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
8867 // TODO: Handle shufflevector here!
8872 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8873 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8874 /// that computes V and the LHS value of the shuffle.
8875 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8877 assert(isa<VectorType>(V->getType()) &&
8878 (RHS == 0 || V->getType() == RHS->getType()) &&
8879 "Invalid shuffle!");
8880 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8882 if (isa<UndefValue>(V)) {
8883 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8885 } else if (isa<ConstantAggregateZero>(V)) {
8886 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
8888 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8889 // If this is an insert of an extract from some other vector, include it.
8890 Value *VecOp = IEI->getOperand(0);
8891 Value *ScalarOp = IEI->getOperand(1);
8892 Value *IdxOp = IEI->getOperand(2);
8894 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8895 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8896 EI->getOperand(0)->getType() == V->getType()) {
8897 unsigned ExtractedIdx =
8898 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8899 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8901 // Either the extracted from or inserted into vector must be RHSVec,
8902 // otherwise we'd end up with a shuffle of three inputs.
8903 if (EI->getOperand(0) == RHS || RHS == 0) {
8904 RHS = EI->getOperand(0);
8905 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8906 Mask[InsertedIdx & (NumElts-1)] =
8907 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
8912 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8913 // Everything but the extracted element is replaced with the RHS.
8914 for (unsigned i = 0; i != NumElts; ++i) {
8915 if (i != InsertedIdx)
8916 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
8921 // If this insertelement is a chain that comes from exactly these two
8922 // vectors, return the vector and the effective shuffle.
8923 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8924 return EI->getOperand(0);
8929 // TODO: Handle shufflevector here!
8931 // Otherwise, can't do anything fancy. Return an identity vector.
8932 for (unsigned i = 0; i != NumElts; ++i)
8933 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8937 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8938 Value *VecOp = IE.getOperand(0);
8939 Value *ScalarOp = IE.getOperand(1);
8940 Value *IdxOp = IE.getOperand(2);
8942 // If the inserted element was extracted from some other vector, and if the
8943 // indexes are constant, try to turn this into a shufflevector operation.
8944 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8945 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8946 EI->getOperand(0)->getType() == IE.getType()) {
8947 unsigned NumVectorElts = IE.getType()->getNumElements();
8948 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8949 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8951 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8952 return ReplaceInstUsesWith(IE, VecOp);
8954 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8955 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8957 // If we are extracting a value from a vector, then inserting it right
8958 // back into the same place, just use the input vector.
8959 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8960 return ReplaceInstUsesWith(IE, VecOp);
8962 // We could theoretically do this for ANY input. However, doing so could
8963 // turn chains of insertelement instructions into a chain of shufflevector
8964 // instructions, and right now we do not merge shufflevectors. As such,
8965 // only do this in a situation where it is clear that there is benefit.
8966 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8967 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8968 // the values of VecOp, except then one read from EIOp0.
8969 // Build a new shuffle mask.
8970 std::vector<Constant*> Mask;
8971 if (isa<UndefValue>(VecOp))
8972 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
8974 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8975 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
8978 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8979 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8980 ConstantVector::get(Mask));
8983 // If this insertelement isn't used by some other insertelement, turn it
8984 // (and any insertelements it points to), into one big shuffle.
8985 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8986 std::vector<Constant*> Mask;
8988 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8989 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8990 // We now have a shuffle of LHS, RHS, Mask.
8991 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9000 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9001 Value *LHS = SVI.getOperand(0);
9002 Value *RHS = SVI.getOperand(1);
9003 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9005 bool MadeChange = false;
9007 // Undefined shuffle mask -> undefined value.
9008 if (isa<UndefValue>(SVI.getOperand(2)))
9009 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9011 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9012 // the undef, change them to undefs.
9013 if (isa<UndefValue>(SVI.getOperand(1))) {
9014 // Scan to see if there are any references to the RHS. If so, replace them
9015 // with undef element refs and set MadeChange to true.
9016 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9017 if (Mask[i] >= e && Mask[i] != 2*e) {
9024 // Remap any references to RHS to use LHS.
9025 std::vector<Constant*> Elts;
9026 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9028 Elts.push_back(UndefValue::get(Type::Int32Ty));
9030 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9032 SVI.setOperand(2, ConstantVector::get(Elts));
9036 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9037 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9038 if (LHS == RHS || isa<UndefValue>(LHS)) {
9039 if (isa<UndefValue>(LHS) && LHS == RHS) {
9040 // shuffle(undef,undef,mask) -> undef.
9041 return ReplaceInstUsesWith(SVI, LHS);
9044 // Remap any references to RHS to use LHS.
9045 std::vector<Constant*> Elts;
9046 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9048 Elts.push_back(UndefValue::get(Type::Int32Ty));
9050 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9051 (Mask[i] < e && isa<UndefValue>(LHS)))
9052 Mask[i] = 2*e; // Turn into undef.
9054 Mask[i] &= (e-1); // Force to LHS.
9055 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9058 SVI.setOperand(0, SVI.getOperand(1));
9059 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9060 SVI.setOperand(2, ConstantVector::get(Elts));
9061 LHS = SVI.getOperand(0);
9062 RHS = SVI.getOperand(1);
9066 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9067 bool isLHSID = true, isRHSID = true;
9069 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9070 if (Mask[i] >= e*2) continue; // Ignore undef values.
9071 // Is this an identity shuffle of the LHS value?
9072 isLHSID &= (Mask[i] == i);
9074 // Is this an identity shuffle of the RHS value?
9075 isRHSID &= (Mask[i]-e == i);
9078 // Eliminate identity shuffles.
9079 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9080 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9082 // If the LHS is a shufflevector itself, see if we can combine it with this
9083 // one without producing an unusual shuffle. Here we are really conservative:
9084 // we are absolutely afraid of producing a shuffle mask not in the input
9085 // program, because the code gen may not be smart enough to turn a merged
9086 // shuffle into two specific shuffles: it may produce worse code. As such,
9087 // we only merge two shuffles if the result is one of the two input shuffle
9088 // masks. In this case, merging the shuffles just removes one instruction,
9089 // which we know is safe. This is good for things like turning:
9090 // (splat(splat)) -> splat.
9091 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9092 if (isa<UndefValue>(RHS)) {
9093 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9095 std::vector<unsigned> NewMask;
9096 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9098 NewMask.push_back(2*e);
9100 NewMask.push_back(LHSMask[Mask[i]]);
9102 // If the result mask is equal to the src shuffle or this shuffle mask, do
9104 if (NewMask == LHSMask || NewMask == Mask) {
9105 std::vector<Constant*> Elts;
9106 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9107 if (NewMask[i] >= e*2) {
9108 Elts.push_back(UndefValue::get(Type::Int32Ty));
9110 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9113 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9114 LHSSVI->getOperand(1),
9115 ConstantVector::get(Elts));
9120 return MadeChange ? &SVI : 0;
9126 /// TryToSinkInstruction - Try to move the specified instruction from its
9127 /// current block into the beginning of DestBlock, which can only happen if it's
9128 /// safe to move the instruction past all of the instructions between it and the
9129 /// end of its block.
9130 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9131 assert(I->hasOneUse() && "Invariants didn't hold!");
9133 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9134 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9136 // Do not sink alloca instructions out of the entry block.
9137 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
9140 // We can only sink load instructions if there is nothing between the load and
9141 // the end of block that could change the value.
9142 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9143 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9145 if (Scan->mayWriteToMemory())
9149 BasicBlock::iterator InsertPos = DestBlock->begin();
9150 while (isa<PHINode>(InsertPos)) ++InsertPos;
9152 I->moveBefore(InsertPos);
9158 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9159 /// all reachable code to the worklist.
9161 /// This has a couple of tricks to make the code faster and more powerful. In
9162 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9163 /// them to the worklist (this significantly speeds up instcombine on code where
9164 /// many instructions are dead or constant). Additionally, if we find a branch
9165 /// whose condition is a known constant, we only visit the reachable successors.
9167 static void AddReachableCodeToWorklist(BasicBlock *BB,
9168 SmallPtrSet<BasicBlock*, 64> &Visited,
9170 const TargetData *TD) {
9171 // We have now visited this block! If we've already been here, bail out.
9172 if (!Visited.insert(BB)) return;
9174 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9175 Instruction *Inst = BBI++;
9177 // DCE instruction if trivially dead.
9178 if (isInstructionTriviallyDead(Inst)) {
9180 DOUT << "IC: DCE: " << *Inst;
9181 Inst->eraseFromParent();
9185 // ConstantProp instruction if trivially constant.
9186 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9187 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9188 Inst->replaceAllUsesWith(C);
9190 Inst->eraseFromParent();
9194 IC.AddToWorkList(Inst);
9197 // Recursively visit successors. If this is a branch or switch on a constant,
9198 // only visit the reachable successor.
9199 TerminatorInst *TI = BB->getTerminator();
9200 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9201 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9202 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9203 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, IC, TD);
9206 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9207 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9208 // See if this is an explicit destination.
9209 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9210 if (SI->getCaseValue(i) == Cond) {
9211 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, IC, TD);
9215 // Otherwise it is the default destination.
9216 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, IC, TD);
9221 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9222 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, IC, TD);
9225 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9226 bool Changed = false;
9227 TD = &getAnalysis<TargetData>();
9229 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9230 << F.getNameStr() << "\n");
9233 // Do a depth-first traversal of the function, populate the worklist with
9234 // the reachable instructions. Ignore blocks that are not reachable. Keep
9235 // track of which blocks we visit.
9236 SmallPtrSet<BasicBlock*, 64> Visited;
9237 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9239 // Do a quick scan over the function. If we find any blocks that are
9240 // unreachable, remove any instructions inside of them. This prevents
9241 // the instcombine code from having to deal with some bad special cases.
9242 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9243 if (!Visited.count(BB)) {
9244 Instruction *Term = BB->getTerminator();
9245 while (Term != BB->begin()) { // Remove instrs bottom-up
9246 BasicBlock::iterator I = Term; --I;
9248 DOUT << "IC: DCE: " << *I;
9251 if (!I->use_empty())
9252 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9253 I->eraseFromParent();
9258 while (!Worklist.empty()) {
9259 Instruction *I = RemoveOneFromWorkList();
9260 if (I == 0) continue; // skip null values.
9262 // Check to see if we can DCE the instruction.
9263 if (isInstructionTriviallyDead(I)) {
9264 // Add operands to the worklist.
9265 if (I->getNumOperands() < 4)
9266 AddUsesToWorkList(*I);
9269 DOUT << "IC: DCE: " << *I;
9271 I->eraseFromParent();
9272 RemoveFromWorkList(I);
9276 // Instruction isn't dead, see if we can constant propagate it.
9277 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9278 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9280 // Add operands to the worklist.
9281 AddUsesToWorkList(*I);
9282 ReplaceInstUsesWith(*I, C);
9285 I->eraseFromParent();
9286 RemoveFromWorkList(I);
9290 // See if we can trivially sink this instruction to a successor basic block.
9291 if (I->hasOneUse()) {
9292 BasicBlock *BB = I->getParent();
9293 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9294 if (UserParent != BB) {
9295 bool UserIsSuccessor = false;
9296 // See if the user is one of our successors.
9297 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9298 if (*SI == UserParent) {
9299 UserIsSuccessor = true;
9303 // If the user is one of our immediate successors, and if that successor
9304 // only has us as a predecessors (we'd have to split the critical edge
9305 // otherwise), we can keep going.
9306 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9307 next(pred_begin(UserParent)) == pred_end(UserParent))
9308 // Okay, the CFG is simple enough, try to sink this instruction.
9309 Changed |= TryToSinkInstruction(I, UserParent);
9313 // Now that we have an instruction, try combining it to simplify it...
9314 if (Instruction *Result = visit(*I)) {
9316 // Should we replace the old instruction with a new one?
9318 DOUT << "IC: Old = " << *I
9319 << " New = " << *Result;
9321 // Everything uses the new instruction now.
9322 I->replaceAllUsesWith(Result);
9324 // Push the new instruction and any users onto the worklist.
9325 AddToWorkList(Result);
9326 AddUsersToWorkList(*Result);
9328 // Move the name to the new instruction first.
9329 Result->takeName(I);
9331 // Insert the new instruction into the basic block...
9332 BasicBlock *InstParent = I->getParent();
9333 BasicBlock::iterator InsertPos = I;
9335 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9336 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9339 InstParent->getInstList().insert(InsertPos, Result);
9341 // Make sure that we reprocess all operands now that we reduced their
9343 AddUsesToWorkList(*I);
9345 // Instructions can end up on the worklist more than once. Make sure
9346 // we do not process an instruction that has been deleted.
9347 RemoveFromWorkList(I);
9349 // Erase the old instruction.
9350 InstParent->getInstList().erase(I);
9352 DOUT << "IC: MOD = " << *I;
9354 // If the instruction was modified, it's possible that it is now dead.
9355 // if so, remove it.
9356 if (isInstructionTriviallyDead(I)) {
9357 // Make sure we process all operands now that we are reducing their
9359 AddUsesToWorkList(*I);
9361 // Instructions may end up in the worklist more than once. Erase all
9362 // occurrences of this instruction.
9363 RemoveFromWorkList(I);
9364 I->eraseFromParent();
9367 AddUsersToWorkList(*I);
9374 assert(WorklistMap.empty() && "Worklist empty, but map not?");
9379 bool InstCombiner::runOnFunction(Function &F) {
9380 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
9382 bool EverMadeChange = false;
9384 // Iterate while there is work to do.
9385 unsigned Iteration = 0;
9386 while (DoOneIteration(F, Iteration++))
9387 EverMadeChange = true;
9388 return EverMadeChange;
9391 FunctionPass *llvm::createInstructionCombiningPass() {
9392 return new InstCombiner();